From a5aa64ab6b734cdbcd8dd96b0a1741973ec514f8 Mon Sep 17 00:00:00 2001 From: Davids048 Date: Sun, 4 Oct 2026 07:14:31 +0000 Subject: [PATCH] [refactor]: decide every config value before the config freezes, and drop the from_pretrained keywords VideoGenerator.from_pretrained(model_path, config) takes the same nested settings as from_config; the 24 flat keywords, from_pretrained_kwargs_to_config, FROM_PRETRAINED_KWARGS, and every flat_field in the schema are deleted. nvfp4_fa4 is the typed field engine.attention.nvfp4_fa4, applied by a resolution step. 65 call sites and the docs use the nested form; the three kwargs golden cases are config cases with byte-identical results. Every value is decided during resolution; no runtime code calls with_override. The device offload policy (unified memory, MPS, layerwise conflicts, lazy module load) runs as resolution steps in the main process (fastvideo/api/device_policy.py), and the worker runs on the config it receives. The LTX-2 refine defaults from model_index.json and the MiniMax-H3 schedule from fastvideo_inference.json are resolution steps (fastvideo/api/checkpoint_defaults.py) that read local checkpoints or download only those files; the H3 pipeline validates the schedule against the loaded schedulers but no longer writes it. The preprocessing entry point passes the downloaded local path as run state instead of a config override. Teacher and critic models load with override_transformer_cls_name as a load argument. The test isolation turns the device policy off, as it blocks downloads, so that the goldens do not depend on the machine. Against the 3f6893a0 goldens, every field still matches except boundary_ratio and ltx2_vae_tiling (DESIGN section 9) and the fields that no longer exist. The launcher and trainer YAML verifications report 0 unexplained differences. Co-Authored-By: Claude Fable 5.1 --- .../templates/pipeline_parity_test.py | 16 +- README.md | 2 +- apps/dreamverse/dreamverse/ltx2_generation.py | 2 +- docs/contributing/coding_agents.md | 2 +- .../inference_schema_parity_inventory.yaml | 59 +++-- docs/design/overview.md | 2 +- docs/design/server_contracts/dynamo.md | 14 +- .../installation/spark_pair.md | 6 +- docs/inference/architecture.md | 6 +- docs/inference/configuration.md | 14 +- docs/inference/inference_quick_start.md | 14 +- docs/inference/offloading.md | 50 ++-- docs/inference/optimizations.md | 36 +-- examples/inference/basic/README.md | 2 +- examples/inference/basic/basic.py | 20 +- .../inference/basic/basic_cosmos2_5_i2w.py | 18 +- .../inference/basic/basic_cosmos2_5_t2w.py | 18 +- .../inference/basic/basic_cosmos2_5_v2w.py | 18 +- examples/inference/basic/basic_gamecraft.py | 18 +- examples/inference/basic/basic_gen3c.py | 18 +- examples/inference/basic/basic_hy15.py | 20 +- examples/inference/basic/basic_hy15_1080p.py | 20 +- examples/inference/basic/basic_hyworld.py | 20 +- .../inference/basic/basic_kandinsky5_i2v.py | 20 +- .../inference/basic/basic_kandinsky5_t2v.py | 20 +- .../inference/basic/basic_lingbot_video.py | 16 +- .../basic/basic_lingbotworld_base_cam.py | 22 +- examples/inference/basic/basic_ltx2.py | 6 +- .../inference/basic/basic_ltx2_distilled.py | 6 +- examples/inference/basic/basic_lucy_edit.py | 18 +- examples/inference/basic/basic_matrixgame2.py | 22 +- .../basic/basic_matrixgame2_streaming.py | 22 +- examples/inference/basic/basic_matrixgame3.py | 18 +- examples/inference/basic/basic_ray.py | 22 +- .../basic/basic_self_forcing_causal.py | 14 +- .../inference/basic/basic_stable_audio.py | 6 +- .../inference/basic/basic_stable_audio_a2a.py | 6 +- .../basic/basic_stable_audio_inpaint.py | 6 +- .../basic/basic_stable_audio_small.py | 6 +- .../inference/basic/basic_turbodiffusion.py | 15 +- .../basic/basic_turbodiffusion_14b.py | 6 +- .../basic/basic_turbodiffusion_i2v.py | 6 +- examples/inference/basic/basic_wan2_2.py | 22 +- examples/inference/basic/basic_wan2_2_Fun.py | 22 +- examples/inference/basic/basic_wan2_2_i2v.py | 22 +- examples/inference/basic/basic_wan2_2_ti2v.py | 20 +- examples/inference/eval/all_metrics_demo.py | 6 +- .../inference/eval/basic_ltx2_audio_eval.py | 6 +- examples/inference/eval/basic_ltx2_eval.py | 6 +- examples/inference/eval/bench_physics_iq.py | 9 +- examples/inference/eval/bench_vbench.py | 9 +- examples/inference/eval/eval_ltx2_vbench.py | 9 +- .../local/gradio_local_demo_matrixgame2.py | 18 +- .../optimizations/attention_example.py | 18 +- .../optimizations/fp4_attn_wan2_1_1_3b.py | 26 ++- .../optimizations/torch_compile_example.py | 14 +- fastvideo/api/checkpoint_defaults.py | 215 ++++++++++++++++++ fastvideo/api/compat.py | 94 +------- fastvideo/api/device_policy.py | 171 +++++++------- fastvideo/api/inference_resolution.py | 43 +++- fastvideo/api/schema.py | 68 +++--- fastvideo/api/training_schema.py | 29 +-- fastvideo/attention/backends/flash_attn.py | 2 +- fastvideo/configs/pipelines/base.py | 4 +- fastvideo/dataset/__init__.py | 10 +- fastvideo/dataset/preprocessing_datasets.py | 11 +- fastvideo/entrypoints/streaming/server.py | 2 +- fastvideo/entrypoints/streaming/worker.py | 2 +- fastvideo/entrypoints/video_generator.py | 59 ++--- fastvideo/models/loader/component_loader.py | 42 ++-- fastvideo/models/pinned_offload.py | 2 +- .../pipelines/basic/ltx2/ltx2_pipeline.py | 89 -------- .../basic/minimax_h3/minimax_h3_pipeline.py | 21 +- .../stable_audio/stages/latent_preparation.py | 2 +- fastvideo/pipelines/composed_pipeline_base.py | 30 +-- ...ame2_preprocess_pipeline_ode_trajectory.py | 2 +- .../preprocess/preprocess_pipeline_base.py | 2 +- .../preprocess_pipeline_ode_trajectory.py | 2 +- .../preprocess/preprocess_pipeline_text.py | 2 +- .../pipelines/preprocess/v1_preprocess.py | 16 +- fastvideo/pipelines/stages/kandinsky5.py | 2 +- fastvideo/tests/api/config_snapshot.py | 94 ++++---- .../attention_precision_and_flow_shift.json | 13 +- .../config/boundary_ratio.json | 13 +- .../{kwargs => config}/disable_autocast.json | 13 +- .../config/experimental_keys.json | 13 +- .../config/ltx2_refine_lora_disabled.json | 13 +- .../config/ltx2_refine_lora_path.json | 13 +- .../config/ltx2_vae_tile_sizes.json | 13 +- ...er_port_refine_enabled_boundary_ratio.json | 10 +- .../config/pipeline_config_json_path.json | 13 +- .../torch_compile_kwargs_typed_and_extra.json | 13 +- .../typed_offload_and_parallelism.json | 13 +- .../config/vae_tiling_typed.json | 13 +- .../environment/attention_backend.json | 13 +- ..._argument_wins_over_attention_backend.json | 13 +- .../environment/inference_torch_compile.json | 13 +- .../environment/vae_parallel.json | 13 +- ...tVideo__DreamX-World-5B-Cam-Diffusers.json | 13 +- .../FastVideo__DreamX-World-5B-Diffusers.json | 13 +- .../FastVideo__FastHunyuan-diffusers.json | 13 +- ...FastVideo__FastVideo-FastH3-8-Step-V2.json | 13 +- ...FastVideo-Minimax-FastH3-Preview-v0.2.json | 13 +- ...tVideo__FastWan2.1-T2V-1.3B-Diffusers.json | 13 +- ...eo__FastWan2.1-T2V-14B-480P-Diffusers.json | 13 +- ...stVideo__FastWan2.2-TI2V-5B-Diffusers.json | 13 +- ...FastWan2.2-TI2V-5B-FullAttn-Diffusers.json | 13 +- .../FastVideo__GEN3C-Cosmos-7B-Diffusers.json | 13 +- ..._HY-WorldPlay-Bidirectional-Diffusers.json | 13 +- ...FastVideo__HunyuanGameCraft-Diffusers.json | 13 +- ...astVideo__LTX-2.3-Distilled-Diffusers.json | 13 +- .../registry/FastVideo__LTX2-Diffusers.json | 13 +- .../FastVideo__LTX2-Distilled-Diffusers.json | 13 +- .../registry/FastVideo__LTX2-base.json | 13 +- .../registry/FastVideo__LTX2.3-Diffusers.json | 13 +- ...FastVideo__LTX2.3-Distilled-Diffusers.json | 13 +- .../registry/FastVideo__LTX2.3-base.json | 13 +- ...o__LingBot-Video-Dense-1.3B-Diffusers.json | 13 +- ...__LingBot-Video-MoE-30B-A3B-Diffusers.json | 13 +- ...deo__LingBot-World-Base-Cam-Diffusers.json | 13 +- ...astVideo__LongCat-Video-I2V-Diffusers.json | 13 +- ...astVideo__LongCat-Video-T2V-Diffusers.json | 13 +- ...FastVideo__LongCat-Video-VC-Diffusers.json | 13 +- ...Video__MMAudio-large-44k-v2-Diffusers.json | 13 +- ...Video__Matrix-Game-2.0-Base-Diffusers.json | 13 +- ...rix-Game-2.0-Base-Distilled-Diffusers.json | 13 +- ...tVideo__Matrix-Game-2.0-GTA-Diffusers.json | 13 +- ...trix-Game-2.0-GTA-Distilled-Diffusers.json | 13 +- ...__Matrix-Game-2.0-TempleRun-Diffusers.json | 13 +- ...ame-2.0-TempleRun-Distilled-Diffusers.json | 13 +- ...rix-Game-3.0-Base-Distilled-Diffusers.json | 13 +- ...__SFWan2.2-I2V-A14B-Preview-Diffusers.json | 13 +- ...eo__Wan2.1-VSA-T2V-14B-720P-Diffusers.json | 13 +- ...ideo__stable-audio-open-1.0-Diffusers.json | 13 +- ...eo__stable-audio-open-small-Diffusers.json | 13 +- ...en__Wan2.1-Fun-1.3B-Control-Diffusers.json | 13 +- ...eShao__Cosmos-Predict2.5-2B-Diffusers.json | 13 +- .../registry/Lightricks__LTX-2.3.json | 13 +- .../registry/Lightricks__LTX-2.json | 13 +- .../registry/MiniMaxAI__MiniMax-H3.json | 13 +- .../registry/Tongyi-MAI__Z-Image-Turbo.json | 13 +- ...Wan-AI__Wan2.1-I2V-14B-480P-Diffusers.json | 13 +- ...Wan-AI__Wan2.1-I2V-14B-720P-Diffusers.json | 13 +- .../Wan-AI__Wan2.1-T2V-1.3B-Diffusers.json | 13 +- .../Wan-AI__Wan2.1-T2V-14B-Diffusers.json | 13 +- .../Wan-AI__Wan2.2-I2V-A14B-Diffusers.json | 13 +- .../Wan-AI__Wan2.2-T2V-A14B-Diffusers.json | 13 +- .../Wan-AI__Wan2.2-TI2V-5B-Diffusers.json | 13 +- .../black-forest-labs__FLUX.1-dev.json | 13 +- .../black-forest-labs__FLUX.2-dev.json | 13 +- .../black-forest-labs__FLUX.2-klein-4B.json | 13 +- .../black-forest-labs__FLUX.2-klein-9B.json | 13 +- .../decart-ai__Lucy-Edit-1.1-Dev.json | 13 +- .../registry/decart-ai__Lucy-Edit-Dev.json | 13 +- ...1.5-Diffusers-480p_i2v_step_distilled.json | 13 +- ...__HunyuanVideo-1.5-Diffusers-480p_t2v.json | 13 +- ...ideo-1.5-Diffusers-720p_i2v_distilled.json | 13 +- ...__HunyuanVideo-1.5-Diffusers-720p_t2v.json | 13 +- .../hunyuanvideo-community__HunyuanVideo.json | 13 +- ...__Kandinsky-5.0-I2V-Lite-5s-Diffusers.json | 13 +- ...ky-5.0-I2V-Pro-distilled-5s-Diffusers.json | 13 +- ...andinsky-5.0-I2V-Pro-sft-5s-Diffusers.json | 13 +- ...2V-Lite-distilled16steps-5s-Diffusers.json | 13 +- ...ndinsky-5.0-T2V-Lite-sft-5s-Diffusers.json | 13 +- ...ky-5.0-T2V-Pro-distilled-5s-Diffusers.json | 13 +- ...andinsky-5.0-T2V-Pro-sft-5s-Diffusers.json | 13 +- ...ashid__TurboWan2.1-T2V-1.3B-Diffusers.json | 13 +- ...rashid__TurboWan2.1-T2V-14B-Diffusers.json | 13 +- ...ashid__TurboWan2.2-I2V-A14B-Diffusers.json | 13 +- .../mignonjia__mg_bidirectional_zelda.json | 13 +- .../registry/mignonjia__mg_causal_zelda.json | 13 +- ...nonjia__mg_longtuning_distilled_zelda.json | 13 +- .../mignonjia__mg_sf_distilled_zelda.json | 13 +- ...onjia__mg_sf_distilled_zelda_1k_steps.json | 13 +- ...vidia__Cosmos-Predict2-2B-Video2World.json | 13 +- .../nvidia__Cosmos-Predict2.5-14B.json | 13 +- ...rand0nmr__SFWan2.2-T2V-A14B-Diffusers.json | 13 +- ...ant__lingbot-world-v2-14b-causal-fast.json | 13 +- ...bilityai__stable-diffusion-3.5-medium.json | 13 +- ..._HunyuanVideo-1.5-Diffusers-1080p-2SR.json | 13 +- ...u03__HunyuanVideo-1.5-Diffusers-1080p.json | 13 +- ...zhou03__Wan2.1-Fun-1.3B-InP-Diffusers.json | 13 +- ...wlsaidhi__SFWan2.1-T2V-1.3B-Diffusers.json | 13 +- .../registry/zai-org__GLM-Image.json | 13 +- ...e__serve_configs__streaming_demo.yaml.json | 1 + ...rence__ltx2_3__i2v_5s2_1024x1536.yaml.json | 1 + ...erence__ltx2_3__i2v_5s2_1280x832.yaml.json | 1 + ...ference__ltx2_3__i2v_5s2_512x768.yaml.json | 1 + ...erence__ltx2_3__i2v_5s2_768x1280.yaml.json | 1 + ...rence__ltx2_3__i2v_8s3_1024x1536.yaml.json | 1 + ...erence__ltx2_3__i2v_8s3_1280x832.yaml.json | 1 + ...ference__ltx2_3__i2v_8s3_512x768.yaml.json | 1 + ...erence__ltx2_3__i2v_8s3_768x1280.yaml.json | 1 + ...rence__ltx2_3__t2v_5s2_1024x1536.yaml.json | 1 + ...erence__ltx2_3__t2v_5s2_1280x832.yaml.json | 1 + ...ference__ltx2_3__t2v_5s2_512x768.yaml.json | 1 + ...erence__ltx2_3__t2v_5s2_768x1280.yaml.json | 1 + ...rence__ltx2_3__t2v_8s3_1024x1536.yaml.json | 1 + ...erence__ltx2_3__t2v_8s3_1280x832.yaml.json | 1 + ...ference__ltx2_3__t2v_8s3_512x768.yaml.json | 1 + ...erence__ltx2_3__t2v_8s3_768x1280.yaml.json | 1 + ...examples__serving__openai_fasth3.yaml.json | 13 +- ...es__serving__openai_fasth3_8step.yaml.json | 13 +- ...les__serving__openai_fasth3_lora.yaml.json | 13 +- ...__serving__openai_fastwan21_1_3b.yaml.json | 13 +- ...s__serving__openai_wan21_i2v_14b.yaml.json | 13 +- ...__serving__openai_wan22_t2v_a14b.yaml.json | 13 +- ...s__serving__openai_wan22_ti2v_5b.yaml.json | 13 +- ...xamples__serving__streaming_demo.yaml.json | 1 + ...inference__inference_fasthunyuan.yaml.json | 13 +- ...inference__inference_flux2_klein.yaml.json | 13 +- ...ts__inference__inference_hunyuan.yaml.json | 13 +- ...ts__inference__inference_longcat.yaml.json | 13 +- ...rence__inference_longcat_distill.yaml.json | 13 +- ...inference__inference_longcat_i2v.yaml.json | 13 +- ...ference_longcat_refine_fromvideo.yaml.json | 13 +- ..._inference__inference_longcat_vc.yaml.json | 13 +- ...cripts__inference__inference_wan.yaml.json | 13 +- ...erence__inference_wan_1.3B_VMoba.yaml.json | 13 +- ...ts__inference__inference_wan_VSA.yaml.json | 13 +- ...__inference_wan_VSA_DMD_14B_480P.yaml.json | 13 +- ...__inference_wan_VSA_DMD_14B_720P.yaml.json | 13 +- ...ence__inference_wan_VSA_DMD_1_3B.yaml.json | 13 +- ...e__inference_wan_VSA_DMD_5B_720P.yaml.json | 13 +- ...ts__inference__inference_wan_i2v.yaml.json | 13 +- .../tests/api/test_compat_translation.py | 42 +--- .../api/test_inference_resolution_steps.py | 7 + .../api/test_ltx2_gpu_pool_translation.py | 26 ++- fastvideo/tests/api/test_parser.py | 1 + .../tests/api/test_resolved_runtime_config.py | 172 ++++++++------ fastvideo/tests/contract/test_env_policy.py | 1 - .../tests/entrypoints/test_video_generator.py | 34 ++- ...est_text_encoder_unified_memory_offload.py | 99 ++++---- .../loader/test_transformer_loader_config.py | 20 ++ .../test_direct_pipeline_unified_memory.py | 24 +- .../test_ltx2_refine_checkpoint_defaults.py | 110 ++++++--- .../platforms/test_unified_memory_offload.py | 193 ++++++++-------- .../tests/ssim/test_gamecraft_similarity.py | 40 ++-- .../tests/ssim/test_matrixgame2_similarity.py | 22 +- .../tests/ssim/test_matrixgame3_similarity.py | 22 +- .../test_ltx2_refine_upsampler_resolve.py | 2 +- .../test_minimax_h3_distilled_schedule.py | 64 ++++-- .../test_minimax_h3_sequential_start.py | 3 +- fastvideo/tests/worker/test_gpu_worker.py | 32 +-- fastvideo/train/callbacks/validation.py | 19 +- fastvideo/train/utils/moduleloader.py | 15 +- fastvideo/training/distillation_pipeline.py | 41 +++- ...ame2_self_forcing_distillation_pipeline.py | 11 +- fastvideo/worker/gpu_worker.py | 11 +- tests/local_tests/flux2/README.md | 10 +- .../test_gamecraft_pipeline_parity.py | 36 ++- .../gen3c/test_gen3c_pipeline_smoke.py | 16 +- .../glm_image/test_edit_pipeline_parity.py | 10 +- .../glm_image/test_pipeline_parity.py | 10 +- .../test_magi_human_pipeline_smoke.py | 16 +- .../test_lingbot_video_pipeline_smoke.py | 28 ++- .../pipelines/test_zimage_pipeline_parity.py | 30 ++- .../pipelines/test_zimage_pipeline_smoke.py | 30 ++- .../test_stable_audio_a2a_parity.py | 16 +- .../test_stable_audio_inpaint_parity.py | 16 +- .../test_stable_audio_pipeline_parity.py | 16 +- 261 files changed, 2955 insertions(+), 1805 deletions(-) create mode 100644 fastvideo/api/checkpoint_defaults.py rename fastvideo/tests/api/config_snapshot_goldens/{kwargs => config}/disable_autocast.json (97%) rename fastvideo/tests/api/config_snapshot_goldens/{kwargs => config}/torch_compile_kwargs_typed_and_extra.json (97%) rename fastvideo/tests/api/config_snapshot_goldens/{kwargs => config}/typed_offload_and_parallelism.json (97%) diff --git a/.agents/skills/add-model-09-pipeline/templates/pipeline_parity_test.py b/.agents/skills/add-model-09-pipeline/templates/pipeline_parity_test.py index 34c260923..100fbe861 100644 --- a/.agents/skills/add-model-09-pipeline/templates/pipeline_parity_test.py +++ b/.agents/skills/add-model-09-pipeline/templates/pipeline_parity_test.py @@ -80,11 +80,17 @@ def _run_fastvideo_pipeline(model_path: Path, params: dict[str, Any]) -> Any: generator = VideoGenerator.from_pretrained( str(model_path), - num_gpus=1, - use_fsdp_inference=False, - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "dit": False, + "vae": False, + "text_encoder": False, + }, + }, + }, ) try: return generator.generate({ diff --git a/README.md b/README.md index a580fa7d1..f9590bc87 100644 --- a/README.md +++ b/README.md @@ -135,7 +135,7 @@ def main(): # Create a video generator with a pre-trained model generator = VideoGenerator.from_pretrained( "FastVideo/FastWan2.1-T2V-1.3B-Diffusers", - num_gpus=1, # Adjust based on your hardware + {"engine": {"num_gpus": 1}}, # Adjust based on your hardware ) # Define a prompt for your video diff --git a/apps/dreamverse/dreamverse/ltx2_generation.py b/apps/dreamverse/dreamverse/ltx2_generation.py index ea70a11f5..91a7bd942 100644 --- a/apps/dreamverse/dreamverse/ltx2_generation.py +++ b/apps/dreamverse/dreamverse/ltx2_generation.py @@ -310,7 +310,7 @@ class LTX2GenerationBackend: ), ) - self.generator = VideoGenerator.from_pretrained(config=generator_config) + self.generator = VideoGenerator.from_config(generator_config) print(f"[GPU {self.gpu_id}] After model load: {self._gpu_mem()}") lora_stack = DREAMVERSE_LORA_STACK or ([(DREAMVERSE_LORA_PATH, diff --git a/docs/contributing/coding_agents.md b/docs/contributing/coding_agents.md index 450f52d4b..257e3ca86 100644 --- a/docs/contributing/coding_agents.md +++ b/docs/contributing/coding_agents.md @@ -59,7 +59,7 @@ Minimal usage example (based on `examples/inference/basic/basic.py`): from fastvideo import VideoGenerator model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers" # or official_weights// -generator = VideoGenerator.from_pretrained(model_id, num_gpus=1) +generator = VideoGenerator.from_pretrained(model_id, {"engine": {"num_gpus": 1}}) video = generator.generate({ "prompt": "A vibrant city street at sunset.", diff --git a/docs/design/inference_schema_parity_inventory.yaml b/docs/design/inference_schema_parity_inventory.yaml index 2427f17cf..7c3391552 100644 --- a/docs/design/inference_schema_parity_inventory.yaml +++ b/docs/design/inference_schema_parity_inventory.yaml @@ -2,7 +2,6 @@ status_definitions: kept: "Public field remains on a public adapter surface with the same meaning." moved: "Public field remains supported but normalizes into a different nested path." preset_owned: "Public field remains supported only through a model/preset-specific surface." - from_pretrained_keyword: "Typed config field that an argument of VideoGenerator.from_pretrained also sets; the entry value names that argument." compatibility_only: "Public field that an adapter or an open mapping accepts outside the typed fields of the canonical schema." private_only: "Field should only be handled by private adapters and is not a public FastVideo compatibility promise." internal_only: "Field is runtime/config plumbing that model code, config resolution, or the runtime fills; it is not a public input." @@ -10,37 +9,37 @@ status_definitions: surfaces: generator_config: - from_pretrained_keyword: - model_path: "model_path (positional argument)" - revision: revision - trust_remote_code: trust_remote_code - engine.num_gpus: num_gpus - engine.execution_backend: distributed_executor_backend - engine.parallelism.tp_size: tp_size - engine.parallelism.sp_size: sp_size - engine.parallelism.hsdp_replicate_dim: hsdp_replicate_dim - engine.parallelism.hsdp_shard_dim: hsdp_shard_dim - engine.parallelism.dist_timeout: dist_timeout - engine.offload.dit: dit_cpu_offload - engine.offload.dit_layerwise: dit_layerwise_offload - engine.offload.text_encoder: text_encoder_cpu_offload - engine.offload.image_encoder: image_encoder_cpu_offload - engine.offload.vae: vae_cpu_offload - engine.offload.pin_cpu_memory: pin_cpu_memory - engine.compile.enabled: enable_torch_compile - engine.compile.backend: "torch_compile_kwargs['backend']" - engine.compile.fullgraph: "torch_compile_kwargs['fullgraph']" - engine.compile.mode: "torch_compile_kwargs['mode']" - engine.compile.dynamic: "torch_compile_kwargs['dynamic']" - engine.compile.extras: "torch_compile_kwargs keys other than backend, fullgraph, mode, and dynamic" - engine.enable_stage_verification: enable_stage_verification - engine.use_fsdp_inference: use_fsdp_inference - engine.disable_autocast: disable_autocast - pipeline.components.lora_path: lora_path - pipeline.components.lora_strength: lora_strength - pipeline.output_type: output_type kept: + - model_path - mode + - revision + - trust_remote_code + - engine.num_gpus + - engine.execution_backend + - engine.parallelism.tp_size + - engine.parallelism.sp_size + - engine.parallelism.hsdp_replicate_dim + - engine.parallelism.hsdp_shard_dim + - engine.parallelism.dist_timeout + - engine.offload.dit + - engine.offload.dit_layerwise + - engine.offload.text_encoder + - engine.offload.image_encoder + - engine.offload.vae + - engine.offload.pin_cpu_memory + - engine.compile.enabled + - engine.compile.backend + - engine.compile.fullgraph + - engine.compile.mode + - engine.compile.dynamic + - engine.compile.extras + - engine.enable_stage_verification + - engine.use_fsdp_inference + - engine.disable_autocast + - engine.attention.nvfp4_fa4 + - pipeline.components.lora_path + - pipeline.components.lora_strength + - pipeline.output_type - engine.parallelism.master_port - engine.offload.lazy_module_load - engine.compile.text_encoder_enabled diff --git a/docs/design/overview.md b/docs/design/overview.md index 8a88475f8..a1d2b5af2 100644 --- a/docs/design/overview.md +++ b/docs/design/overview.md @@ -30,7 +30,7 @@ Minimal usage (from `examples/inference/basic/basic.py`): from fastvideo import VideoGenerator model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers" # or official_weights// -generator = VideoGenerator.from_pretrained(model_id, num_gpus=1) +generator = VideoGenerator.from_pretrained(model_id, {"engine": {"num_gpus": 1}}) video = generator.generate({ "prompt": "A vibrant city street at sunset.", diff --git a/docs/design/server_contracts/dynamo.md b/docs/design/server_contracts/dynamo.md index c163cab2e..5600b0fd4 100644 --- a/docs/design/server_contracts/dynamo.md +++ b/docs/design/server_contracts/dynamo.md @@ -32,7 +32,7 @@ from fastvideo.api import ( | Surface | Availability | Notes | | --- | --- | --- | -| `VideoGenerator.from_pretrained(model_path, **typed_kwargs)` | Today | `typed_kwargs` is a stable subset from `GeneratorConfig` — no flat legacy LTX-2 kwargs (guaranteed after PR 6) | +| `VideoGenerator.from_pretrained(model_path, config)` | Today | `config` is a nested `GeneratorConfig` mapping without `model_path`; no flat keywords | | `VideoGenerator.generate(request: GenerationRequest) -> GenerationResult` | Today | Aggregated; Dynamo wraps in `asyncio.to_thread` under `asyncio.Lock` | | `VideoGenerator.generate_async(request) -> AsyncGenerator[VideoEvent, None]` | **PR 7.10** | Canonical execution substrate; sync wrapper reroutes through this | | `VideoGenerator.default_health_check_request() -> GenerationRequest` | **PR 7.10** | 256x256 / 8 frames / 1 step; lets Dynamo build its health payload without knowing any FastVideo internals | @@ -225,7 +225,7 @@ async def init_video_generation(runtime, config, shutdown_endpoints): from fastvideo.api import config_to_dict server_args, dynamo_args = config.server_args, config.dynamo_args - generator = VideoGenerator.from_pretrained(**config.fastvideo_kwargs()) + generator = VideoGenerator.from_config(build_generator_config(server_args)) dump_config(dynamo_args.dump_config_to, config) @@ -309,11 +309,11 @@ re-chase FastVideo drift: 2. `ContinuationState.payload` is JSON-serializable or references opaque blob ids. Dynamo can round-trip it through RPC without special-casing torch tensors. -3. `VideoGenerator.from_pretrained` accepts a typed `GeneratorConfig`. - Its only flat keywords are the convenience keywords in - `fastvideo.api.compat.FROM_PRETRAINED_KWARGS`; any other keyword - raises `TypeError` that points to `VideoGenerator.from_config(...)`, - which takes every setting at its typed config path. +3. `VideoGenerator.from_pretrained(model_path, config)` takes a typed + `GeneratorConfig` or its nested mapping; any flat keyword raises + `TypeError` that points to the nested config. + `VideoGenerator.from_config(...)` takes the same settings with + `model_path` inside. 4. `generate_async` (PR 7.10+) emits events in order `Progress* → Partial* → Final`; the final event always has exactly one occurrence per request. diff --git a/docs/getting_started/installation/spark_pair.md b/docs/getting_started/installation/spark_pair.md index 1d63aa1f1..d4d8b1c7a 100644 --- a/docs/getting_started/installation/spark_pair.md +++ b/docs/getting_started/installation/spark_pair.md @@ -17,7 +17,7 @@ xDiT vendor. Do not install xDiT for this path. | Goal | How | Use two Sparks? | |---|---|---| -| Two independent videos at once | One process per box, `num_gpus=1` | Throughput only. Each clip still takes the 1-GPU time for that size. | +| Two independent videos at once | One process per box, `engine.num_gpus: 1` | Throughput only. Each clip still takes the 1-GPU time for that size. | | One clip, faster | Ray + `sp_size=2` + parallel VAE | **Yes.** One 768×1344×124 recipe was 292 s vs 374 s on one GB10. | | One clip, longer | Same, more frames | **Yes.** 345 frames (~14.4 s at 24 fps) finished in 587 s at 768×1344. | @@ -35,7 +35,7 @@ over ~21 GB/s RoCE. QSFP; do not download 100+ GB twice over Wi-Fi. - Ray in the FastVideo venv (`uv pip install ray` if it is not already there). -Each Spark has **one** GPU. `num_gpus=2` therefore means two nodes, which is +Each Spark has **one** GPU. `engine.num_gpus: 2` therefore means two nodes, which is why the executor must be Ray (`mp` only works inside one process tree). ## 1. Put IPv4 on the QSFP NICs @@ -198,7 +198,7 @@ sm_100a VSA kernel is not on this chip, so denoise is slower than a GB200 | Gloo `connectFullMesh` / `remote=[127.0.0.1]` | Two 1-GPU nodes must not use loopback as the Gloo store. Source `spark_pair_env.sh` so `GLOO_SOCKET_IFNAME` is the QSFP NIC on **each** box. FastVideo no longer copies that NIC name from the driver onto workers. | | Second `generate()` crashes `NoneType.parameters` | Sequential load used to drop the text encoder without reloading it. This branch reloads Qwen for later requests so `--warmup --repeats N` works. | | OOM / `earlyoom` prefers Python | Lazy module load must stay on (do not pass `--no-lazy-module-load` to `basic_fasth3.py` or set `engine.offload.lazy_module_load: false`). Peak GPU during 345-frame denoise is ~90 GiB/node. | -| `num_gpus=2` on one Spark | Each Spark has one GPU. Use Ray across two nodes, or `num_gpus=1` on one box. | +| `engine.num_gpus: 2` on one Spark | Each Spark has one GPU. Use Ray across two nodes, or `engine.num_gpus: 1` on one box. | ## What we are not claiming diff --git a/docs/inference/architecture.md b/docs/inference/architecture.md index e22913cda..84783952e 100644 --- a/docs/inference/architecture.md +++ b/docs/inference/architecture.md @@ -376,8 +376,8 @@ modules are skipped. ```python generator = VideoGenerator.from_pretrained( - model_path="Wan-AI/Wan2.1-T2V-14B-Diffusers", - num_gpus=1, tp_size=1, sp_size=1, + "Wan-AI/Wan2.1-T2V-14B-Diffusers", + {"engine": {"num_gpus": 1, "parallelism": {"tp_size": 1, "sp_size": 1}}}, ) result = generator.generate({ "prompt": "A cat dancing", @@ -414,7 +414,7 @@ registry, applies a JSON config if provided, and freezes it. ## End-to-End Inference Flow ``` -User: VideoGenerator.from_pretrained(model_path, **kwargs) +User: VideoGenerator.from_pretrained(model_path, config) │ ├─ resolve_inference_config() → PipelineConfig resolved via registry ├─ get_model_info() → ModelInfo(pipeline_cls, sampling_param_cls, ...) diff --git a/docs/inference/configuration.md b/docs/inference/configuration.md index 767bb7d99..1f4b2b6a6 100644 --- a/docs/inference/configuration.md +++ b/docs/inference/configuration.md @@ -8,7 +8,7 @@ FastVideo automatically distributes the generation process when multiple GPUs ar # Will use 4 GPUs in parallel for faster generation generator = VideoGenerator.from_pretrained( "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", - num_gpus=4, + {"engine": {"num_gpus": 4}}, ) ``` @@ -30,10 +30,10 @@ bring-up: [Pair two NVIDIA DGX Sparks](../getting_started/installation/spark_pai ## Customizing Generation -`VideoGenerator.from_pretrained` accepts only common engine and offload -keywords, such as `num_gpus` and `dit_cpu_offload`. The full list is -`fastvideo.api.compat.FROM_PRETRAINED_KWARGS`. Pass every other startup -setting to `VideoGenerator.from_config` at its config path, and pass +`VideoGenerator.from_pretrained(model_path, config)` takes the startup +settings as a nested mapping at their typed config paths, such as +`{"engine": {"num_gpus": 2, "offload": {"dit": False}}}`; it is +`VideoGenerator.from_config` with `model_path` added to the mapping. Pass generation settings to `VideoGenerator.generate` as a request: ```python @@ -137,8 +137,8 @@ result.resolved_request.provenance("sampling.num_frames") # PathProvenance(..., value=81, source='fill_sampling_defaults[preset wan_t2v_1_3b]', explicit=False) ``` -`resolved_config.provenance_table()` lists every path. After startup, `resolved_config` is read-only; a later -change is a recorded override (`resolved_config.with_override(source, values)`) that returns a new object. +`resolved_config.provenance_table()` lists every path. Every value is decided before resolution ends, including the +device offload policy and the checkpoint defaults; after that, `resolved_config` is read-only. ## Performance Optimization diff --git a/docs/inference/inference_quick_start.md b/docs/inference/inference_quick_start.md index 3224408f9..8e9ab03d2 100644 --- a/docs/inference/inference_quick_start.md +++ b/docs/inference/inference_quick_start.md @@ -38,7 +38,7 @@ def main(): # Create a video generator with a pre-trained model generator = VideoGenerator.from_pretrained( "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", - num_gpus=1, # Adjust based on your hardware + {"engine": {"num_gpus": 1}}, # Adjust based on your hardware ) # Define a prompt for your video @@ -82,7 +82,7 @@ from fastvideo import VideoGenerator def main(): # Create the generator model_name = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers" - generator = VideoGenerator.from_pretrained(model_name, num_gpus=1) + generator = VideoGenerator.from_pretrained(model_name, {"engine": {"num_gpus": 1}}) # Generate video based on the image prompt = "A photograph coming to life with gentle movement" @@ -109,12 +109,12 @@ Common issues and their solutions: If you encounter CUDA out of memory errors: - Reduce `num_frames` or video resolution -- Enable FastVideo offloading options such as `dit_layerwise_offload=True` - (single GPU) or `use_fsdp_inference=True` (multi-GPU) +- Enable FastVideo offloading options such as `engine.offload.dit_layerwise: true` + (single GPU) or `engine.use_fsdp_inference: true` (multi-GPU) - Try a smaller model or use distilled versions -- Use `num_gpus` > 1 if multiple GPUs are available -- Try enabling FSDP inference with `use_fsdp_inference=True` (may slow down generation) -- Try enabling DiT layerwise offload with `dit_layerwise_offload=True` (now only a few models support this, but may introduce less overhead than FSDP) +- Use `engine.num_gpus` > 1 if multiple GPUs are available +- Try enabling FSDP inference with `engine.use_fsdp_inference: true` (may slow down generation) +- Try enabling DiT layerwise offload with `engine.offload.dit_layerwise: true` (now only a few models support this, but may introduce less overhead than FSDP) ### Slow Generation diff --git a/docs/inference/offloading.md b/docs/inference/offloading.md index af4035f5d..f57785fab 100644 --- a/docs/inference/offloading.md +++ b/docs/inference/offloading.md @@ -208,15 +208,21 @@ from fastvideo import VideoGenerator generator = VideoGenerator.from_pretrained( "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", - num_gpus=1, - # Recommended for single GPU - dit_layerwise_offload=True, - # Enable if OOM happens - vae_cpu_offload=True, - image_encoder_cpu_offload=True, - text_encoder_cpu_offload=True, - # Speeds up CPU-GPU transfer - pin_cpu_memory=True, + { + "engine": { + "num_gpus": 1, + "offload": { + # Recommended for single GPU + "dit_layerwise": True, + # Enable if OOM happens + "vae": True, + "image_encoder": True, + "text_encoder": True, + # Speeds up CPU-GPU transfer + "pin_cpu_memory": True, + }, + }, + }, ) prompt = "A curious raccoon peers through a vibrant field of yellow sunflowers." @@ -230,16 +236,22 @@ from fastvideo import VideoGenerator generator = VideoGenerator.from_pretrained( "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", - num_gpus=2, - # Recommended for multi-GPU - use_fsdp_inference=True, - dit_layerwise_offload=False, - dit_cpu_offload=False, - # Enable if OOM happens - vae_cpu_offload=True, - image_encoder_cpu_offload=True, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, + { + "engine": { + "num_gpus": 2, + # Recommended for multi-GPU + "use_fsdp_inference": True, + "offload": { + "dit_layerwise": False, + "dit": False, + # Enable if OOM happens + "vae": True, + "image_encoder": True, + "text_encoder": True, + "pin_cpu_memory": True, + }, + }, + }, ) prompt = "A majestic lion strides across the golden savanna." diff --git a/docs/inference/optimizations.md b/docs/inference/optimizations.md index 0ae7309e0..91a1da8e4 100644 --- a/docs/inference/optimizations.md +++ b/docs/inference/optimizations.md @@ -166,22 +166,26 @@ Enable FP4 attention via the `--nvfp4_fa4` flag: python examples/inference/optimizations/fp4_attn_wan2_1_1_3b.py --nvfp4_fa4 ``` -Or in Python via the `nvfp4_fa4` kwarg (sets env vars automatically): +Or in Python via the `engine.attention.nvfp4_fa4` field (resolution sets the env vars): ```python from fastvideo import VideoGenerator gen = VideoGenerator.from_pretrained( "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", - nvfp4_fa4=True, - num_gpus=1, - use_fsdp_inference=False, # FSDP is incompatible with FP4 pointer path + { + "engine": { + "attention": {"nvfp4_fa4": True}, + "num_gpus": 1, + "use_fsdp_inference": False, # FSDP is incompatible with FP4 pointer path + }, + }, ) gen.generate(request={"prompt": "A raccoon in sunflowers", "output": {"save_video": True}}) ``` #### Known Limitations -- `use_fsdp_inference=True` is incompatible with the FP4 path (FSDP shards invalidate tensor pointers) +- `engine.use_fsdp_inference: true` is incompatible with the FP4 path (FSDP shards invalidate tensor pointers) - Per-call cosine similarity vs BF16: ~0.99 (slight quantization error accumulates over denoising steps) - Only supports `headdim >= 128` @@ -354,7 +358,7 @@ end-to-end speedup. It is **off by default** and enabled per-run. ```python generator = VideoGenerator.from_pretrained( "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", - enable_torch_compile=True, + {"engine": {"compile": {"enabled": True}}}, ) ``` @@ -405,7 +409,7 @@ compile wins for the DiT. ### What to expect from generic compile The Wan result below measures the existing generic -`enable_torch_compile=True` path. It is useful evidence that compile can help, +`engine.compile.enabled: true` path. It is useful evidence that compile can help, but it is **not** a benchmark or numerical gate for the stricter regional fullgraph path above. @@ -448,12 +452,12 @@ not asserted by any standing SSIM regression here — the SSIM tests in run with `enable_torch_compile` disabled. If you depend on compile output staying close to eager (or your previous compiled run), run an MS-SSIM gate on *your* config, especially when combining -`enable_torch_compile=True` with other numerics-affecting flags +`engine.compile.enabled: true` with other numerics-affecting flags (quantized attention backends, FP4, layerwise offload edge cases). ### Known interactions -- **Layerwise CPU offload** (`dit_layerwise_offload=True`, the default): +- **Layerwise CPU offload** (`engine.offload.dit_layerwise: true`, the default): the offload hook previously caused an implicit graph break once per transformer layer, fragmenting the compiled region. Addressed in hao-ai-lab/FastVideo#1365 — keep that fix to get a clean compiled @@ -468,19 +472,17 @@ MS-SSIM gate on *your* config, especially when combining grad-enabled path remain outside it. Use the default inductor mode shown above unless your exact configuration has its own gate. -Extra `torch.compile` options are passed through `torch_compile_kwargs` -(a dict), accepted by `VideoGenerator.from_pretrained(...)`. In a config -file or a CLI dotted override, set them at `engine.compile.backend`, -`fullgraph`, `mode`, and `dynamic`, and put any other `torch.compile` -kwargs in `engine.compile.extras` (for example +Extra `torch.compile` options live at `engine.compile.backend`, +`fullgraph`, `mode`, and `dynamic`; any other `torch.compile` kwargs go in +`engine.compile.extras`. Set them in the nested config, in a config file, +or as a CLI dotted override (for example `--generator.engine.compile.mode reduce-overhead`). Example (currently **not** recommended — see the CUDA-graphs caveat above): ```python VideoGenerator.from_pretrained( "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", - enable_torch_compile=True, - torch_compile_kwargs={"mode": "reduce-overhead"}, # may error today + {"engine": {"compile": {"enabled": True, "mode": "reduce-overhead"}}}, # may error today ) ``` @@ -493,7 +495,7 @@ config; **discard the first generation** (graph build): import time from fastvideo import VideoGenerator -gen = VideoGenerator.from_pretrained("your-model-id", enable_torch_compile=True) +gen = VideoGenerator.from_pretrained("your-model-id", {"engine": {"compile": {"enabled": True}}}) req = {"prompt": "Your prompt", "sampling": {"seed": 1024}, "output": {"save_video": False}} gen.generate(req) # warmup: graph build, discard diff --git a/examples/inference/basic/README.md b/examples/inference/basic/README.md index 2544fe48f..16f926d2b 100644 --- a/examples/inference/basic/README.md +++ b/examples/inference/basic/README.md @@ -177,7 +177,7 @@ from fastvideo import VideoGenerator def main(): generator = VideoGenerator.from_pretrained( "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", - num_gpus=1, + {"engine": {"num_gpus": 1}}, ) prompt = ("A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " diff --git a/examples/inference/basic/basic.py b/examples/inference/basic/basic.py index 428c38f17..2cae41493 100644 --- a/examples/inference/basic/basic.py +++ b/examples/inference/basic/basic.py @@ -11,13 +11,19 @@ def main(): generator = VideoGenerator.from_pretrained( "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", # FastVideo will automatically handle distributed setup - num_gpus=1, - use_fsdp_inference=False, # set to True if GPU is out of memory - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" - # image_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set to True if GPU is out of memory + "offload": { + "dit": False, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + # "image_encoder": False, + }, + }, + }, ) # Generate videos with the same simple API, regardless of GPU count diff --git a/examples/inference/basic/basic_cosmos2_5_i2w.py b/examples/inference/basic/basic_cosmos2_5_i2w.py index 366f2b862..7b7768a06 100644 --- a/examples/inference/basic/basic_cosmos2_5_i2w.py +++ b/examples/inference/basic/basic_cosmos2_5_i2w.py @@ -8,12 +8,18 @@ def main(): generator = VideoGenerator.from_pretrained( model_path, - num_gpus=1, - use_fsdp_inference=False, # set True if GPU is out of memory - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set True if GPU is out of memory + "offload": { + "dit": False, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, + }, + }, + }, ) # image2world example from official repo diff --git a/examples/inference/basic/basic_cosmos2_5_t2w.py b/examples/inference/basic/basic_cosmos2_5_t2w.py index 84e9b178f..aa522d42e 100644 --- a/examples/inference/basic/basic_cosmos2_5_t2w.py +++ b/examples/inference/basic/basic_cosmos2_5_t2w.py @@ -7,12 +7,18 @@ def main(): generator = VideoGenerator.from_pretrained( model_path, - num_gpus=1, - use_fsdp_inference=False, # set True if GPU is out of memory - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set True if GPU is out of memory + "offload": { + "dit": False, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, + }, + }, + }, ) prompt = ( diff --git a/examples/inference/basic/basic_cosmos2_5_v2w.py b/examples/inference/basic/basic_cosmos2_5_v2w.py index 45a689c3e..1cc21cd01 100644 --- a/examples/inference/basic/basic_cosmos2_5_v2w.py +++ b/examples/inference/basic/basic_cosmos2_5_v2w.py @@ -8,12 +8,18 @@ def main(): generator = VideoGenerator.from_pretrained( model_path, - num_gpus=1, - use_fsdp_inference=False, # set True if GPU is out of memory - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set True if GPU is out of memory + "offload": { + "dit": False, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, + }, + }, + }, ) # video2world example from official repo diff --git a/examples/inference/basic/basic_gamecraft.py b/examples/inference/basic/basic_gamecraft.py index 2782827ea..4a442df1a 100644 --- a/examples/inference/basic/basic_gamecraft.py +++ b/examples/inference/basic/basic_gamecraft.py @@ -56,12 +56,18 @@ def main(): # FastVideo will automatically download weights from HuggingFace generator = VideoGenerator.from_pretrained( MODEL_PATH, - num_gpus=1, - use_fsdp_inference=True, - dit_cpu_offload=True, - vae_cpu_offload=True, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": True, + "offload": { + "dit": True, + "vae": True, + "text_encoder": True, + "pin_cpu_memory": True, + }, + }, + }, ) # Video parameters diff --git a/examples/inference/basic/basic_gen3c.py b/examples/inference/basic/basic_gen3c.py index 05b1ea57b..ee444565d 100644 --- a/examples/inference/basic/basic_gen3c.py +++ b/examples/inference/basic/basic_gen3c.py @@ -60,12 +60,18 @@ def main(): generator = VideoGenerator.from_pretrained( args.model_path, - num_gpus=1, - use_fsdp_inference=False, - dit_cpu_offload=False, - vae_cpu_offload=True, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "dit": False, + "vae": True, + "text_encoder": True, + "pin_cpu_memory": True, + }, + }, + }, ) video = generator.generate({ diff --git a/examples/inference/basic/basic_hy15.py b/examples/inference/basic/basic_hy15.py index eda138126..54ee55eb4 100644 --- a/examples/inference/basic/basic_hy15.py +++ b/examples/inference/basic/basic_hy15.py @@ -13,13 +13,19 @@ def main(): generator = VideoGenerator.from_pretrained( "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v", # FastVideo will automatically handle distributed setup - num_gpus=1, - use_fsdp_inference=False, # set to True if GPU is out of memory - dit_cpu_offload=True, - vae_cpu_offload=True, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" - # image_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set to True if GPU is out of memory + "offload": { + "dit": True, + "vae": True, + "text_encoder": True, + "pin_cpu_memory": True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + # "image_encoder": False, + }, + }, + }, ) prompt = ("A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " diff --git a/examples/inference/basic/basic_hy15_1080p.py b/examples/inference/basic/basic_hy15_1080p.py index 439da4a98..a326d9f2c 100644 --- a/examples/inference/basic/basic_hy15_1080p.py +++ b/examples/inference/basic/basic_hy15_1080p.py @@ -14,13 +14,19 @@ def main(): "weizhou03/HunyuanVideo-1.5-Diffusers-1080p-2SR", # 480p -> 720p -> 1080p # or "weizhou03/HunyuanVideo-1.5-Diffusers-1080p" # 720p -> 1080p # FastVideo will automatically handle distributed setup - num_gpus=1, - use_fsdp_inference=False, # set to True if GPU is out of memory - dit_cpu_offload=True, - vae_cpu_offload=True, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" - # image_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set to True if GPU is out of memory + "offload": { + "dit": True, + "vae": True, + "text_encoder": True, + "pin_cpu_memory": True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + # "image_encoder": False, + }, + }, + }, ) prompt = ("A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " diff --git a/examples/inference/basic/basic_hyworld.py b/examples/inference/basic/basic_hyworld.py index 9a99ed6f5..68e251142 100644 --- a/examples/inference/basic/basic_hyworld.py +++ b/examples/inference/basic/basic_hyworld.py @@ -35,13 +35,19 @@ def main(): print("\nInitializing VideoGenerator for HYWorld...") generator = VideoGenerator.from_pretrained( "FastVideo/HY-WorldPlay-Bidirectional-Diffusers", - num_gpus=1, - use_fsdp_inference=True, - dit_cpu_offload=True, - vae_cpu_offload=True, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, - image_encoder_cpu_offload=True, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": True, + "offload": { + "dit": True, + "vae": True, + "text_encoder": True, + "pin_cpu_memory": True, + "image_encoder": True, + }, + }, + }, ) # Generate video diff --git a/examples/inference/basic/basic_kandinsky5_i2v.py b/examples/inference/basic/basic_kandinsky5_i2v.py index 86579913c..b64625e16 100644 --- a/examples/inference/basic/basic_kandinsky5_i2v.py +++ b/examples/inference/basic/basic_kandinsky5_i2v.py @@ -10,13 +10,19 @@ def main(): "kandinskylab/Kandinsky-5.0-I2V-Pro-distilled-5s-Diffusers", # "kandinskylab/Kandinsky-5.0-I2V-Pro-sft-5s-Diffusers" # "kandinskylab/Kandinsky-5.0-I2V-Lite-5s-Diffusers" - num_gpus=1, - use_fsdp_inference=False, - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, - # image_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "dit": False, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, + # "image_encoder": False, + }, + }, + }, ) prompt = ("A woman stands up and walks away") diff --git a/examples/inference/basic/basic_kandinsky5_t2v.py b/examples/inference/basic/basic_kandinsky5_t2v.py index a4dcc175a..b5d9481f7 100644 --- a/examples/inference/basic/basic_kandinsky5_t2v.py +++ b/examples/inference/basic/basic_kandinsky5_t2v.py @@ -9,13 +9,19 @@ def main(): # "kandinskylab/Kandinsky-5.0-T2V-Pro-sft-5s-Diffusers" # "kandinskylab/Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers" # "kandinskylab/Kandinsky-5.0-T2V-Pro-distilled-5s-Diffusers" - num_gpus=1, - use_fsdp_inference=False, - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, - # image_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "dit": False, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, + # "image_encoder": False, + }, + }, + }, ) prompt = ("A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " diff --git a/examples/inference/basic/basic_lingbot_video.py b/examples/inference/basic/basic_lingbot_video.py index a39fd0b4f..a57d81757 100644 --- a/examples/inference/basic/basic_lingbot_video.py +++ b/examples/inference/basic/basic_lingbot_video.py @@ -29,11 +29,17 @@ def main() -> None: args = parse_args() generator = VideoGenerator.from_pretrained( str(args.model_path), - num_gpus=1, - use_fsdp_inference=False, - text_encoder_cpu_offload=True, - vae_cpu_offload=False, - pin_cpu_memory=True, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "text_encoder": True, + "vae": False, + "pin_cpu_memory": True, + }, + }, + }, ) try: generator.generate({ diff --git a/examples/inference/basic/basic_lingbotworld_base_cam.py b/examples/inference/basic/basic_lingbotworld_base_cam.py index 7f586ab79..766788ad9 100644 --- a/examples/inference/basic/basic_lingbotworld_base_cam.py +++ b/examples/inference/basic/basic_lingbotworld_base_cam.py @@ -13,14 +13,20 @@ def main(): generator = VideoGenerator.from_pretrained( "FastVideo/LingBot-World-Base-Cam-Diffusers", # FastVideo will automatically handle distributed setup - num_gpus=1, - use_fsdp_inference=False, # set to True if GPU is out of memory - dit_cpu_offload=True, # DiT need to be offloaded for MoE - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer - pin_cpu_memory=True, - # image_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set to True if GPU is out of memory + "offload": { + "dit": True, # DiT need to be offloaded for MoE + "vae": False, + "text_encoder": True, + # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer + "pin_cpu_memory": True, + # "image_encoder": False, + }, + }, + }, ) num_frames = 81 diff --git a/examples/inference/basic/basic_ltx2.py b/examples/inference/basic/basic_ltx2.py index b1da6b0ec..fd11cc59f 100644 --- a/examples/inference/basic/basic_ltx2.py +++ b/examples/inference/basic/basic_ltx2.py @@ -17,7 +17,11 @@ def main() -> None: # Uses FastVideo default sampling settings for LTX2 base. generator = VideoGenerator.from_pretrained( "Davids048/LTX2-Base-Diffusers", - num_gpus=1, + { + "engine": { + "num_gpus": 1, + }, + }, ) output_path = "outputs_video/ltx2_basic/output_ltx2_base_t2v_1088_1920_1.1.mp4" diff --git a/examples/inference/basic/basic_ltx2_distilled.py b/examples/inference/basic/basic_ltx2_distilled.py index 817522b14..a6d2bf696 100644 --- a/examples/inference/basic/basic_ltx2_distilled.py +++ b/examples/inference/basic/basic_ltx2_distilled.py @@ -19,7 +19,11 @@ os.environ["FASTVIDEO_ATTENTION_BACKEND"] = "FLASH_ATTN" def main() -> None: generator = VideoGenerator.from_pretrained( "FastVideo/LTX2-Distilled-Diffusers", - num_gpus=4, + { + "engine": { + "num_gpus": 4, + }, + }, ) output_path = "outputs_video/ltx2_basic/output_ltx2_distilled_t2v.mp4" diff --git a/examples/inference/basic/basic_lucy_edit.py b/examples/inference/basic/basic_lucy_edit.py index 6329c4846..42bda0f4c 100644 --- a/examples/inference/basic/basic_lucy_edit.py +++ b/examples/inference/basic/basic_lucy_edit.py @@ -6,12 +6,18 @@ OUTPUT_PATH = "video_samples_lucy_edit" def main(): generator = VideoGenerator.from_pretrained( "decart-ai/Lucy-Edit-Dev", - num_gpus=1, - use_fsdp_inference=False, - dit_cpu_offload=True, - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "dit": True, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, + }, + }, + }, ) prompt = ("Change the apron and blouse to a classic clown costume: satin " diff --git a/examples/inference/basic/basic_matrixgame2.py b/examples/inference/basic/basic_matrixgame2.py index bab669d80..5343611d8 100644 --- a/examples/inference/basic/basic_matrixgame2.py +++ b/examples/inference/basic/basic_matrixgame2.py @@ -51,14 +51,20 @@ def main(): generator = VideoGenerator.from_pretrained( config["model_path"], # FastVideo will automatically handle distributed setup - num_gpus=1, - use_fsdp_inference=False, # set to True if GPU is out of memory - dit_cpu_offload=True, # DiT need to be offloaded for MoE - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer - pin_cpu_memory=True, - # image_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set to True if GPU is out of memory + "offload": { + "dit": True, # DiT need to be offloaded for MoE + "vae": False, + "text_encoder": True, + # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer + "pin_cpu_memory": True, + # "image_encoder": False, + }, + }, + }, ) num_frames = 597 diff --git a/examples/inference/basic/basic_matrixgame2_streaming.py b/examples/inference/basic/basic_matrixgame2_streaming.py index 454dafb29..c795c01b9 100644 --- a/examples/inference/basic/basic_matrixgame2_streaming.py +++ b/examples/inference/basic/basic_matrixgame2_streaming.py @@ -58,14 +58,20 @@ async def main(): generator = StreamingVideoGenerator.from_pretrained( config["model_path"], # FastVideo will automatically handle distributed setup - num_gpus=1, - use_fsdp_inference=False, # set to True if GPU is out of memory - dit_cpu_offload=True, # DiT need to be offloaded for MoE - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer - pin_cpu_memory=True, - # image_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set to True if GPU is out of memory + "offload": { + "dit": True, # DiT need to be offloaded for MoE + "vae": False, + "text_encoder": True, + # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer + "pin_cpu_memory": True, + # "image_encoder": False, + }, + }, + }, ) max_blocks = 50 diff --git a/examples/inference/basic/basic_matrixgame3.py b/examples/inference/basic/basic_matrixgame3.py index e3d8fa2f7..c5c8cdd0d 100644 --- a/examples/inference/basic/basic_matrixgame3.py +++ b/examples/inference/basic/basic_matrixgame3.py @@ -9,12 +9,18 @@ OUTPUT_PATH = "video_samples_matrixgame3" def main(): generator = VideoGenerator.from_pretrained( MODEL_PATH, - num_gpus=1, - use_fsdp_inference=False, - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "dit": False, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, + }, + }, + }, ) generator.generate({ diff --git a/examples/inference/basic/basic_ray.py b/examples/inference/basic/basic_ray.py index 05f2a558f..a05f1a202 100644 --- a/examples/inference/basic/basic_ray.py +++ b/examples/inference/basic/basic_ray.py @@ -13,14 +13,20 @@ def main(): generator = VideoGenerator.from_pretrained( "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", # FastVideo will automatically handle distributed setup - num_gpus=2, - use_fsdp_inference=True, - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" - distributed_executor_backend="ray", - # image_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 2, + "use_fsdp_inference": True, + "offload": { + "dit": False, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + # "image_encoder": False, + }, + "execution_backend": "ray", + }, + }, ) # Generate videos with the same simple API, regardless of GPU count diff --git a/examples/inference/basic/basic_self_forcing_causal.py b/examples/inference/basic/basic_self_forcing_causal.py index 1ea839882..7bbe940c0 100644 --- a/examples/inference/basic/basic_self_forcing_causal.py +++ b/examples/inference/basic/basic_self_forcing_causal.py @@ -14,10 +14,16 @@ def main(): generator = VideoGenerator.from_pretrained( model_name, # FastVideo will automatically handle distributed setup - num_gpus=1, - use_fsdp_inference=False, # set to True if GPU is out of memory - text_encoder_cpu_offload=False, - dit_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set to True if GPU is out of memory + "offload": { + "text_encoder": False, + "dit": False, + }, + }, + }, ) prompt = ("A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " diff --git a/examples/inference/basic/basic_stable_audio.py b/examples/inference/basic/basic_stable_audio.py index 5142c49e9..2e31c0b09 100644 --- a/examples/inference/basic/basic_stable_audio.py +++ b/examples/inference/basic/basic_stable_audio.py @@ -58,7 +58,11 @@ PROMPT = "Lo-fi hip hop instrumental with vinyl crackle and gentle piano." def main() -> None: generator = VideoGenerator.from_pretrained( "FastVideo/stable-audio-open-1.0-Diffusers", - num_gpus=1, + { + "engine": { + "num_gpus": 1, + }, + }, ) output_path = "outputs_audio/stable_audio_basic/output_stable_audio.wav" generator.generate({ diff --git a/examples/inference/basic/basic_stable_audio_a2a.py b/examples/inference/basic/basic_stable_audio_a2a.py index fc2e0096c..c83b36fe7 100644 --- a/examples/inference/basic/basic_stable_audio_a2a.py +++ b/examples/inference/basic/basic_stable_audio_a2a.py @@ -60,7 +60,11 @@ INIT_AUDIO_STRENGTH = 0.6 def main() -> None: generator = VideoGenerator.from_pretrained( "FastVideo/stable-audio-open-1.0-Diffusers", - num_gpus=1, + { + "engine": { + "num_gpus": 1, + }, + }, ) generator.generate({ "prompt": PROMPT, diff --git a/examples/inference/basic/basic_stable_audio_inpaint.py b/examples/inference/basic/basic_stable_audio_inpaint.py index 459192cd1..3f01fc916 100644 --- a/examples/inference/basic/basic_stable_audio_inpaint.py +++ b/examples/inference/basic/basic_stable_audio_inpaint.py @@ -65,7 +65,11 @@ def main() -> None: "m4a/flac) before running.") generator = VideoGenerator.from_pretrained( "FastVideo/stable-audio-open-1.0-Diffusers", - num_gpus=1, + { + "engine": { + "num_gpus": 1, + }, + }, ) generator.generate({ "prompt": PROMPT, diff --git a/examples/inference/basic/basic_stable_audio_small.py b/examples/inference/basic/basic_stable_audio_small.py index 94bbd97f8..dfce9a5d2 100644 --- a/examples/inference/basic/basic_stable_audio_small.py +++ b/examples/inference/basic/basic_stable_audio_small.py @@ -35,7 +35,11 @@ PROMPT = "Lo-fi hip hop instrumental with vinyl crackle and gentle piano." def main() -> None: generator = VideoGenerator.from_pretrained( "FastVideo/stable-audio-open-small-Diffusers", - num_gpus=1, + { + "engine": { + "num_gpus": 1, + }, + }, ) output_path = "outputs_audio/stable_audio_small/output_stable_audio_small.wav" generator.generate({ diff --git a/examples/inference/basic/basic_turbodiffusion.py b/examples/inference/basic/basic_turbodiffusion.py index 608ad91e4..9fbd8ccfe 100644 --- a/examples/inference/basic/basic_turbodiffusion.py +++ b/examples/inference/basic/basic_turbodiffusion.py @@ -14,11 +14,16 @@ def main() -> None: generator = VideoGenerator.from_pretrained( "loayrashid/TurboWan2.1-T2V-1.3B-Diffusers", # FastVideo will automatically handle distributed setup - num_gpus=1, - use_fsdp_inference=False, # set to True if GPU is out of memory - - # set to false if using RTX 4090 - # pin_cpu_memory=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set to True if GPU is out of memory + "offload": { + # set to false if using RTX 4090 + # "pin_cpu_memory": False, + }, + }, + }, ) # Generate videos with the same simple API, regardless of GPU count diff --git a/examples/inference/basic/basic_turbodiffusion_14b.py b/examples/inference/basic/basic_turbodiffusion_14b.py index 0daaaf262..e29d6f9df 100644 --- a/examples/inference/basic/basic_turbodiffusion_14b.py +++ b/examples/inference/basic/basic_turbodiffusion_14b.py @@ -14,7 +14,11 @@ def main() -> None: generator = VideoGenerator.from_pretrained( "loayrashid/TurboWan2.1-T2V-14B-Diffusers", # 14B model needs more GPUs - num_gpus=2, + { + "engine": { + "num_gpus": 2, + }, + }, ) prompt = ("A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " diff --git a/examples/inference/basic/basic_turbodiffusion_i2v.py b/examples/inference/basic/basic_turbodiffusion_i2v.py index 0a874a614..4030e98d9 100644 --- a/examples/inference/basic/basic_turbodiffusion_i2v.py +++ b/examples/inference/basic/basic_turbodiffusion_i2v.py @@ -14,7 +14,11 @@ def main() -> None: # TurboDiffusion I2V: 1-4 step image-to-video generation generator = VideoGenerator.from_pretrained( MODEL_PATH, - num_gpus=2, + { + "engine": { + "num_gpus": 2, + }, + }, ) # Example prompt and image for I2V diff --git a/examples/inference/basic/basic_wan2_2.py b/examples/inference/basic/basic_wan2_2.py index 26366700c..aabe03d7f 100644 --- a/examples/inference/basic/basic_wan2_2.py +++ b/examples/inference/basic/basic_wan2_2.py @@ -11,14 +11,20 @@ def main(): generator = VideoGenerator.from_pretrained( "Wan-AI/Wan2.2-T2V-A14B-Diffusers", # FastVideo will automatically handle distributed setup - num_gpus=2, - use_fsdp_inference=False, # set to True if GPU is out of memory - dit_cpu_offload=True, # DiT need to be offloaded for MoE - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer - pin_cpu_memory=True, - # image_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 2, + "use_fsdp_inference": False, # set to True if GPU is out of memory + "offload": { + "dit": True, # DiT need to be offloaded for MoE + "vae": False, + "text_encoder": True, + # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer + "pin_cpu_memory": True, + # "image_encoder": False, + }, + }, + }, ) # Generate videos with the same simple API, regardless of GPU count diff --git a/examples/inference/basic/basic_wan2_2_Fun.py b/examples/inference/basic/basic_wan2_2_Fun.py index 9185e4766..c2d4fd847 100644 --- a/examples/inference/basic/basic_wan2_2_Fun.py +++ b/examples/inference/basic/basic_wan2_2_Fun.py @@ -15,14 +15,20 @@ def main(): "IRMChen/Wan2.1-Fun-1.3B-Control-Diffusers", # "alibaba-pai/Wan2.2-Fun-A14B-Control", # FastVideo will automatically handle distributed setup - num_gpus=1, - use_fsdp_inference=False, # set to True if GPU is out of memory - dit_cpu_offload=True, # DiT need to be offloaded for MoE - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer - pin_cpu_memory=True, - # image_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set to True if GPU is out of memory + "offload": { + "dit": True, # DiT need to be offloaded for MoE + "vae": False, + "text_encoder": True, + # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer + "pin_cpu_memory": True, + # "image_encoder": False, + }, + }, + }, ) prompt = "一位年轻女性穿着一件粉色的连衣裙，裙子上有白色的装饰和粉色的纽扣。她的头发是紫色的，头上戴着一个红色的大蝴蝶结，显得非常可爱和精致。她还戴着一个红色的领结，整体造型充满了少女感和活力。她的表情温柔，双手轻轻交叉放在身前，姿态优雅。背景是简单的灰色，没有任何多余的装饰，使得人物更加突出。她的妆容清淡自然，突显了她的清新气质。整体画面给人一种甜美、梦幻的感觉，仿佛置身于童话世界中。" diff --git a/examples/inference/basic/basic_wan2_2_i2v.py b/examples/inference/basic/basic_wan2_2_i2v.py index fe7f653ad..39e9538b0 100644 --- a/examples/inference/basic/basic_wan2_2_i2v.py +++ b/examples/inference/basic/basic_wan2_2_i2v.py @@ -13,14 +13,20 @@ def main(): generator = VideoGenerator.from_pretrained( "Wan-AI/Wan2.2-I2V-A14B-Diffusers", # FastVideo will automatically handle distributed setup - num_gpus=1, - use_fsdp_inference=False, # set to True if GPU is out of memory - dit_cpu_offload=True, # DiT need to be offloaded for MoE - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer - pin_cpu_memory=True, - # image_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set to True if GPU is out of memory + "offload": { + "dit": True, # DiT need to be offloaded for MoE + "vae": False, + "text_encoder": True, + # Set pin_cpu_memory to false if CPU RAM is limited and there're no frequent CPU-GPU transfer + "pin_cpu_memory": True, + # "image_encoder": False, + }, + }, + }, ) prompt = "Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside." diff --git a/examples/inference/basic/basic_wan2_2_ti2v.py b/examples/inference/basic/basic_wan2_2_ti2v.py index 40b82eaad..02f459797 100644 --- a/examples/inference/basic/basic_wan2_2_ti2v.py +++ b/examples/inference/basic/basic_wan2_2_ti2v.py @@ -12,13 +12,19 @@ def main(): generator = VideoGenerator.from_pretrained( model_name, # FastVideo will automatically handle distributed setup - num_gpus=1, - use_fsdp_inference=False, # set to True if GPU is out of memory - dit_cpu_offload=True, - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" - # image_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, # set to True if GPU is out of memory + "offload": { + "dit": True, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + # "image_encoder": False, + }, + }, + }, ) # I2V is triggered just by passing in an image_path argument diff --git a/examples/inference/eval/all_metrics_demo.py b/examples/inference/eval/all_metrics_demo.py index d7c7a6c4b..62f4483bc 100644 --- a/examples/inference/eval/all_metrics_demo.py +++ b/examples/inference/eval/all_metrics_demo.py @@ -55,7 +55,11 @@ def generate_one_ltx2_video() -> str: # audio track is silence/noise — unusable for audio.* metrics). generator = VideoGenerator.from_pretrained( "Davids048/LTX2-Base-Diffusers", - num_gpus=1, + { + "engine": { + "num_gpus": 1, + }, + }, ) generator.generate({ "prompt": PROMPT, diff --git a/examples/inference/eval/basic_ltx2_audio_eval.py b/examples/inference/eval/basic_ltx2_audio_eval.py index ec9aacd0e..6bc454d8d 100644 --- a/examples/inference/eval/basic_ltx2_audio_eval.py +++ b/examples/inference/eval/basic_ltx2_audio_eval.py @@ -39,7 +39,11 @@ METRICS = [ def main() -> None: generator = VideoGenerator.from_pretrained( "Davids048/LTX2-Base-Diffusers", - num_gpus=1, + { + "engine": { + "num_gpus": 1, + }, + }, ) output_path = "outputs_video/ltx2_audio_eval/output.mp4" diff --git a/examples/inference/eval/basic_ltx2_eval.py b/examples/inference/eval/basic_ltx2_eval.py index 43cf58774..66c0c5beb 100644 --- a/examples/inference/eval/basic_ltx2_eval.py +++ b/examples/inference/eval/basic_ltx2_eval.py @@ -58,7 +58,11 @@ def main() -> None: # ----- generation (matches examples/inference/basic/basic_ltx2.py) ----- generator = VideoGenerator.from_pretrained( "Davids048/LTX2-Base-Diffusers", - num_gpus=1, + { + "engine": { + "num_gpus": 1, + }, + }, ) output_path = "outputs_video/ltx2_basic/output_ltx2_base_t2v_1088_1920_1.1.mp4" diff --git a/examples/inference/eval/bench_physics_iq.py b/examples/inference/eval/bench_physics_iq.py index 018c198b9..9eb5f4ea9 100644 --- a/examples/inference/eval/bench_physics_iq.py +++ b/examples/inference/eval/bench_physics_iq.py @@ -54,7 +54,14 @@ def _generate_videos(rows: list[dict], videos_dir: Path, model: str, num_gpus: i print(f"[gen] {len(todo)}/{len(rows)} scenarios to render with {model} " f"({num_frames}x{height}x{width})...") - gen = VideoGenerator.from_pretrained(model, num_gpus=num_gpus) + gen = VideoGenerator.from_pretrained( + model, + { + "engine": { + "num_gpus": num_gpus, + }, + }, + ) try: for row, out_path in todo: gen.generate({ diff --git a/examples/inference/eval/bench_vbench.py b/examples/inference/eval/bench_vbench.py index c87095cad..61a2898cf 100644 --- a/examples/inference/eval/bench_vbench.py +++ b/examples/inference/eval/bench_vbench.py @@ -52,7 +52,14 @@ def _generate_videos(prompts: list[str], videos_dir: Path, model: str, num_gpus: print(f"[gen] {len(todo)}/{len(prompts)} prompts to render with {model} " f"({num_frames}x{height}x{width})...") - gen = VideoGenerator.from_pretrained(model, num_gpus=num_gpus) + gen = VideoGenerator.from_pretrained( + model, + { + "engine": { + "num_gpus": num_gpus, + }, + }, + ) try: for prompt, out_path in todo: gen.generate({ diff --git a/examples/inference/eval/eval_ltx2_vbench.py b/examples/inference/eval/eval_ltx2_vbench.py index 0a1af5f76..6d7f43a75 100644 --- a/examples/inference/eval/eval_ltx2_vbench.py +++ b/examples/inference/eval/eval_ltx2_vbench.py @@ -97,7 +97,14 @@ def generate(args: argparse.Namespace) -> Path: out.parent.mkdir(parents=True, exist_ok=True) print(f"[gen] loading {args.model} ({args.num_gpus} GPU)...") - generator = VideoGenerator.from_pretrained(args.model, num_gpus=args.num_gpus) + generator = VideoGenerator.from_pretrained( + args.model, + { + "engine": { + "num_gpus": args.num_gpus, + }, + }, + ) try: print(f"[gen] generating to {out}...") generator.generate({ diff --git a/examples/inference/gradio/local/gradio_local_demo_matrixgame2.py b/examples/inference/gradio/local/gradio_local_demo_matrixgame2.py index f39c1b81b..6676a99b0 100644 --- a/examples/inference/gradio/local/gradio_local_demo_matrixgame2.py +++ b/examples/inference/gradio/local/gradio_local_demo_matrixgame2.py @@ -581,12 +581,18 @@ def main(): setup_model_environment(model_path) generator = StreamingVideoGenerator.from_pretrained( model_path, - num_gpus=1, - use_fsdp_inference=True, - dit_cpu_offload=True, - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": True, + "offload": { + "dit": True, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, + }, + }, + }, ) generators = {model_path: generator} diff --git a/examples/inference/optimizations/attention_example.py b/examples/inference/optimizations/attention_example.py index ea2a23601..d7e466254 100644 --- a/examples/inference/optimizations/attention_example.py +++ b/examples/inference/optimizations/attention_example.py @@ -10,12 +10,18 @@ def main(): start_time = time.perf_counter() gen = VideoGenerator.from_pretrained( - model_path="Wan-AI/Wan2.1-T2V-1.3B-Diffusers", - num_gpus=1, - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", + { + "engine": { + "num_gpus": 1, + "offload": { + "dit": False, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, # set to false if low CPU RAM or hit obscure "CUDA error: Invalid argument" + }, + }, + }, ) load_time = time.perf_counter() - start_time print(f"Model loading time: {load_time:.2f} seconds") diff --git a/examples/inference/optimizations/fp4_attn_wan2_1_1_3b.py b/examples/inference/optimizations/fp4_attn_wan2_1_1_3b.py index 1221e27c1..a3292d563 100644 --- a/examples/inference/optimizations/fp4_attn_wan2_1_1_3b.py +++ b/examples/inference/optimizations/fp4_attn_wan2_1_1_3b.py @@ -38,14 +38,24 @@ def main(): generator = VideoGenerator.from_pretrained( args.model, - num_gpus=args.num_gpus, - nvfp4_fa4=args.nvfp4_fa4, - use_fsdp_inference=not args.nvfp4_fa4, - dit_cpu_offload=False, - dit_layerwise_offload=False, - vae_cpu_offload=True, - text_encoder_cpu_offload=True, - enable_torch_compile=args.compile, + { + "engine": { + "num_gpus": args.num_gpus, + "attention": { + "nvfp4_fa4": args.nvfp4_fa4, + }, + "use_fsdp_inference": not args.nvfp4_fa4, + "offload": { + "dit": False, + "dit_layerwise": False, + "vae": True, + "text_encoder": True, + }, + "compile": { + "enabled": args.compile, + }, + }, + }, ) prompt = ("A curious raccoon peers through a vibrant field of yellow sunflowers, its eyes " diff --git a/examples/inference/optimizations/torch_compile_example.py b/examples/inference/optimizations/torch_compile_example.py index f9c39ce03..a49634e4c 100644 --- a/examples/inference/optimizations/torch_compile_example.py +++ b/examples/inference/optimizations/torch_compile_example.py @@ -1,6 +1,6 @@ """torch.compile A/B example for FastVideo. -`enable_torch_compile=True` compiles the DiT submodules that declare +`engine.compile.enabled: true` compiles the DiT submodules that declare `_compile_conditions` for a substantial end-to-end speedup (e.g. Wan2.1-T2V-1.3B on A100: ~-24% e2e). It is off by default. @@ -35,14 +35,20 @@ def main() -> None: args = parser.parse_args() mode = "COMPILE" if args.compile else "BASELINE" - print(f"Mode: {mode} (enable_torch_compile={args.compile})") + print(f"Mode: {mode} (engine.compile.enabled={args.compile})") os.makedirs("video_samples", exist_ok=True) generator = VideoGenerator.from_pretrained( args.model, - num_gpus=args.num_gpus, - enable_torch_compile=args.compile, + { + "engine": { + "num_gpus": args.num_gpus, + "compile": { + "enabled": args.compile, + }, + }, + }, ) def _run(tag: str) -> float: diff --git a/fastvideo/api/checkpoint_defaults.py b/fastvideo/api/checkpoint_defaults.py new file mode 100644 index 000000000..ada415fba --- /dev/null +++ b/fastvideo/api/checkpoint_defaults.py @@ -0,0 +1,215 @@ +# SPDX-License-Identifier: Apache-2.0 +"""Resolution steps that fill typed fields from files that a checkpoint bundles. + +Distilled LTX-2 checkpoints carry ``fastvideo_refine_*`` defaults in ``model_index.json``, and distilled MiniMax-H3 +checkpoints carry their trained DMD schedule in ``fastvideo_inference.json``. The steps here read those files during +resolution, so the frozen config already holds the checkpoint's values when the pipeline loads. Each field fills only +while it is ``None``: a value from the input always wins. + +A local ``model_path`` is read from disk. A Hub id downloads only the files that the steps need: the manifest through +``fastvideo.registry.maybe_download_model_index`` (the download that the test isolation blocks), and the schedule file +or the component ``config.json`` files through ``huggingface_hub``. When the manifest download fails, the steps leave +their fields unset; the pipeline's own download reports the error. +""" +from __future__ import annotations + +import json +import os +from typing import Any + +from fastvideo.api.resolution import ResolutionStep, ResolutionView +from fastvideo.logger import init_logger +from fastvideo.utils import _split_hf_repo_subfolder, maybe_download_model, verify_model_config_and_directory + +logger = init_logger(__name__) + +# Refine component paths, by the model_index.json key that supplies the checkpoint's default. +LTX2_REFINE_PATH_KEYS = { + "pipeline.ltx2.refine.transformer_path": "fastvideo_refine_transformer_path", + "pipeline.ltx2.refine.lora_path": "fastvideo_refine_lora_path", + "pipeline.ltx2.refine.noise_path": "fastvideo_refine_noise_path", + "pipeline.ltx2.refine.audio_noise_path": "fastvideo_refine_audio_noise_path", +} +# Refine switches, by the model_index.json key that supplies the checkpoint's default and the value's type. +LTX2_REFINE_SWITCH_KEYS = ( + ("pipeline.ltx2.refine.num_inference_steps", "fastvideo_refine_num_inference_steps", int), + ("pipeline.ltx2.refine.guidance_scale", "fastvideo_refine_guidance_scale", float), + ("pipeline.ltx2.refine.add_noise", "fastvideo_refine_add_noise", bool), +) +# Directory names under which published LTX-2 checkpoints bundle the refine upsampler without declaring it. +_UPSAMPLER_DIRECTORY_NAMES = ("spatial_upscaler", "spatial_upsampler") +# The MiniMax-H3 schedule file; the pipeline validates its contents against the loaded schedulers. +H3_SCHEDULE_FILENAME = "fastvideo_inference.json" + + +def _local_manifest(model_path: str) -> dict[str, Any] | None: + """The manifest of a local checkpoint directory, or ``None`` when the directory has no manifest file.""" + manifest_names = ("model_index.json", "modular_model_index.json") + if not any(os.path.isfile(os.path.join(model_path, name)) for name in manifest_names): + return None + return verify_model_config_and_directory(model_path, required_component_dirs=[]) + + +def _hub_manifest(model_path: str, revision: str | None) -> dict[str, Any] | None: + """The manifest of a Hub checkpoint, or ``None`` when its download fails. + + The download goes through ``fastvideo.registry`` so that a test isolation which blocks registry downloads blocks + this one too; the steps then leave their fields unset. + """ + import fastvideo.registry as registry + + try: + return registry.maybe_download_model_index(model_path, revision=revision) + except Exception as error: + logger.debug("Skipping the checkpoint defaults of %s: the manifest is unavailable (%s)", model_path, error) + return None + + +def _checkpoint_manifest(model_path: str, revision: str | None) -> dict[str, Any] | None: + """The checkpoint manifest of ``model_path``, or ``None`` when it cannot be read.""" + if os.path.exists(model_path): + return _local_manifest(model_path) + return _hub_manifest(model_path, revision) + + +def _resolve_refine_path(root: str, value: str | None) -> str | None: + """A refine component path; a relative value that exists inside the checkpoint root becomes that path.""" + if value is None: + return None + if os.path.isabs(value): + return value + candidate = os.path.join(root, value) + if os.path.exists(candidate): + return candidate + return value + + +def _resolve_refine_upsampler_path(root: str, model_index: dict[str, Any]) -> str | None: + """The refine upsampler directory of a checkpoint. + + The ``model_index.json`` keys ``fastvideo_refine_upsampler_path`` and then ``spatial_upsampler`` are the documented + override and win when present. Otherwise the checkpoint root is probed: published LTX-2 distilled checkpoints + bundle the upsampler weights on disk (spelled ``spatial_upscaler``) without declaring them in ``model_index.json``. + """ + path = _resolve_refine_path(root, model_index.get("fastvideo_refine_upsampler_path")) + if path is None and "spatial_upsampler" in model_index: + path = _resolve_refine_path(root, "spatial_upsampler") + if path is None: + for dirname in _UPSAMPLER_DIRECTORY_NAMES: + candidate = os.path.join(root, dirname) + if os.path.isdir(candidate): + return candidate + return path + + +def _ltx2_checkpoint_root(model_path: str, revision: str | None, model_index: dict[str, Any]) -> str: + """The directory that relative refine paths of ``model_index`` resolve against. + + A local ``model_path`` is that directory. For a Hub id, a partial snapshot with every component ``config.json`` + and every relative refine path of the manifest is downloaded; its directory is the one that the pipeline's full + download fills later, so the resolved paths stay valid. + """ + if os.path.exists(model_path): + return model_path + relative_values = [ + value for key in (*LTX2_REFINE_PATH_KEYS.values(), "fastvideo_refine_upsampler_path") + for value in (model_index.get(key), ) if isinstance(value, str) and not os.path.isabs(value) + ] + return maybe_download_model(model_path, revision=revision, allow_patterns=["*/config.json", *relative_values]) + + +def ltx2_refine_checkpoint_step(defaults: Any) -> ResolutionStep: + """Build the step that fills the unset LTX-2 refine fields from the checkpoint's ``model_index.json``. + + ``defaults`` is the model's ``PipelineConfig``; the step decides nothing for a model that is not LTX-2. + ``fastvideo_refine_enabled`` can only turn refine on. The upsampler falls back to a directory of the checkpoint. + """ + + def fill_ltx2_refine_from_checkpoint(view: ResolutionView) -> dict[str, Any]: + from fastvideo.pipelines.basic.ltx2.pipeline_configs import LTX2T2VConfig + + if not isinstance(defaults, LTX2T2VConfig): + return {} + model_path, revision = view.get("model_path"), view.get("revision") + model_index = _checkpoint_manifest(model_path, revision) + if model_index is None: + return {} + root = _ltx2_checkpoint_root(model_path, revision, model_index) + values: dict[str, Any] = {} + if model_index.get("fastvideo_refine_enabled") is True and view.get("pipeline.ltx2.refine.enabled") is None: + values["pipeline.ltx2.refine.enabled"] = True + if view.get("pipeline.components.upsampler_weights") is None: + upsampler_path = _resolve_refine_upsampler_path(root, model_index) + if upsampler_path is not None: + values["pipeline.components.upsampler_weights"] = upsampler_path + for path, key in LTX2_REFINE_PATH_KEYS.items(): + if view.get(path) is None: + component_path = _resolve_refine_path(root, model_index.get(key)) + if component_path is not None: + values[path] = component_path + for path, key, convert in LTX2_REFINE_SWITCH_KEYS: + if view.get(path) is None and model_index.get(key) is not None: + values[path] = convert(model_index[key]) + return values + + fill_ltx2_refine_from_checkpoint.__qualname__ = "fill_ltx2_refine_from_checkpoint" + return fill_ltx2_refine_from_checkpoint + + +def _checkpoint_json(model_path: str, revision: str | None, filename: str) -> Any: + """The parsed JSON file ``filename`` of the checkpoint, or ``None`` when the checkpoint does not have it. + + For a Hub id, the manifest download gates the file download (see :func:`_hub_manifest`). + """ + if os.path.exists(model_path): + path = os.path.join(model_path, filename) + else: + if _hub_manifest(model_path, revision) is None: + return None + from huggingface_hub import hf_hub_download + from huggingface_hub.utils import EntryNotFoundError + + repo_id, subfolder = _split_hf_repo_subfolder(model_path) + try: + path = hf_hub_download(repo_id=repo_id, + filename=f"{subfolder}/{filename}" if subfolder else filename, + revision=revision) + except EntryNotFoundError: + return None + if not os.path.isfile(path): + return None + with open(path, encoding="utf-8") as handle: + return json.load(handle) + + +def dmd_schedule_checkpoint_step(defaults: Any) -> ResolutionStep: + """Build the step that fills an unset ``pipeline.dmd_denoising_steps`` from a MiniMax-H3 checkpoint's + ``fastvideo_inference.json``. + + ``defaults`` is the model's ``PipelineConfig``; the step decides nothing for a model that is not MiniMax-H3. The + step copies the ``dmd_denoising_steps`` list as written; the pipeline validates the file against the loaded + schedulers and raises on a malformed or conflicting schedule. + """ + + def fill_dmd_schedule_from_checkpoint(view: ResolutionView) -> dict[str, Any]: + from fastvideo.configs.pipelines.minimax_h3 import MiniMaxH3PipelineConfig + + if not isinstance(defaults, MiniMaxH3PipelineConfig) or view.get("pipeline.dmd_denoising_steps") is not None: + return {} + contract = _checkpoint_json(view.get("model_path"), view.get("revision"), H3_SCHEDULE_FILENAME) + steps = contract.get("dmd_denoising_steps") if isinstance(contract, dict) else None + if not isinstance(steps, list) or not steps: + return {} + return {"pipeline.dmd_denoising_steps": list(steps)} + + fill_dmd_schedule_from_checkpoint.__qualname__ = "fill_dmd_schedule_from_checkpoint" + return fill_dmd_schedule_from_checkpoint + + +__all__ = [ + "H3_SCHEDULE_FILENAME", + "LTX2_REFINE_PATH_KEYS", + "LTX2_REFINE_SWITCH_KEYS", + "dmd_schedule_checkpoint_step", + "ltx2_refine_checkpoint_step", +] diff --git a/fastvideo/api/compat.py b/fastvideo/api/compat.py index 6a29be034..1447a6771 100644 --- a/fastvideo/api/compat.py +++ b/fastvideo/api/compat.py @@ -1,11 +1,11 @@ # SPDX-License-Identifier: Apache-2.0 from __future__ import annotations -from collections.abc import Iterator, Mapping +from collections.abc import Mapping from copy import deepcopy -from dataclasses import fields, is_dataclass +from dataclasses import is_dataclass from pathlib import Path -from typing import Any, get_args, get_type_hints +from typing import Any from fastvideo.api.overrides import apply_overrides, normalize_overrides from fastvideo.api.parser import config_to_dict, load_raw_config, parse_config @@ -16,7 +16,6 @@ from fastvideo.api.request_metadata import ( reset_tracking_roots, ) from fastvideo.api.schema import ( - FLAT_NAME, ContinuationState, GenerationRequest, GeneratorConfig, @@ -34,36 +33,6 @@ REQUEST_BATCH_EXTRA_PASSTHROUGH_FIELDS = ( "vsa_dense_first_n_steps", "vsa_dense_layers", ) -# The VideoGenerator.from_pretrained keywords besides model_path, config, and log_queue. Every other setting goes -# through VideoGenerator.from_config at its typed path. -FROM_PRETRAINED_KWARGS = frozenset({ - "num_gpus", - "revision", - "trust_remote_code", - "distributed_executor_backend", - "tp_size", - "sp_size", - "hsdp_replicate_dim", - "hsdp_shard_dim", - "dist_timeout", - "use_fsdp_inference", - "disable_autocast", - "enable_stage_verification", - "dit_cpu_offload", - "dit_layerwise_offload", - "text_encoder_cpu_offload", - "image_encoder_cpu_offload", - "vae_cpu_offload", - "pin_cpu_memory", - "enable_torch_compile", - "torch_compile_kwargs", - "lora_path", - "lora_strength", - "output_type", - "nvfp4_fa4", -}) -# torch.compile kwargs that map to first-class CompileConfig fields. -_COMPILE_TYPED_KEYS = ("backend", "fullgraph", "mode", "dynamic") def normalize_generator_config(config: GeneratorConfig | Mapping[str, Any], ) -> GeneratorConfig: @@ -93,38 +62,6 @@ def load_generator_config_from_file( return parse_config(GeneratorConfig, raw) -def from_pretrained_kwargs_to_config( - model_path: str, - kwargs: Mapping[str, Any], -) -> GeneratorConfig: - """Build a ``GeneratorConfig`` from ``VideoGenerator.from_pretrained`` keyword arguments. - - A keyword that a schema field declares as its ``from_pretrained`` keyword sets that field. ``torch_compile_kwargs`` - is split across ``engine.compile``, and the keywords without a field are kept in ``pipeline.experimental``. A - keyword outside ``FROM_PRETRAINED_KWARGS`` raises ``TypeError``. - """ - unsupported = sorted(set(kwargs) - FROM_PRETRAINED_KWARGS) - if unsupported: - raise TypeError(f"VideoGenerator.from_pretrained(...) does not accept {', '.join(unsupported)}. It accepts " - f"only these keywords: {', '.join(sorted(FROM_PRETRAINED_KWARGS))}. Pass every other setting " - "to VideoGenerator.from_config(...) at its config path.") - - raw: dict[str, Any] = {"model_path": model_path} - for key, value in kwargs.items(): - if key == "torch_compile_kwargs": - remaining: dict[str, Any] = (dict(deepcopy(value)) if isinstance(value, Mapping) else {}) - for first_class in _COMPILE_TYPED_KEYS: - if first_class in remaining: - _set_dotted_path(raw, ["engine", "compile", first_class], remaining.pop(first_class)) - if remaining: - _set_dotted_path(raw, ["engine", "compile", "extras"], remaining) - elif key in _FROM_PRETRAINED_FIELDS: - _set_dotted_path(raw, _FROM_PRETRAINED_FIELDS[key].split("."), value) - else: - _set_dotted_path(raw, ["pipeline", "experimental", key], deepcopy(value)) - return parse_config(GeneratorConfig, raw) - - def normalize_generation_request(request: GenerationRequest | Mapping[str, Any], ) -> GenerationRequest: normalized = (request if isinstance(request, GenerationRequest) else parse_config(GenerationRequest, request)) @@ -190,29 +127,6 @@ def _looks_like_run_or_serve_config(raw: Mapping[str, Any]) -> bool: return isinstance(raw.get("generator"), Mapping) -def _schema_fields(config_type: type, prefix: str = "") -> Iterator[tuple[str, Any]]: - """Yield ``(dotted path, field)`` for every field under ``config_type`` that is not a nested config. - - A field whose type is a dataclass, or an optional dataclass, is a nested config and is walked into. - """ - type_hints = get_type_hints(config_type) - for config_field in fields(config_type): - dotted_path = f"{prefix}{config_field.name}" - annotation = type_hints[config_field.name] - nested = [arg for arg in (annotation, *get_args(annotation)) if isinstance(arg, type) and is_dataclass(arg)] - if nested: - yield from _schema_fields(nested[0], f"{dotted_path}.") - else: - yield dotted_path, config_field - - -# ``from_pretrained`` keyword -> dotted path of the GeneratorConfig field that declares it. -_FROM_PRETRAINED_FIELDS: dict[str, str] = { - config_field.metadata[FLAT_NAME]: dotted_path - for dotted_path, config_field in _schema_fields(GeneratorConfig) if FLAT_NAME in config_field.metadata -} - - def request_to_batch_extra(request: GenerationRequest) -> dict[str, Any]: """Extract typed-request extensions consumed through ``ForwardBatch.extra``.""" return { @@ -405,11 +319,9 @@ def _validate_batched_input_length( __all__ = [ - "FROM_PRETRAINED_KWARGS", "REQUEST_BATCH_EXTRA_PASSTHROUGH_FIELDS", "explicit_request_raw", "explicit_request_updates", - "from_pretrained_kwargs_to_config", "load_generator_config_from_file", "normalize_generation_request", "normalize_generator_config", diff --git a/fastvideo/api/device_policy.py b/fastvideo/api/device_policy.py index 9b55f3f60..98bd47755 100644 --- a/fastvideo/api/device_policy.py +++ b/fastvideo/api/device_policy.py @@ -1,22 +1,35 @@ # SPDX-License-Identifier: Apache-2.0 -"""Offload decisions that depend on the device a worker binds, as overrides of a resolved config. +"""Offload decisions that depend on the device class of the local process, as resolution steps. -Resolution runs before any device is chosen, so it cannot tell whether host and device memory are one physical pool. -A worker calls :func:`finalize_device_offload_policy` once after it binds its device and keeps the returned config. -Each decision is a recorded override (``device_policy:``) on a new config object; the input is unchanged. +The steps run in the main process during resolution, so every worker starts from a config whose offload settings are +final. They query the platform for ``LOCAL_DEVICE_ID`` (device 0 of the local process) and assume that every device of +a node has the same memory class: a node mixes no unified-memory and discrete devices. Importing ``fastvideo`` +initializes the CUDA driver already, so the query adds no new initialization; when the platform cannot answer (for +example ``torch.cuda`` is unavailable), the steps decide "not unified". -:func:`offload_disabled_on_unified_memory` answers the same question without overriding, for a loader that chooses a -target device for one component. +``UNIFIED_MEMORY_OFFLOAD_PATHS`` names the offload settings that the unified-memory policy turns off, by the flag name +that loaders and log messages use. + +``APPLY_DEVICE_POLICY`` is the switch of every step here. The test isolation +(``fastvideo/tests/api/config_snapshot.py::isolated_environment``) sets it to ``False`` the same way it blocks +downloads, so the golden snapshots do not depend on the machine that produced them; the steps then decide nothing and +the offload settings keep their input values. """ from __future__ import annotations from typing import Any -from fastvideo.api.resolution import ResolvedGeneratorConfig +from fastvideo.api.resolution import ResolutionStep, ResolutionView +from fastvideo.api.schema import ExecutionMode from fastvideo.logger import init_logger logger = init_logger(__name__) +# The device whose memory class decides the policy: device 0 of the local process (see the module docstring). +LOCAL_DEVICE_ID = 0 +# Whether the device-policy steps decide anything; the test isolation turns it off (see the module docstring). +APPLY_DEVICE_POLICY = True + # Offload settings that trade device memory for host memory, by the flag name that loaders and log messages use. All # of them are a loss on a device whose host and device memory are the same physical pool. UNIFIED_MEMORY_OFFLOAD_PATHS = { @@ -40,107 +53,91 @@ def _device_name(device_id: int) -> str: return current_platform.device_name -def has_unified_memory(device_id: int) -> bool: - """Whether the bound device shares one memory pool with the host. - - CUDA's probe reads runtime device properties and may initialize a CUDA context, so call it only in a process - that owns ``device_id``. - """ +def has_unified_memory(device_id: int = LOCAL_DEVICE_ID) -> bool: + """Whether ``device_id`` shares one memory pool with the host; ``False`` when the platform cannot answer.""" from fastvideo.platforms import current_platform - return bool(current_platform.has_unified_memory(device_id)) - - -def offload_disabled_on_unified_memory(device_id: int, offload_flag: str | None = None) -> bool: - """Whether the unified-memory policy turns off ``offload_flag`` (any offload when ``None``) on ``device_id``.""" - if not has_unified_memory(device_id): + try: + return bool(current_platform.has_unified_memory(device_id)) + except Exception as error: + logger.debug("Treating device %d as discrete: the unified-memory query failed (%s)", device_id, error) return False - return offload_flag is None or offload_flag in UNIFIED_MEMORY_OFFLOAD_PATHS -def disable_offload_on_unified_memory(resolved_config: ResolvedGeneratorConfig, - device_id: int, - *, - unified: bool | None = None) -> ResolvedGeneratorConfig: - """Turn off every enabled host offload when ``device_id`` has unified memory. - - ``unified`` is the probe result when the caller already has it. The decision is the override - ``device_policy:unified_memory``. - """ - if unified is None: - unified = has_unified_memory(device_id) - if not unified: - return resolved_config - enabled = [flag for flag, path in UNIFIED_MEMORY_OFFLOAD_PATHS.items() if _typed_value(resolved_config, path)] +def apply_unified_memory_offload_policy(view: ResolutionView) -> dict[str, Any]: + """Every enabled host offload of ``UNIFIED_MEMORY_OFFLOAD_PATHS`` turns off when the local device has unified + memory: moving weights to the host duplicates them instead of freeing device memory.""" + if not APPLY_DEVICE_POLICY or not has_unified_memory(LOCAL_DEVICE_ID): + return {} + enabled = [flag for flag, path in UNIFIED_MEMORY_OFFLOAD_PATHS.items() if view.get(path)] if not enabled: - return resolved_config - device_name = _device_name(device_id) + return {} + device_name = _device_name(LOCAL_DEVICE_ID) for flag in enabled: logger.info( "Disabling %s: %s has unified memory, so moving weights to the host duplicates them rather than " "freeing device memory.", flag, device_name) - return resolved_config.with_override("device_policy:unified_memory", - {UNIFIED_MEMORY_OFFLOAD_PATHS[flag]: False - for flag in enabled}) + return {UNIFIED_MEMORY_OFFLOAD_PATHS[flag]: False for flag in enabled} -def resolve_device_offload_conflicts(resolved_config: ResolvedGeneratorConfig) -> ResolvedGeneratorConfig: - """Turn off offload modes that cannot run together on this platform. +def fill_lazy_module_load(view: ResolutionView) -> dict[str, Any]: + """An unset ``engine.offload.lazy_module_load`` becomes ``True`` on unified memory for inference and ``False`` + otherwise.""" + if not APPLY_DEVICE_POLICY or view.get("engine.offload.lazy_module_load") is not None: + return {} + training = view.get("mode") in (ExecutionMode.FINETUNING, ExecutionMode.DISTILLATION) + lazy_module_load = not training and has_unified_memory(LOCAL_DEVICE_ID) + if lazy_module_load: + logger.info( + "Enabling lazy_module_load: %s has unified memory, so encoder, DiT, and VAEs cannot stay resident " + "together. Set engine.offload.lazy_module_load to false to keep every component loaded.", + _device_name(LOCAL_DEVICE_ID), + ) + return {"engine.offload.lazy_module_load": lazy_module_load} - On MPS, FSDP inference and layerwise offload turn off. With layerwise offload on, FSDP inference and DiT CPU - offload turn off. - """ + +def apply_mps_offload_policy(view: ResolutionView) -> dict[str, Any]: + """On MPS, FSDP inference and layerwise offload turn off.""" + if not APPLY_DEVICE_POLICY: + return {} from fastvideo.platforms import current_platform - if current_platform.is_mps(): - resolved_config = resolved_config.with_override("device_policy:mps", { - "engine.use_fsdp_inference": False, - "engine.offload.dit_layerwise": False - }) - if _typed_value(resolved_config, "engine.offload.dit_layerwise"): - if _typed_value(resolved_config, "engine.use_fsdp_inference"): - logger.warning("dit_layerwise_offload is enabled, automatically disabling use_fsdp_inference.") - resolved_config = resolved_config.with_override("device_policy:layerwise_offload", - {"engine.use_fsdp_inference": False}) - if _typed_value(resolved_config, "engine.offload.dit"): - logger.warning("dit_layerwise_offload is enabled, automatically disabling dit_cpu_offload.") - resolved_config = resolved_config.with_override("device_policy:layerwise_offload", - {"engine.offload.dit": False}) - return resolved_config + if not current_platform.is_mps(): + return {} + return {"engine.use_fsdp_inference": False, "engine.offload.dit_layerwise": False} -def finalize_device_offload_policy(resolved_config: ResolvedGeneratorConfig, - device_id: int = 0) -> ResolvedGeneratorConfig: - """Apply the device-local memory policy of ``device_id``, then resolve incompatible offload modes. - - On unified memory every host offload turns off. An unset ``engine.offload.lazy_module_load`` becomes ``True`` on - unified memory for inference and ``False`` otherwise. Applying the policy to its own result changes nothing. - """ - unified = has_unified_memory(device_id) - resolved_config = disable_offload_on_unified_memory(resolved_config, device_id, unified=unified) - if _typed_value(resolved_config, "engine.offload.lazy_module_load") is None: - lazy_module_load = unified and not resolved_config.training_mode - resolved_config = resolved_config.with_override("device_policy:lazy_module_load", - {"engine.offload.lazy_module_load": lazy_module_load}) - if lazy_module_load: - logger.info( - "Enabling lazy_module_load: %s has unified memory, so encoder, DiT, and VAEs cannot stay " - "resident together. Set engine.offload.lazy_module_load to false to keep every component loaded.", - _device_name(device_id), - ) - return resolve_device_offload_conflicts(resolved_config) +def apply_layerwise_offload_conflicts(view: ResolutionView) -> dict[str, Any]: + """With layerwise offload on, FSDP inference and DiT CPU offload turn off.""" + if not APPLY_DEVICE_POLICY or not view.get("engine.offload.dit_layerwise"): + return {} + values: dict[str, Any] = {} + if view.get("engine.use_fsdp_inference"): + logger.warning("dit_layerwise_offload is enabled, automatically disabling use_fsdp_inference.") + values["engine.use_fsdp_inference"] = False + if view.get("engine.offload.dit"): + logger.warning("dit_layerwise_offload is enabled, automatically disabling dit_cpu_offload.") + values["engine.offload.dit"] = False + return values -def _typed_value(resolved_config: ResolvedGeneratorConfig, path: str) -> Any: - """The resolved value of a dotted path.""" - return resolved_config.provenance(path).value - +# The device-policy steps in the order that they run: the unified-memory policy first, so the conflict steps see the +# offload settings that survive it. +DEVICE_POLICY_STEPS: tuple[ResolutionStep, ...] = ( + apply_unified_memory_offload_policy, + fill_lazy_module_load, + apply_mps_offload_policy, + apply_layerwise_offload_conflicts, +) __all__ = [ + "APPLY_DEVICE_POLICY", + "DEVICE_POLICY_STEPS", + "LOCAL_DEVICE_ID", "UNIFIED_MEMORY_OFFLOAD_PATHS", - "disable_offload_on_unified_memory", - "finalize_device_offload_policy", + "apply_layerwise_offload_conflicts", + "apply_mps_offload_policy", + "apply_unified_memory_offload_policy", + "fill_lazy_module_load", "has_unified_memory", - "offload_disabled_on_unified_memory", - "resolve_device_offload_conflicts", ] diff --git a/fastvideo/api/inference_resolution.py b/fastvideo/api/inference_resolution.py index d9cc44206..e7c83d00d 100644 --- a/fastvideo/api/inference_resolution.py +++ b/fastvideo/api/inference_resolution.py @@ -11,9 +11,13 @@ precedence over a later one: user input, then environment variables, then model 1. Environment variables fill typed fields. 2. Model defaults from the model's ``PipelineConfig`` fill the typed fields of ``PIPELINE_CONFIG_MIRRORS`` that are still unset. -3. ``pipeline.preset_overrides.refine`` sets the LTX-2 refine fields. +3. ``pipeline.preset_overrides.refine`` sets the LTX-2 refine fields, and the checkpoint's bundled files fill the + LTX-2 refine fields and the MiniMax-H3 DMD schedule that are still unset (``fastvideo.api.checkpoint_defaults``). 4. Derived values replace placeholders and load the files that a path names. -5. Validation steps raise on inconsistent values and decide nothing. +5. The device policy settles the offload settings for the memory class of the local device + (``fastvideo.api.device_policy``). +6. Validation steps raise on inconsistent values and decide nothing. +7. ``fill_runtime_defaults`` gives every field that is still unset its runtime default. ``fastvideo.api.training_schema`` resolves the training and preprocessing roots with the same steps plus their own. """ @@ -27,6 +31,8 @@ from dataclasses import fields, is_dataclass from typing import Any import fastvideo.envs as envs +from fastvideo.api.checkpoint_defaults import dmd_schedule_checkpoint_step, ltx2_refine_checkpoint_step +from fastvideo.api.device_policy import DEVICE_POLICY_STEPS from fastvideo.api.parser import parse_config from fastvideo.api.resolution import (ResolutionStep, ResolutionView, ResolvedGeneratorConfig, resolve_generator_config, thaw) @@ -302,6 +308,18 @@ def warn_deprecated_environment_variables(view: ResolutionView) -> dict[str, Any return {} +def apply_nvfp4_fa4_env(view: ResolutionView) -> dict[str, Any]: + """``engine.attention.nvfp4_fa4`` exports the environment that the NVFP4 FlashAttention-4 path reads. + + Sets ``FASTVIDEO_NVFP4_FA4=1`` and the ``CUTE_DSL_ENABLE_TVM_FFI`` default of the CuTe DSL kernels when the field + is true; decides nothing. + """ + if view.get("engine.attention.nvfp4_fa4"): + envs.FASTVIDEO_NVFP4_FA4.set(True) + envs.setdefault_external("CUTE_DSL_ENABLE_TVM_FFI", "1") + return {} + + def torch_compile_kwargs(resolved_config: ResolvedGeneratorConfig) -> dict[str, Any]: """The keyword arguments for ``torch.compile`` that ``engine.compile`` describes. @@ -317,9 +335,8 @@ def torch_compile_kwargs(resolved_config: ResolvedGeneratorConfig) -> dict[str, return kwargs -# Values that runtime code uses for typed fields that no earlier step decided. ``pipeline.ltx2.refine.*`` and -# ``engine.offload.lazy_module_load`` are not listed: ``None`` there means "decide later" (the LTX-2 checkpoint defaults -# and the device policy). +# Values that runtime code uses for typed fields that no earlier step decided. The LTX-2 refine switches are the +# stage-2 defaults of the LTX-2 pipeline; the checkpoint's model_index.json fills them first when it declares them. RUNTIME_DEFAULTS: dict[str, Any] = { "engine.attention.vsa_sparsity": 0.0, "engine.attention.vsa_tile_size": 256, @@ -335,6 +352,10 @@ RUNTIME_DEFAULTS: dict[str, Any] = { "pipeline.minimax_h3.video_decode_backend": "h3-vae", "pipeline.ltx2.legacy_native_noise_order": False, "pipeline.ltx2.use_distilled_sigmas": True, + "pipeline.ltx2.refine.enabled": False, + "pipeline.ltx2.refine.add_noise": True, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, } @@ -356,6 +377,7 @@ VALIDATION_STEPS: tuple[ResolutionStep, ...] = ( validate_attention_backend, validate_vae_parallel_decode_strategy, warn_deprecated_environment_variables, + apply_nvfp4_fa4_env, ) @@ -368,9 +390,10 @@ def generator_resolution_steps( ) -> tuple[ResolutionStep, ...]: """The resolution steps of any generator root, in the order that they run. - ``before_placeholders`` runs after the refine copy and before ``derive_parallel_sizes``, for checks that need the - -1 placeholders; ``after`` runs last. ``validate_parallel_sizes`` checks the sizes before - ``derive_num_gpus_from_parallel_sizes`` raises ``num_gpus``. + ``before_placeholders`` runs after the checkpoint fills and before ``derive_parallel_sizes``, for checks that need + the -1 placeholders; ``after`` runs last. The device-policy steps run after every fill and derivation and before + the validations. ``validate_parallel_sizes`` checks the sizes before ``derive_num_gpus_from_parallel_sizes`` raises + ``num_gpus``. """ if defaults is None: defaults = build_model_pipeline_config(config) @@ -378,11 +401,14 @@ def generator_resolution_steps( *ENVIRONMENT_STEPS, pipeline_config_defaults_step(config, defaults), copy_refine_preset_overrides, + ltx2_refine_checkpoint_step(defaults), + dmd_schedule_checkpoint_step(defaults), *before_placeholders, derive_parallel_sizes, derive_vae_tiling_from_ltx2_tile_sizes, vae_tiling_default_step(defaults), load_moba_config, + *DEVICE_POLICY_STEPS, *VALIDATION_STEPS, validate_parallel_sizes, derive_num_gpus_from_parallel_sizes, @@ -537,6 +563,7 @@ __all__ = [ "ENVIRONMENT_STEPS", "PIPELINE_CONFIG_MIRRORS", "VALIDATION_STEPS", + "apply_nvfp4_fa4_env", "build_model_pipeline_config", "copy_refine_preset_overrides", "derive_num_gpus_from_parallel_sizes", diff --git a/fastvideo/api/schema.py b/fastvideo/api/schema.py index 64fab5b76..3d40df2c2 100644 --- a/fastvideo/api/schema.py +++ b/fastvideo/api/schema.py @@ -1,20 +1,10 @@ # SPDX-License-Identifier: Apache-2.0 from __future__ import annotations -from dataclasses import MISSING, dataclass, field +from dataclasses import dataclass, field from enum import Enum from typing import Any, Literal -# Field metadata key that holds the ``VideoGenerator.from_pretrained`` keyword that sets a field. -# ``fastvideo.api.compat.from_pretrained_kwargs_to_config`` maps each keyword of ``FROM_PRETRAINED_KWARGS`` to the -# field that declares it. -FLAT_NAME = "flat_name" - - -def flat_field(flat_name: str, default: Any = MISSING, *, default_factory: Any = MISSING) -> Any: - """Declare a dataclass field that the ``VideoGenerator.from_pretrained`` keyword ``flat_name`` sets.""" - return field(default=default, default_factory=default_factory, metadata={FLAT_NAME: flat_name}) - class ExecutionMode(str, Enum): """ @@ -69,23 +59,23 @@ class ServerConfig: @dataclass class ParallelismConfig: - tp_size: int = flat_field("tp_size", -1) - sp_size: int = flat_field("sp_size", -1) - hsdp_replicate_dim: int = flat_field("hsdp_replicate_dim", 1) - hsdp_shard_dim: int = flat_field("hsdp_shard_dim", -1) - dist_timeout: int | None = flat_field("dist_timeout", None) + tp_size: int = -1 + sp_size: int = -1 + hsdp_replicate_dim: int = 1 + hsdp_shard_dim: int = -1 + dist_timeout: int | None = None master_port: int | None = None """Port of the rendezvous that the executor opens for its workers. ``None`` picks an open port.""" @dataclass class OffloadConfig: - dit: bool = flat_field("dit_cpu_offload", True) - dit_layerwise: bool = flat_field("dit_layerwise_offload", True) - text_encoder: bool = flat_field("text_encoder_cpu_offload", True) - image_encoder: bool = flat_field("image_encoder_cpu_offload", True) - vae: bool = flat_field("vae_cpu_offload", True) - pin_cpu_memory: bool = flat_field("pin_cpu_memory", True) + dit: bool = True + dit_layerwise: bool = True + text_encoder: bool = True + image_encoder: bool = True + vae: bool = True + pin_cpu_memory: bool = True # Not a CPU offload: loads each heavy component on first use and frees it # after the last stage that needs it, so peak memory is the largest # overlapping set rather than the sum. Grouped here because it is the same @@ -111,7 +101,7 @@ class CompileConfig: leaving it empty inherits the master kwargs. """ - enabled: bool = flat_field("enable_torch_compile", False) + enabled: bool = False backend: str | None = None fullgraph: bool | None = None mode: str | None = None @@ -144,6 +134,9 @@ class AttentionConfig: """Path to a JSON config for V-MoBA attention.""" moba_config: dict[str, Any] | None = None """V-MoBA attention settings. Resolution loads them from ``moba_config_path`` when that path is set.""" + nvfp4_fa4: bool = False + """FlashAttention-4 with Q and K quantized to NVFP4. Resolution exports ``FASTVIDEO_NVFP4_FA4=1`` and the + ``CUTE_DSL_ENABLE_TVM_FFI`` default that the CuTe DSL kernels need when this is true.""" Precision = Literal["fp32", "fp16", "bf16"] @@ -169,16 +162,16 @@ class QuantizationConfig: @dataclass class EngineConfig: - num_gpus: int = flat_field("num_gpus", 1) - execution_backend: Literal["mp", "ray"] = flat_field("distributed_executor_backend", "mp") + num_gpus: int = 1 + execution_backend: Literal["mp", "ray"] = "mp" parallelism: ParallelismConfig = field(default_factory=ParallelismConfig) offload: OffloadConfig = field(default_factory=OffloadConfig) compile: CompileConfig = field(default_factory=CompileConfig) attention: AttentionConfig = field(default_factory=AttentionConfig) precision: PrecisionConfig = field(default_factory=PrecisionConfig) - enable_stage_verification: bool = flat_field("enable_stage_verification", True) - use_fsdp_inference: bool = flat_field("use_fsdp_inference", False) - disable_autocast: bool = flat_field("disable_autocast", False) + enable_stage_verification: bool = True + use_fsdp_inference: bool = False + disable_autocast: bool = False quantization: QuantizationConfig | None = None @@ -191,9 +184,9 @@ class ComponentConfig: transformer_2_weights: str | None = None vae_weights: str | None = None upsampler_weights: str | None = None - lora_path: str | None = flat_field("lora_path", None) + lora_path: str | None = None lora_nickname: str = "default" - lora_strength: float = flat_field("lora_strength", 1.0) + lora_strength: float = 1.0 lora_target_modules: list[str] | None = None """Module name substrings that restrict LoRA injection, such as ``["q_proj", "v_proj"]``. ``None`` adapts the default modules.""" @@ -205,10 +198,11 @@ class ComponentConfig: class LTX2RefineOptions: """Stage-2 refine settings. ``pipeline.components.upsampler_weights`` holds the refine upsampler. - Resolution copies ``pipeline.preset_overrides.refine`` into the fields of the same name. When the pipeline loads, - the ``fastvideo_refine_*`` defaults of the checkpoint's ``model_index.json`` fill each field that is still ``None`` - (``fastvideo_refine_enabled`` can only turn refine on); ``enabled``, ``add_noise``, ``guidance_scale``, and - ``num_inference_steps`` that are still ``None`` then become ``False``, ``True``, 1.0, and 3. + Resolution copies ``pipeline.preset_overrides.refine`` into the fields of the same name; then the step + ``fill_ltx2_refine_from_checkpoint`` fills each field that is still ``None`` from the ``fastvideo_refine_*`` + defaults of the checkpoint's ``model_index.json`` (``fastvideo_refine_enabled`` can only turn refine on); and + ``fill_runtime_defaults`` gives ``enabled``, ``add_noise``, ``guidance_scale``, and ``num_inference_steps`` that + are still ``None`` the values ``False``, ``True``, 1.0, and 3. """ enabled: bool | None = None @@ -305,7 +299,7 @@ class PipelineSelection: """Timesteps of a few-step distilled (DMD) sampler. ``None`` keeps the model's default.""" boundary_ratio: float | None = None """Mixture-of-experts switch point of a two-transformer model. ``None`` keeps the model's default.""" - output_type: str = flat_field("output_type", "pil") + output_type: str = "pil" """Output of the decoding stage: ``pil`` for decoded frames, ``latent`` to skip the VAE decode.""" dit: dict[str, Any] = field(default_factory=dict) """Overrides for fields of the model's DiT config, such as ``prefix``.""" @@ -323,8 +317,8 @@ class GeneratorConfig: model_path: str mode: ExecutionMode = ExecutionMode.INFERENCE """What the run does: inference, preprocessing, finetuning, or distillation.""" - revision: str | None = flat_field("revision", None) - trust_remote_code: bool = flat_field("trust_remote_code", False) + revision: str | None = None + trust_remote_code: bool = False engine: EngineConfig = field(default_factory=EngineConfig) pipeline: PipelineSelection = field(default_factory=PipelineSelection) diff --git a/fastvideo/api/training_schema.py b/fastvideo/api/training_schema.py index 229a077dd..172caeda2 100644 --- a/fastvideo/api/training_schema.py +++ b/fastvideo/api/training_schema.py @@ -226,27 +226,6 @@ class PreprocessRunConfig(GeneratorConfig): preprocess: PreprocessConfig = field(default_factory=PreprocessConfig) -def resolve_training_offload_conflicts(view: ResolutionView) -> dict[str, Any]: - """Settle offload modes that cannot run together in training, before any device is bound. - - On MPS, FSDP inference and layerwise offload turn off. With layerwise offload on, FSDP inference and DiT CPU - offload turn off. - """ - from fastvideo.platforms import current_platform - - values: dict[str, Any] = {} - if current_platform.is_mps(): - values.update({"engine.use_fsdp_inference": False, "engine.offload.dit_layerwise": False}) - if values.get("engine.offload.dit_layerwise", view.get("engine.offload.dit_layerwise")): - if view.get("engine.use_fsdp_inference"): - logger.warning("dit_layerwise_offload is enabled, automatically disabling use_fsdp_inference.") - values["engine.use_fsdp_inference"] = False - if view.get("engine.offload.dit"): - logger.warning("dit_layerwise_offload is enabled, automatically disabling dit_cpu_offload.") - values["engine.offload.dit"] = False - return values - - def keep_training_dit_on_device(view: ResolutionView) -> dict[str, Any]: """A finetuning or distillation run keeps the DiT on the device, so ``engine.offload.dit`` turns off.""" training = view.get("mode") in (ExecutionMode.FINETUNING, ExecutionMode.DISTILLATION) @@ -297,13 +276,12 @@ def validate_preprocess_options(view: ResolutionView) -> dict[str, Any]: def training_resolution_steps(config: GeneratorConfig, defaults: Any = None) -> tuple[ResolutionStep, ...]: """The resolution steps of a ``TrainingRunConfig``, in the order that they run. - The offload conflicts and the parallel-size check run on the placeholders, before ``derive_parallel_sizes`` - replaces them. + The parallel-size check runs on the placeholders, before ``derive_parallel_sizes`` replaces them. The offload + conflicts are settled by the device-policy steps of the generator list. """ return generator_resolution_steps(config, defaults, - before_placeholders=(resolve_training_offload_conflicts, - validate_training_parallel_sizes), + before_placeholders=(validate_training_parallel_sizes, ), after=(derive_lora_alpha_from_rank, keep_training_dit_on_device)) @@ -391,7 +369,6 @@ __all__ = [ "preprocess_resolution_steps", "resolve_preprocess_config", "resolve_training_config", - "resolve_training_offload_conflicts", "training_resolution_steps", "validate_preprocess_options", "validate_training_parallel_sizes", diff --git a/fastvideo/attention/backends/flash_attn.py b/fastvideo/attention/backends/flash_attn.py index 5ff725a40..5f6f2dd85 100644 --- a/fastvideo/attention/backends/flash_attn.py +++ b/fastvideo/attention/backends/flash_attn.py @@ -36,7 +36,7 @@ logger.info("Worker %s Using FlashAttention-%s backend", local_main_process_only=False) # FP4 FA4 support: quantize Q/K to NVFP4 E2M1 for block-scaled MMA on Blackwell. -# Requires: flash-attention-fp4, flashinfer, cutlass-dsl. Enable via nvfp4_fa4=True kwarg. +# Requires: flash-attention-fp4, flashinfer, cutlass-dsl. Enable via ``engine.attention.nvfp4_fa4``. # The FP4 path uses a dedicated custom_op wrapper (flash_attn_fp4_func) so that # torch.compile treats the CuTeDSL kernel as an opaque boundary. try: diff --git a/fastvideo/configs/pipelines/base.py b/fastvideo/configs/pipelines/base.py index 7f20a8b36..184f4c423 100644 --- a/fastvideo/configs/pipelines/base.py +++ b/fastvideo/configs/pipelines/base.py @@ -90,8 +90,8 @@ class PipelineConfig: def __setattr__(self, name: str, value: Any) -> None: if self.__dict__.get("_frozen") and name in self.__dataclass_fields__: raise AttributeError(f"{type(self).__name__}.{name} was decided by config resolution and is read-only; " - f"to change pipeline_config.{name}, override its typed path with " - "resolved_config.with_override(source, {path: value}) and use the returned config") + f"set the typed path of pipeline_config.{name} in the input config, or decide it in a " + "resolution step") super().__setattr__(name, value) def freeze(self) -> None: diff --git a/fastvideo/dataset/__init__.py b/fastvideo/dataset/__init__.py index 30f36bab0..36881e54f 100644 --- a/fastvideo/dataset/__init__.py +++ b/fastvideo/dataset/__init__.py @@ -9,9 +9,9 @@ from fastvideo.dataset.transform import (CenterCropResizeVideo, Normalize255, Te from fastvideo.dataset.validation_dataset import ValidationDataset -def getdataset(preprocess_config) -> VideoCaptionMergedDataset: +def getdataset(preprocess_config, model_path: str) -> VideoCaptionMergedDataset: """Build the video-caption dataset of ``preprocess_config``, the ``preprocess`` section of a resolved - ``PreprocessRunConfig``.""" + ``PreprocessRunConfig``. ``model_path`` is the local checkpoint directory that holds the tokenizer.""" temporal_sample = (TemporalRandomCrop(preprocess_config.num_frames) if preprocess_config.do_temporal_sample else None) # 16 x norm_fun = Lambda(lambda x: 2.0 * x - 1.0) @@ -32,15 +32,19 @@ def getdataset(preprocess_config) -> VideoCaptionMergedDataset: ]) return VideoCaptionMergedDataset(data_merge_path=preprocess_config.data_merge_path, preprocess_config=preprocess_config, + model_path=model_path, transform=transform, temporal_sample=temporal_sample, transform_topcrop=transform_topcrop, seed=preprocess_config.seed) -def gettextdataset(preprocess_config) -> TextDataset: +def gettextdataset(preprocess_config, model_path: str) -> TextDataset: + """Build the text dataset of ``preprocess_config``; ``model_path`` is the local checkpoint directory that holds + the tokenizer.""" return TextDataset(data_merge_path=preprocess_config.data_merge_path, preprocess_config=preprocess_config, + model_path=model_path, seed=preprocess_config.seed) diff --git a/fastvideo/dataset/preprocessing_datasets.py b/fastvideo/dataset/preprocessing_datasets.py index 8d7cf9c6c..71060c243 100644 --- a/fastvideo/dataset/preprocessing_datasets.py +++ b/fastvideo/dataset/preprocessing_datasets.py @@ -387,6 +387,7 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset, torch.distribu def __init__(self, data_merge_path: str, preprocess_config, + model_path: str, transform, temporal_sample, transform_topcrop, @@ -397,8 +398,8 @@ class VideoCaptionMergedDataset(torch.utils.data.IterableDataset, torch.distribu self.temporal_sample = temporal_sample self.seed = seed - # Initialize tokenizer - tokenizer_path = os.path.join(preprocess_config.model_path, "tokenizer") + # Initialize tokenizer from the local checkpoint directory + tokenizer_path = os.path.join(model_path, "tokenizer") tokenizer = None if os.path.exists(tokenizer_path): try: @@ -565,13 +566,13 @@ class TextDataset(torch.utils.data.IterableDataset, torch.distributed.checkpoint This dataset processes text data through text encoding stages only. """ - def __init__(self, data_merge_path: str, preprocess_config, start_idx: int = 0, seed: int = 42): + def __init__(self, data_merge_path: str, preprocess_config, model_path: str, start_idx: int = 0, seed: int = 42): self.data_merge_path = data_merge_path self.start_idx = start_idx self.seed = seed - # Initialize tokenizer - tokenizer_path = os.path.join(preprocess_config.model_path, "tokenizer") + # Initialize tokenizer from the local checkpoint directory + tokenizer_path = os.path.join(model_path, "tokenizer") tokenizer = AutoTokenizer.from_pretrained(tokenizer_path, cache_dir=preprocess_config.cache_dir) # Initialize text encoding stage diff --git a/fastvideo/entrypoints/streaming/server.py b/fastvideo/entrypoints/streaming/server.py index cabecf614..911314713 100644 --- a/fastvideo/entrypoints/streaming/server.py +++ b/fastvideo/entrypoints/streaming/server.py @@ -189,7 +189,7 @@ def run_server(serve_config: ServeConfig, *, generator: _GeneratorProto | None = if generator is None: from fastvideo import VideoGenerator # lazy to avoid boot cost - generator = VideoGenerator.from_pretrained(config=serve_config.generator) + generator = VideoGenerator.from_config(serve_config.generator) app = build_app(serve_config, generator) uvicorn.run( app, diff --git a/fastvideo/entrypoints/streaming/worker.py b/fastvideo/entrypoints/streaming/worker.py index d90ddb2a3..bbd96a8ad 100644 --- a/fastvideo/entrypoints/streaming/worker.py +++ b/fastvideo/entrypoints/streaming/worker.py @@ -56,7 +56,7 @@ def worker_main( try: from fastvideo import VideoGenerator - generator = VideoGenerator.from_pretrained(config=generator_config) + generator = VideoGenerator.from_config(generator_config) if warmup_config.enabled: _warmup_worker(generator, warmup_config) result_queue.put({"kind": "ready", "worker_id": worker_id}) diff --git a/fastvideo/entrypoints/video_generator.py b/fastvideo/entrypoints/video_generator.py index 283be9ec9..eb50199da 100644 --- a/fastvideo/entrypoints/video_generator.py +++ b/fastvideo/entrypoints/video_generator.py @@ -29,7 +29,6 @@ import fastvideo.envs as envs from fastvideo.api.compat import ( REQUEST_BATCH_EXTRA_PASSTHROUGH_FIELDS, expand_request_prompt_batch, - from_pretrained_kwargs_to_config, load_generator_config_from_file, normalize_generation_request, normalize_generator_config, @@ -150,49 +149,41 @@ class VideoGenerator: @classmethod def from_pretrained( cls, - model_path: str | GeneratorConfig | Mapping[str, Any] | None = None, - **kwargs, + model_path: str, + config: GeneratorConfig | Mapping[str, Any] | None = None, + *, + log_queue=None, + **flat_keywords, ) -> "VideoGenerator": """ Create a video generator from a pretrained model. + ``from_pretrained(model_path, config)`` is ``from_config({"model_path": model_path, **config})``. + Args: model_path: Path or identifier for the pretrained model - **kwargs: The common engine and offload keywords in - ``fastvideo.api.compat.FROM_PRETRAINED_KWARGS``, such as - ``num_gpus`` and ``dit_cpu_offload``. Pass any other setting - through ``VideoGenerator.from_config(...)``. + config: The settings besides ``model_path``, as a nested mapping such as + ``{"engine": {"num_gpus": 2, "offload": {"dit": False}}}`` or a ``GeneratorConfig``. A + ``model_path`` inside it must equal ``model_path``. + log_queue: Optional multiprocessing.Queue to forward worker logs to Returns: The created video generator """ - log_queue = kwargs.pop("log_queue", None) - if kwargs.pop("nvfp4_fa4", False): - import os - os.environ["FASTVIDEO_NVFP4_FA4"] = "1" - envs.setdefault_external("CUTE_DSL_ENABLE_TVM_FFI", "1") - typed_config = kwargs.pop("config", None) - if typed_config is not None: - if model_path is not None: - raise TypeError("Pass either model_path or config to from_pretrained, not both") - if kwargs: - unexpected = ", ".join(sorted(kwargs)) - raise TypeError(f"Unexpected keyword arguments with config: {unexpected}") - return cls.from_config(typed_config, log_queue=log_queue) - - if isinstance(model_path, GeneratorConfig | Mapping): - if kwargs: - unexpected = ", ".join(sorted(kwargs)) - raise TypeError(f"Unexpected keyword arguments with typed config: {unexpected}") - return cls.from_config(model_path, log_queue=log_queue) - - if model_path is None: - raise TypeError("model_path or config is required") - - return cls.from_config( - from_pretrained_kwargs_to_config(model_path, kwargs), - log_queue=log_queue, - ) + if flat_keywords: + unexpected = ", ".join(sorted(flat_keywords)) + raise TypeError(f"VideoGenerator.from_pretrained(...) does not accept {unexpected}. Pass settings as a " + "nested config at their typed paths, for example " + "from_pretrained(model_path, {'engine': {'num_gpus': 2, 'offload': {'dit': False}}}).") + if config is None: + return cls.from_config({"model_path": model_path}, log_queue=log_queue) + if isinstance(config, GeneratorConfig): + if config.model_path != model_path: + raise ValueError(f"config.model_path {config.model_path!r} differs from model_path {model_path!r}") + return cls.from_config(config, log_queue=log_queue) + if config.get("model_path", model_path) != model_path: + raise ValueError(f"config['model_path'] {config['model_path']!r} differs from model_path {model_path!r}") + return cls.from_config({**config, "model_path": model_path}, log_queue=log_queue) @classmethod def from_config( diff --git a/fastvideo/models/loader/component_loader.py b/fastvideo/models/loader/component_loader.py index 906f0af4e..1e5e9beb1 100644 --- a/fastvideo/models/loader/component_loader.py +++ b/fastvideo/models/loader/component_loader.py @@ -19,7 +19,7 @@ from torch.distributed import init_device_mesh from transformers import AutoImageProcessor, AutoProcessor, AutoTokenizer from transformers.utils import SAFE_WEIGHTS_INDEX_NAME -from fastvideo.api.device_policy import UNIFIED_MEMORY_OFFLOAD_PATHS, offload_disabled_on_unified_memory +from fastvideo.api.device_policy import UNIFIED_MEMORY_OFFLOAD_PATHS from fastvideo.api.inference_resolution import torch_compile_kwargs from fastvideo.api.resolution import ResolvedGeneratorConfig from fastvideo.attention.selector import ( @@ -358,27 +358,12 @@ class TextEncoderLoader(ComponentLoader): offload_flag: str = "text_encoder_cpu_offload", ): runtime_device = get_local_torch_device() - device_id = runtime_device.index if runtime_device.index is not None else 0 if cpu_offload is None: - # offload_flag names the device-policy switch of this component; its value is the - # engine.offload field at the same path (text_encoder_cpu_offload -> engine.offload.text_encoder). + # offload_flag names the device-policy switch of this component; its value is the engine.offload field + # at the same path (text_encoder_cpu_offload -> engine.offload.text_encoder), which resolution already + # settled for the memory class of this device. offload_field = UNIFIED_MEMORY_OFFLOAD_PATHS[offload_flag].rsplit(".", 1)[1] - requested_cpu_offload = getattr(resolved_config.engine.offload, offload_field) - else: - requested_cpu_offload = cpu_offload - # The resolved config already carries the worker's offload policy, so only ask whether the policy covers - # this component. - disable_cpu_offload = offload_disabled_on_unified_memory(device_id, offload_flag) - - if requested_cpu_offload and disable_cpu_offload: - # Direct loader callers can choose a CPU target before the worker - # applies its device-local policy. Reset both the request and the - # target so the model is never constructed on the host first. - logger.info("Disabling %s on unified-memory device %d", offload_flag, device_id) - cpu_offload = False - target_device = runtime_device - else: - cpu_offload = requested_cpu_offload + cpu_offload = getattr(resolved_config.engine.offload, offload_field) use_cpu_offload = (cpu_offload and len(getattr(model_config, "_fsdp_shard_conditions", [])) > 0) from fastvideo.platforms import current_platform @@ -1050,11 +1035,14 @@ class TransformerLoader(ComponentLoader): def load(self, model_path: str, resolved_config: ResolvedGeneratorConfig, - loading_teacher_critic_model: bool = False): + loading_teacher_critic_model: bool = False, + override_transformer_cls_name: str | None = None): """Load the transformer based on the model path, and inference args. ``loading_teacher_critic_model`` marks a DMD teacher (real_score) or critic (fake_score) transformer, which - loads its checkpoint weights in full precision with dense attention. + loads its checkpoint weights in full precision with dense attention. ``override_transformer_cls_name`` is the + model class to build for this load; it takes precedence over + ``pipeline.components.override_transformer_cls_name``. """ config = get_diffusers_config(model=model_path) hf_config = deepcopy(config) @@ -1065,7 +1053,8 @@ class TransformerLoader(ComponentLoader): "Only diffusers format is supported.") logger.info("transformer cls_name: %s", cls_name) - override_transformer_cls_name = resolved_config.pipeline.components.override_transformer_cls_name + if override_transformer_cls_name is None: + override_transformer_cls_name = resolved_config.pipeline.components.override_transformer_cls_name if override_transformer_cls_name is not None: cls_name = override_transformer_cls_name logger.info("Overriding transformer cls_name to %s", cls_name) @@ -1392,6 +1381,7 @@ class PipelineComponentLoader: transformers_or_diffusers: str, resolved_config: ResolvedGeneratorConfig, loading_teacher_critic_model: bool = False, + override_transformer_cls_name: str | None = None, ): """ Load a pipeline module. @@ -1403,6 +1393,9 @@ class PipelineComponentLoader: resolved_config: The resolved runtime config loading_teacher_critic_model: Whether a transformer is a DMD teacher or critic; passed to ``TransformerLoader.load`` and ignored by the other loaders + override_transformer_cls_name: The model class to build for a transformer, over + ``pipeline.components.override_transformer_cls_name``; passed to ``TransformerLoader.load`` and + ignored by the other loaders Returns: The loaded module @@ -1438,5 +1431,6 @@ class PipelineComponentLoader: if isinstance(loader, TransformerLoader): return loader.load(component_model_path, resolved_config, - loading_teacher_critic_model=loading_teacher_critic_model) + loading_teacher_critic_model=loading_teacher_critic_model, + override_transformer_cls_name=override_transformer_cls_name) return loader.load(component_model_path, resolved_config) diff --git a/fastvideo/models/pinned_offload.py b/fastvideo/models/pinned_offload.py index 08c45802a..faf56dd14 100644 --- a/fastvideo/models/pinned_offload.py +++ b/fastvideo/models/pinned_offload.py @@ -63,7 +63,7 @@ def load(module: nn.Module, device: torch.device, pin: bool = True) -> nn.Module return module.to(device) missing = [(name, t) for name, t in _tensors(module) if t.data.device != device] if not missing: - # Already resident, which is what ``vae_cpu_offload=False`` leaves. Taking + # Already resident, which is what ``engine.offload.vae: false`` leaves. Taking # host copies here would pull the whole module back over PCIe once and keep # a pinned mirror of it for the life of the process, for a module that never # leaves the device. diff --git a/fastvideo/pipelines/basic/ltx2/ltx2_pipeline.py b/fastvideo/pipelines/basic/ltx2/ltx2_pipeline.py index 75014e674..82d75888e 100644 --- a/fastvideo/pipelines/basic/ltx2/ltx2_pipeline.py +++ b/fastvideo/pipelines/basic/ltx2/ltx2_pipeline.py @@ -18,88 +18,6 @@ from fastvideo.pipelines.stages import (DecodingStage, InputValidationStage, LTX logger = init_logger(__name__) -# Values of the stage-2 refine settings that neither the input nor the checkpoint's model_index.json decided. -_REFINE_DEFAULTS: dict[str, Any] = { - "pipeline.ltx2.refine.enabled": False, - "pipeline.ltx2.refine.add_noise": True, - "pipeline.ltx2.refine.guidance_scale": 1.0, - "pipeline.ltx2.refine.num_inference_steps": 3, -} -# Refine component paths, by the model_index.json key that supplies the checkpoint's default. -_REFINE_PATH_KEYS = { - "pipeline.ltx2.refine.transformer_path": "fastvideo_refine_transformer_path", - "pipeline.ltx2.refine.lora_path": "fastvideo_refine_lora_path", - "pipeline.ltx2.refine.noise_path": "fastvideo_refine_noise_path", - "pipeline.ltx2.refine.audio_noise_path": "fastvideo_refine_audio_noise_path", -} - - -def _resolve_refine_path(model_path: str, value: str | None) -> str | None: - """Resolve a refine component path, preferring paths inside the local - model snapshot when the value is relative.""" - if value is None: - return None - if os.path.isabs(value): - return value - candidate = os.path.join(model_path, value) - if os.path.exists(candidate): - return candidate - return value - - -def _resolve_refine_upsampler_path(model_path: str, model_index: dict[str, Any]) -> str | None: - """Resolve the refine upsampler directory for a local model snapshot. - - model_index.json keys ("fastvideo_refine_upsampler_path", then - "spatial_upsampler") are the documented override and win when present. - Otherwise probe the snapshot root: published LTX-2 distilled checkpoints - bundle the upsampler weights on disk (spelled "spatial_upscaler") - without declaring them in model_index.json. - """ - path = _resolve_refine_path(model_path, model_index.get("fastvideo_refine_upsampler_path")) - if path is None and "spatial_upsampler" in model_index: - path = _resolve_refine_path(model_path, "spatial_upsampler") - if path is None: - for dirname in ("spatial_upscaler", "spatial_upsampler"): - candidate = os.path.join(model_path, dirname) - if os.path.isdir(candidate): - return candidate - return path - - -def _checkpoint_refine_values(resolved_config: ResolvedGeneratorConfig, model_path: str, - model_index: dict[str, Any]) -> dict[str, Any]: - """The refine settings that the checkpoint's ``model_index.json`` and the refine defaults change. - - Distilled checkpoints bundle ``fastvideo_refine_*`` defaults so the pipeline can configure itself. Each one fills - its setting only while the setting is ``None``, so a value from the input always wins; ``fastvideo_refine_enabled`` - can only turn refine on. The upsampler falls back to a directory in the snapshot. The four switches that are still - ``None`` afterwards take ``_REFINE_DEFAULTS``. The result holds only the paths whose value changes. - """ - - def current(path: str) -> Any: - return resolved_config.provenance(path).value - - values: dict[str, Any] = {} - if model_index.get("fastvideo_refine_enabled") is True and current("pipeline.ltx2.refine.enabled") is None: - values["pipeline.ltx2.refine.enabled"] = True - if current("pipeline.components.upsampler_weights") is None: - values["pipeline.components.upsampler_weights"] = _resolve_refine_upsampler_path(model_path, model_index) - for path, key in _REFINE_PATH_KEYS.items(): - if current(path) is None: - values[path] = _resolve_refine_path(model_path, model_index.get(key)) - for path, key, convert in ( - ("pipeline.ltx2.refine.num_inference_steps", "fastvideo_refine_num_inference_steps", int), - ("pipeline.ltx2.refine.guidance_scale", "fastvideo_refine_guidance_scale", float), - ("pipeline.ltx2.refine.add_noise", "fastvideo_refine_add_noise", bool), - ): - if current(path) is None and model_index.get(key) is not None: - values[path] = convert(model_index[key]) - for path, default in _REFINE_DEFAULTS.items(): - if values.get(path, current(path)) is None: - values[path] = default - return {path: value for path, value in values.items() if value != current(path)} - class LTX2Pipeline(LoRAPipeline): @@ -238,13 +156,6 @@ class LTX2Pipeline(LoRAPipeline): model_index = self._load_config(self.model_path) logger.info("Loading pipeline modules from config: %s", model_index) - # Apply the refine defaults bundled in model_index.json, then the remaining refine defaults. The pipeline owns - # its config; later stages read the rebound one. - refine_values = _checkpoint_refine_values(resolved_config, self.model_path, model_index) - if refine_values: - resolved_config = resolved_config.with_override("checkpoint:model_index.json", refine_values) - self.resolved_config = resolved_config - model_index.pop("_class_name") model_index.pop("_diffusers_version") model_index.pop("workload_type", None) diff --git a/fastvideo/pipelines/basic/minimax_h3/minimax_h3_pipeline.py b/fastvideo/pipelines/basic/minimax_h3/minimax_h3_pipeline.py index 8a3f5da90..071ccdf2d 100644 --- a/fastvideo/pipelines/basic/minimax_h3/minimax_h3_pipeline.py +++ b/fastvideo/pipelines/basic/minimax_h3/minimax_h3_pipeline.py @@ -160,10 +160,15 @@ class MiniMaxH3BasePipeline(LoRAPipeline, ComposedPipelineBase): shift = getattr(self.get_module(module_name), "shift", None) if shift is None or not math.isfinite(float(shift)) or float(shift) <= 0: raise ValueError(f"MiniMax-H3 {modality} scheduler must expose a positive finite shift, got {shift}.") - self._load_checkpoint_schedule(resolved_config) + self._validate_checkpoint_schedule(resolved_config) - def _load_checkpoint_schedule(self, resolved_config: ResolvedGeneratorConfig) -> None: - """A distilled export's schedule is explicit; never silently use a uniform grid.""" + def _validate_checkpoint_schedule(self, resolved_config: ResolvedGeneratorConfig) -> None: + """Check the checkpoint's ``fastvideo_inference.json`` against the loaded schedulers and the resolved schedule. + + Resolution fills ``pipeline.dmd_denoising_steps`` from that file (``fill_dmd_schedule_from_checkpoint``); a + distilled export's schedule is explicit, so a config that disagrees with the file, or that has no schedule + although the file declares one, raises rather than silently using a uniform grid. + """ path = Path(self.model_path) / "fastvideo_inference.json" if not path.is_file(): return @@ -190,10 +195,12 @@ class MiniMaxH3BasePipeline(LoRAPipeline, ComposedPipelineBase): or declared <= 0 or float(self.get_module(name).shift) != float(declared)): raise ValueError(f"FastH3 checkpoint {key}={declared!r} disagrees with {name}/scheduler_config.json.") configured_steps = resolved_config.pipeline.dmd_denoising_steps - if configured_steps is not None and list(configured_steps) != steps: + if configured_steps is None: + raise ValueError("The checkpoint's fastvideo_inference.json declares a DMD schedule but " + "pipeline.dmd_denoising_steps is unset; resolve the config with the checkpoint's " + "model_path so that fill_dmd_schedule_from_checkpoint fills it.") + if list(configured_steps) != steps: raise ValueError("Explicit DMD schedule disagrees with the checkpoint's trained DMD rungs.") - self.resolved_config = resolved_config.with_override("checkpoint:fastvideo_inference.json", - {"pipeline.dmd_denoising_steps": list(steps)}) logger.info("FastH3 checkpoint schedule: %d transformer forwards, DMD rungs=%s, video/audio shifts=%s/%s", len(steps), steps, self.get_module("scheduler").shift, @@ -412,7 +419,7 @@ class MiniMaxH3BasePipeline(LoRAPipeline, ComposedPipelineBase): self._add_denoise_stages(ref2va=ref2va) def forward(self, batch: ForwardBatch, resolved_config: ResolvedGeneratorConfig) -> ForwardBatch: - resolved_config = self._post_init_before_forward(resolved_config) + self._post_init_before_forward() # Sequential encode-then-release is the H3-only fallback. Lazy and the # fully-resident discrete-GPU path both keep a complete stage list and diff --git a/fastvideo/pipelines/basic/stable_audio/stages/latent_preparation.py b/fastvideo/pipelines/basic/stable_audio/stages/latent_preparation.py index febaa07f8..609a00b2e 100644 --- a/fastvideo/pipelines/basic/stable_audio/stages/latent_preparation.py +++ b/fastvideo/pipelines/basic/stable_audio/stages/latent_preparation.py @@ -94,7 +94,7 @@ class StableAudioLatentPreparationStage(PipelineStage): have to. """ assert self.vae is not None, "VAE required for init_audio / inpaint_audio encoding" - # VAE may be CPU-parked under `vae_cpu_offload=True`. + # VAE may be CPU-parked under `engine.offload.vae: true`. self.vae = self.vae.to(device) if isinstance(audio, str | os.PathLike): audio = _decode_audio_file(audio, target_sr=self.sample_rate) diff --git a/fastvideo/pipelines/composed_pipeline_base.py b/fastvideo/pipelines/composed_pipeline_base.py index bc7cc132d..7bb3967f1 100644 --- a/fastvideo/pipelines/composed_pipeline_base.py +++ b/fastvideo/pipelines/composed_pipeline_base.py @@ -13,11 +13,10 @@ from typing import Any, cast import torch -from fastvideo.api.device_policy import finalize_device_offload_policy from fastvideo.api.inference_resolution import torch_compile_kwargs from fastvideo.api.resolution import ResolvedGeneratorConfig, thaw from fastvideo.distributed import ( - get_local_torch_device, + get_local_torch_device, # noqa: F401 # the MiniMax-H3 pipeline reads the device through this module get_world_group, maybe_init_distributed_environment_and_model_parallel, ) @@ -155,16 +154,6 @@ class ComposedPipelineBase(ABC): maybe_init_distributed_environment_and_model_parallel(resolved_config.engine.parallelism.tp_size, resolved_config.engine.parallelism.sp_size) - # VideoGenerator applies this in each Worker before building the - # pipeline. Keep direct from_pretrained/build_pipeline callers aligned, - # but only after distributed setup has selected this process's device. - # On a config that the worker already decided, the policy changes nothing. - if resolved_config.inference_mode: - local_device = get_local_torch_device() - device_id = local_device.index if local_device.index is not None else 0 - resolved_config = finalize_device_offload_policy(resolved_config, device_id) - self.resolved_config = resolved_config - # Torch profiler. Enabled and configured through env vars: # FASTVIDEO_TORCH_PROFILER_DIR=/path/to/save/trace trace_dir = envs.FASTVIDEO_TORCH_PROFILER_DIR.get() @@ -727,17 +716,10 @@ class ComposedPipelineBase(ABC): stage_name) self._install_lazy_release_hooks() - def _post_init_before_forward(self, resolved_config: ResolvedGeneratorConfig) -> ResolvedGeneratorConfig: - """Run ``post_init`` before the first forward, and return the config that the stages read. - - ``post_init`` can rebind ``self.resolved_config`` to an override of it, such as a schedule from the - checkpoint. A caller that passed the pipeline's own config then runs with the rebound one. - """ - if self.post_init_called: - return resolved_config - passed_own_config = resolved_config is self.resolved_config - self.post_init() - return self.resolved_config if passed_own_config else resolved_config + def _post_init_before_forward(self) -> None: + """Run ``post_init`` once, before the first forward.""" + if not self.post_init_called: + self.post_init() # TODO(will): don't hardcode no_grad @torch.no_grad() @@ -755,7 +737,7 @@ class ComposedPipelineBase(ABC): Returns: ForwardBatch: The batch with the generated video or image. """ - resolved_config = self._post_init_before_forward(resolved_config) + self._post_init_before_forward() # Execute each stage logger.info("Running pipeline stages: %s", self._stage_name_mapping.keys()) diff --git a/fastvideo/pipelines/preprocess/matrixgame2/matrixgame2_preprocess_pipeline_ode_trajectory.py b/fastvideo/pipelines/preprocess/matrixgame2/matrixgame2_preprocess_pipeline_ode_trajectory.py index a5ebeaf1f..7ecc729dc 100644 --- a/fastvideo/pipelines/preprocess/matrixgame2/matrixgame2_preprocess_pipeline_ode_trajectory.py +++ b/fastvideo/pipelines/preprocess/matrixgame2/matrixgame2_preprocess_pipeline_ode_trajectory.py @@ -364,7 +364,7 @@ class PreprocessPipeline_MatrixGame2_ODE_Trajectory(BasePreprocessPipeline): os.makedirs(self.combined_parquet_dir, exist_ok=True) # Loading dataset - train_dataset = getdataset(preprocess_config) + train_dataset = getdataset(preprocess_config, self.model_path) self.preprocess_dataloader = DataLoader( train_dataset, diff --git a/fastvideo/pipelines/preprocess/preprocess_pipeline_base.py b/fastvideo/pipelines/preprocess/preprocess_pipeline_base.py index fcefb8f48..90366dfe5 100644 --- a/fastvideo/pipelines/preprocess/preprocess_pipeline_base.py +++ b/fastvideo/pipelines/preprocess/preprocess_pipeline_base.py @@ -251,7 +251,7 @@ class BasePreprocessPipeline(ComposedPipelineBase): start_idx += table.num_rows # Loading dataset - train_dataset = getdataset(preprocess_config) + train_dataset = getdataset(preprocess_config, self.model_path) train_dataloader = DataLoader( train_dataset, diff --git a/fastvideo/pipelines/preprocess/preprocess_pipeline_ode_trajectory.py b/fastvideo/pipelines/preprocess/preprocess_pipeline_ode_trajectory.py index 23cb12059..c0c029415 100644 --- a/fastvideo/pipelines/preprocess/preprocess_pipeline_ode_trajectory.py +++ b/fastvideo/pipelines/preprocess/preprocess_pipeline_ode_trajectory.py @@ -259,7 +259,7 @@ class PreprocessPipeline_ODE_Trajectory(BasePreprocessPipeline): os.makedirs(self.combined_parquet_dir, exist_ok=True) # Loading dataset - train_dataset = gettextdataset(preprocess_config) + train_dataset = gettextdataset(preprocess_config, self.model_path) self.preprocess_dataloader = DataLoader( train_dataset, diff --git a/fastvideo/pipelines/preprocess/preprocess_pipeline_text.py b/fastvideo/pipelines/preprocess/preprocess_pipeline_text.py index 73b701e21..0f94b7807 100644 --- a/fastvideo/pipelines/preprocess/preprocess_pipeline_text.py +++ b/fastvideo/pipelines/preprocess/preprocess_pipeline_text.py @@ -152,7 +152,7 @@ class PreprocessPipeline_Text(BasePreprocessPipeline): os.makedirs(self.combined_parquet_dir, exist_ok=True) # Loading text dataset - train_dataset = gettextdataset(preprocess_config) + train_dataset = gettextdataset(preprocess_config, self.model_path) self.preprocess_dataloader = DataLoader( train_dataset, diff --git a/fastvideo/pipelines/preprocess/v1_preprocess.py b/fastvideo/pipelines/preprocess/v1_preprocess.py index b44a96c3c..5947ab121 100644 --- a/fastvideo/pipelines/preprocess/v1_preprocess.py +++ b/fastvideo/pipelines/preprocess/v1_preprocess.py @@ -12,22 +12,16 @@ from fastvideo.pipelines.preprocess.preprocess_pipeline_text import (PreprocessP from fastvideo.pipelines.preprocess.matrixgame2.matrixgame2_preprocess_pipeline import (PreprocessPipeline_MatrixGame2) from fastvideo.pipelines.preprocess.matrixgame2.matrixgame2_preprocess_pipeline_ode_trajectory import ( PreprocessPipeline_MatrixGame2_ODE_Trajectory) -from fastvideo.utils import maybe_download_model logger = init_logger(__name__) def main(resolved_config: ResolvedGeneratorConfig) -> None: - """Download the model and run the pipeline of ``preprocess.preprocess_task`` on one GPU. + """Run the pipeline of ``preprocess.preprocess_task`` on one GPU. - The local checkpoint directory becomes ``model_path`` and ``preprocess.model_path``, which the pipelines and the - dataset tokenizer read. + The pipeline downloads the checkpoint of ``model_path`` and keeps the local directory as its ``model_path``, which + the dataset tokenizer reads too. """ - model_path = maybe_download_model(resolved_config.model_path) - resolved_config = resolved_config.with_override("v1_preprocess.main", { - "model_path": model_path, - "preprocess.model_path": model_path, - }) maybe_init_distributed_environment_and_model_parallel(1, 1) num_gpus = int(os.environ["WORLD_SIZE"]) assert num_gpus == 1, "Only support 1 GPU" @@ -61,8 +55,8 @@ def main(resolved_config: ResolvedGeneratorConfig) -> None: logger.info("Preprocess task: %s using %s", preprocess_task, PreprocessPipeline.__name__) - pipeline = PreprocessPipeline(model_path, resolved_config) - pipeline.forward(batch=None, resolved_config=pipeline.resolved_config) + pipeline = PreprocessPipeline(resolved_config.model_path, resolved_config) + pipeline.forward(batch=None, resolved_config=resolved_config) if __name__ == "__main__": diff --git a/fastvideo/pipelines/stages/kandinsky5.py b/fastvideo/pipelines/stages/kandinsky5.py index c126e2fb7..6065f1ca7 100644 --- a/fastvideo/pipelines/stages/kandinsky5.py +++ b/fastvideo/pipelines/stages/kandinsky5.py @@ -248,7 +248,7 @@ class Kandinsky5DenoisingStage(PipelineStage): transformer ``ValidationCallback`` reuses from training (whose ``engine.precision.dit`` reflects the fp32 master-weight load dtype, not the actual bf16 compute dtype) and multi-GPU - ``use_fsdp_inference=True`` runs. + ``engine.use_fsdp_inference: true`` runs. For that FSDP case, read the policy itself (``set_mixed_precision_policy`` records it right before diff --git a/fastvideo/tests/api/config_snapshot.py b/fastvideo/tests/api/config_snapshot.py index 3f9b450e5..92e25e848 100644 --- a/fastvideo/tests/api/config_snapshot.py +++ b/fastvideo/tests/api/config_snapshot.py @@ -9,10 +9,8 @@ resolution decisions; a request case records the ``SamplingParam`` and the reque - ``yaml``: every repository config file with a ``generator`` section, through ``load_generator_config_from_file`` and ``resolve_inference_config``; its ``request`` or ``default_request`` through the ``fastvideo generate`` and ``fastvideo serve`` loaders and ``request_to_sampling_param``. -- ``kwargs``: ``VideoGenerator.from_pretrained`` keywords, through ``from_pretrained_kwargs_to_config`` and the same - resolution. -- ``config``: typed ``GeneratorConfig`` mappings, as ``VideoGenerator.from_config`` receives them, through the same - resolution. +- ``config``: typed ``GeneratorConfig`` mappings, as ``VideoGenerator.from_config`` and + ``VideoGenerator.from_pretrained`` receive them, through the same resolution. - ``environment``: environment variables that resolution folds into typed fields. - ``request``: typed ``GenerationRequest`` values, through ``request_to_sampling_param``. @@ -187,13 +185,17 @@ def snapshot_request(request: Any, model_path: str) -> dict[str, Any]: @contextlib.contextmanager def isolated_environment(env_values: dict[str, Any] | None = None) -> Iterator[None]: - """Unset every registered FastVideo environment variable, set ``env_values``, and block model-index downloads. + """Unset every registered FastVideo environment variable, set ``env_values``, block model-index downloads, and + turn the device policy off. The registry's ``override`` context managers restore the previous values on exit. Downloads are blocked by - replacing ``fastvideo.registry.maybe_download_model_index``, the only network call on the configuration path. + replacing ``fastvideo.registry.maybe_download_model_index``, the only network call on the configuration path. The + device policy steps are turned off (``device_policy.APPLY_DEVICE_POLICY``) so that a snapshot does not depend on + the machine that produced it. """ import fastvideo.envs as envs import fastvideo.registry as registry + from fastvideo.api import device_policy def refuse_download(model_path: str, *args: Any, **kwargs: Any) -> dict[str, Any]: raise OfflineResolutionError(f"resolving {model_path!r} needs a model_index.json download") @@ -205,6 +207,9 @@ def isolated_environment(env_values: dict[str, Any] | None = None) -> Iterator[N original = registry.maybe_download_model_index registry.maybe_download_model_index = refuse_download stack.callback(setattr, registry, "maybe_download_model_index", original) + apply_policy = device_policy.APPLY_DEVICE_POLICY + device_policy.APPLY_DEVICE_POLICY = False + stack.callback(setattr, device_policy, "APPLY_DEVICE_POLICY", apply_policy) yield @@ -249,16 +254,6 @@ def _yaml_request(path: Path) -> Any: return build_serve_config(namespace).default_request -def _kwargs_case(model_path: str, kwargs: dict[str, Any]) -> Callable[[], dict[str, Any]]: - - def build() -> dict[str, Any]: - from fastvideo.api.compat import from_pretrained_kwargs_to_config - - return snapshot_resolved_generator_config(from_pretrained_kwargs_to_config(model_path, kwargs)) - - return build - - def _config_case(raw: dict[str, Any]) -> Callable[[], dict[str, Any]]: def build() -> dict[str, Any]: @@ -318,37 +313,46 @@ def collect_cases() -> list[SnapshotCase]: SnapshotCase("yaml", path.relative_to(REPO_ROOT).as_posix(), _yaml_case(path)) for path in _repository_config_files() ] - # One case per kind of branch in from_pretrained_kwargs_to_config: keywords of typed fields and compile sub-keys. - kwargs_cases = { - "typed_offload_and_parallelism": (WAN_T2V, { - "num_gpus": 2, - "sp_size": 2, - "dit_cpu_offload": False, - "vae_cpu_offload": False, - }), - "torch_compile_kwargs_typed_and_extra": (WAN_T2V, { - "enable_torch_compile": True, - "torch_compile_kwargs": { - "backend": "inductor", - "mode": "max-autotune", - "fullgraph": True, - "options": { - "triton.cudagraphs": False - }, - }, - }), - "disable_autocast": (WAN_T2V, { - "disable_autocast": True - }), - } - cases += [ - SnapshotCase("kwargs", name, _kwargs_case(model, kwargs)) for name, (model, kwargs) in kwargs_cases.items() - ] - # Typed configs for the settings that only from_config accepts: typed fields, a model-only PipelineConfig - # attribute and a key that code reads by name in pipeline.experimental, LTX-2 refine and VAE tile fields, and a - # pipeline config JSON path. + # Typed configs: engine, offload, parallelism, and compile fields, a model-only PipelineConfig attribute and a key + # that code reads by name in pipeline.experimental, LTX-2 refine and VAE tile fields, and a pipeline config JSON + # path. pipeline_json = str(REPO_ROOT / "fastvideo/configs/fasthunyuan_t2v.json") config_cases = { + "typed_offload_and_parallelism": { + "model_path": WAN_T2V, + "engine": { + "num_gpus": 2, + "parallelism": { + "sp_size": 2 + }, + "offload": { + "dit": False, + "vae": False + }, + }, + }, + "torch_compile_kwargs_typed_and_extra": { + "model_path": WAN_T2V, + "engine": { + "compile": { + "enabled": True, + "backend": "inductor", + "mode": "max-autotune", + "fullgraph": True, + "extras": { + "options": { + "triton.cudagraphs": False + } + }, + } + }, + }, + "disable_autocast": { + "model_path": WAN_T2V, + "engine": { + "disable_autocast": True + }, + }, "attention_precision_and_flow_shift": { "model_path": WAN_T2V, "engine": { diff --git a/fastvideo/tests/api/config_snapshot_goldens/config/attention_precision_and_flow_shift.json b/fastvideo/tests/api/config_snapshot_goldens/config/attention_precision_and_flow_shift.json index 0034f1ea4..869069cbe 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/config/attention_precision_and_flow_shift.json +++ b/fastvideo/tests/api/config_snapshot_goldens/config/attention_precision_and_flow_shift.json @@ -318,6 +318,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -333,6 +337,7 @@ "backend": "TORCH_SDPA", "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.5, "vsa_tile_size": 256 }, @@ -421,14 +426,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/config/boundary_ratio.json b/fastvideo/tests/api/config_snapshot_goldens/config/boundary_ratio.json index 0c1271ba0..a961ee6f0 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/config/boundary_ratio.json +++ b/fastvideo/tests/api/config_snapshot_goldens/config/boundary_ratio.json @@ -332,6 +332,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -347,6 +351,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -440,14 +445,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/kwargs/disable_autocast.json b/fastvideo/tests/api/config_snapshot_goldens/config/disable_autocast.json similarity index 97% rename from fastvideo/tests/api/config_snapshot_goldens/kwargs/disable_autocast.json rename to fastvideo/tests/api/config_snapshot_goldens/config/disable_autocast.json index 3024fd03e..f2b877cd1 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/kwargs/disable_autocast.json +++ b/fastvideo/tests/api/config_snapshot_goldens/config/disable_autocast.json @@ -321,6 +321,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -336,6 +340,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -424,14 +429,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/config/experimental_keys.json b/fastvideo/tests/api/config_snapshot_goldens/config/experimental_keys.json index 397922720..ef49fb4df 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/config/experimental_keys.json +++ b/fastvideo/tests/api/config_snapshot_goldens/config/experimental_keys.json @@ -321,6 +321,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -336,6 +340,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -431,14 +436,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/config/ltx2_refine_lora_disabled.json b/fastvideo/tests/api/config_snapshot_goldens/config/ltx2_refine_lora_disabled.json index a7e147fd4..62f5c05aa 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/config/ltx2_refine_lora_disabled.json +++ b/fastvideo/tests/api/config_snapshot_goldens/config/ltx2_refine_lora_disabled.json @@ -300,6 +300,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -315,6 +319,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -403,14 +408,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": "", "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/config/ltx2_refine_lora_path.json b/fastvideo/tests/api/config_snapshot_goldens/config/ltx2_refine_lora_path.json index 69c59c336..0f22bc579 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/config/ltx2_refine_lora_path.json +++ b/fastvideo/tests/api/config_snapshot_goldens/config/ltx2_refine_lora_path.json @@ -300,6 +300,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -315,6 +319,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -403,14 +408,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": "/checkpoints/refine_lora.safetensors", "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/config/ltx2_vae_tile_sizes.json b/fastvideo/tests/api/config_snapshot_goldens/config/ltx2_vae_tile_sizes.json index 2fa548055..69cc9c2a3 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/config/ltx2_vae_tile_sizes.json +++ b/fastvideo/tests/api/config_snapshot_goldens/config/ltx2_vae_tile_sizes.json @@ -300,6 +300,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -315,6 +319,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -403,14 +408,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/config/master_port_refine_enabled_boundary_ratio.json b/fastvideo/tests/api/config_snapshot_goldens/config/master_port_refine_enabled_boundary_ratio.json index 0507580ac..019154b48 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/config/master_port_refine_enabled_boundary_ratio.json +++ b/fastvideo/tests/api/config_snapshot_goldens/config/master_port_refine_enabled_boundary_ratio.json @@ -321,6 +321,9 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -336,6 +339,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -424,14 +428,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, "enabled": true, - "guidance_scale": null, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/config/pipeline_config_json_path.json b/fastvideo/tests/api/config_snapshot_goldens/config/pipeline_config_json_path.json index 109b8f0d6..5afd20451 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/config/pipeline_config_json_path.json +++ b/fastvideo/tests/api/config_snapshot_goldens/config/pipeline_config_json_path.json @@ -349,6 +349,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -364,6 +368,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -453,14 +458,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/kwargs/torch_compile_kwargs_typed_and_extra.json b/fastvideo/tests/api/config_snapshot_goldens/config/torch_compile_kwargs_typed_and_extra.json similarity index 97% rename from fastvideo/tests/api/config_snapshot_goldens/kwargs/torch_compile_kwargs_typed_and_extra.json rename to fastvideo/tests/api/config_snapshot_goldens/config/torch_compile_kwargs_typed_and_extra.json index d6317c572..40cea3034 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/kwargs/torch_compile_kwargs_typed_and_extra.json +++ b/fastvideo/tests/api/config_snapshot_goldens/config/torch_compile_kwargs_typed_and_extra.json @@ -321,6 +321,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -336,6 +340,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -428,14 +433,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/kwargs/typed_offload_and_parallelism.json b/fastvideo/tests/api/config_snapshot_goldens/config/typed_offload_and_parallelism.json similarity index 97% rename from fastvideo/tests/api/config_snapshot_goldens/kwargs/typed_offload_and_parallelism.json rename to fastvideo/tests/api/config_snapshot_goldens/config/typed_offload_and_parallelism.json index 0c4bc50fe..327f9f1ed 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/kwargs/typed_offload_and_parallelism.json +++ b/fastvideo/tests/api/config_snapshot_goldens/config/typed_offload_and_parallelism.json @@ -320,6 +320,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -335,6 +339,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -423,14 +428,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/config/vae_tiling_typed.json b/fastvideo/tests/api/config_snapshot_goldens/config/vae_tiling_typed.json index 504f1514f..54605cc41 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/config/vae_tiling_typed.json +++ b/fastvideo/tests/api/config_snapshot_goldens/config/vae_tiling_typed.json @@ -294,6 +294,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -309,6 +313,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -397,14 +402,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/environment/attention_backend.json b/fastvideo/tests/api/config_snapshot_goldens/environment/attention_backend.json index 023c6b062..6374444b7 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/environment/attention_backend.json +++ b/fastvideo/tests/api/config_snapshot_goldens/environment/attention_backend.json @@ -327,6 +327,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -342,6 +346,7 @@ "backend": "TORCH_SDPA", "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -430,14 +435,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/environment/explicit_argument_wins_over_attention_backend.json b/fastvideo/tests/api/config_snapshot_goldens/environment/explicit_argument_wins_over_attention_backend.json index 0961e7c10..1e5f23349 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/environment/explicit_argument_wins_over_attention_backend.json +++ b/fastvideo/tests/api/config_snapshot_goldens/environment/explicit_argument_wins_over_attention_backend.json @@ -321,6 +321,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -336,6 +340,7 @@ "backend": "FLASH_ATTN", "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -424,14 +429,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/environment/inference_torch_compile.json b/fastvideo/tests/api/config_snapshot_goldens/environment/inference_torch_compile.json index ec9a588a6..4ce2ead10 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/environment/inference_torch_compile.json +++ b/fastvideo/tests/api/config_snapshot_goldens/environment/inference_torch_compile.json @@ -326,6 +326,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -341,6 +345,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -429,14 +434,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/environment/vae_parallel.json b/fastvideo/tests/api/config_snapshot_goldens/environment/vae_parallel.json index 5c93b8327..a17087131 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/environment/vae_parallel.json +++ b/fastvideo/tests/api/config_snapshot_goldens/environment/vae_parallel.json @@ -323,6 +323,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.video_decode_backend": "h3-vae", @@ -336,6 +340,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -424,14 +429,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__DreamX-World-5B-Cam-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__DreamX-World-5B-Cam-Diffusers.json index 9c45d73cf..dffb6f5a8 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__DreamX-World-5B-Cam-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__DreamX-World-5B-Cam-Diffusers.json @@ -388,6 +388,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -403,6 +407,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -491,14 +496,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__DreamX-World-5B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__DreamX-World-5B-Diffusers.json index c02918a05..6660f0bcc 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__DreamX-World-5B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__DreamX-World-5B-Diffusers.json @@ -436,6 +436,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -451,6 +455,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -544,14 +549,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastHunyuan-diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastHunyuan-diffusers.json index b02828cac..aed9a49d0 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastHunyuan-diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastHunyuan-diffusers.json @@ -349,6 +349,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -364,6 +368,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -453,14 +458,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastVideo-FastH3-8-Step-V2.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastVideo-FastH3-8-Step-V2.json index a2febd6ea..ed78ef9ba 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastVideo-FastH3-8-Step-V2.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastVideo-FastH3-8-Step-V2.json @@ -429,6 +429,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -444,6 +448,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -532,14 +537,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastVideo-Minimax-FastH3-Preview-v0.2.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastVideo-Minimax-FastH3-Preview-v0.2.json index 43561d37a..9b083fdac 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastVideo-Minimax-FastH3-Preview-v0.2.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastVideo-Minimax-FastH3-Preview-v0.2.json @@ -429,6 +429,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -444,6 +448,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -532,14 +537,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.1-T2V-1.3B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.1-T2V-1.3B-Diffusers.json index c1903eaa9..493688f30 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.1-T2V-1.3B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.1-T2V-1.3B-Diffusers.json @@ -330,6 +330,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -345,6 +349,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -437,14 +442,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.1-T2V-14B-480P-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.1-T2V-14B-480P-Diffusers.json index ca5b98e9b..fd838b585 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.1-T2V-14B-480P-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.1-T2V-14B-480P-Diffusers.json @@ -330,6 +330,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -345,6 +349,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -437,14 +442,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.2-TI2V-5B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.2-TI2V-5B-Diffusers.json index 3a0449d16..af80dba75 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.2-TI2V-5B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.2-TI2V-5B-Diffusers.json @@ -331,6 +331,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -346,6 +350,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -438,14 +443,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.2-TI2V-5B-FullAttn-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.2-TI2V-5B-FullAttn-Diffusers.json index ea5a95ce4..012eda106 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.2-TI2V-5B-FullAttn-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__FastWan2.2-TI2V-5B-FullAttn-Diffusers.json @@ -331,6 +331,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -346,6 +350,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -438,14 +443,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__GEN3C-Cosmos-7B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__GEN3C-Cosmos-7B-Diffusers.json index 2091e2a58..0a669ad86 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__GEN3C-Cosmos-7B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__GEN3C-Cosmos-7B-Diffusers.json @@ -352,6 +352,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -367,6 +371,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -455,14 +460,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__HY-WorldPlay-Bidirectional-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__HY-WorldPlay-Bidirectional-Diffusers.json index 16f95f503..4184182c9 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__HY-WorldPlay-Bidirectional-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__HY-WorldPlay-Bidirectional-Diffusers.json @@ -406,6 +406,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -421,6 +425,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -510,14 +515,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__HunyuanGameCraft-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__HunyuanGameCraft-Diffusers.json index c0b142562..b87c98eb4 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__HunyuanGameCraft-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__HunyuanGameCraft-Diffusers.json @@ -359,6 +359,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -374,6 +378,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -463,14 +468,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX-2.3-Distilled-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX-2.3-Distilled-Diffusers.json index e5c451cf2..f4e9473d0 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX-2.3-Distilled-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX-2.3-Distilled-Diffusers.json @@ -300,6 +300,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -315,6 +319,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -403,14 +408,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2-Diffusers.json index 045ac0f16..49dc1242e 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2-Diffusers.json @@ -300,6 +300,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -315,6 +319,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -403,14 +408,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2-Distilled-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2-Distilled-Diffusers.json index 542e8b142..ed65b43e6 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2-Distilled-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2-Distilled-Diffusers.json @@ -300,6 +300,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -315,6 +319,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -403,14 +408,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2-base.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2-base.json index 467ea4046..ceff7624c 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2-base.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2-base.json @@ -300,6 +300,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -315,6 +319,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -403,14 +408,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2.3-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2.3-Diffusers.json index 65aa2531d..2b446a479 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2.3-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2.3-Diffusers.json @@ -300,6 +300,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -315,6 +319,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -403,14 +408,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2.3-Distilled-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2.3-Distilled-Diffusers.json index 7f137d333..7a0e1dce4 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2.3-Distilled-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2.3-Distilled-Diffusers.json @@ -300,6 +300,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -315,6 +319,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -403,14 +408,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2.3-base.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2.3-base.json index 646d47dbd..2480240c0 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2.3-base.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LTX2.3-base.json @@ -300,6 +300,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -315,6 +319,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -403,14 +408,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LingBot-Video-Dense-1.3B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LingBot-Video-Dense-1.3B-Diffusers.json index 881abde1a..e51e5d3e1 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LingBot-Video-Dense-1.3B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LingBot-Video-Dense-1.3B-Diffusers.json @@ -328,6 +328,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -343,6 +347,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -431,14 +436,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LingBot-Video-MoE-30B-A3B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LingBot-Video-MoE-30B-A3B-Diffusers.json index 959eaa915..bf6fbfede 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LingBot-Video-MoE-30B-A3B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LingBot-Video-MoE-30B-A3B-Diffusers.json @@ -328,6 +328,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -343,6 +347,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -431,14 +436,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LingBot-World-Base-Cam-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LingBot-World-Base-Cam-Diffusers.json index 3c6cf14e5..611cf9a1f 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LingBot-World-Base-Cam-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LingBot-World-Base-Cam-Diffusers.json @@ -352,6 +352,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -367,6 +371,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -455,14 +460,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LongCat-Video-I2V-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LongCat-Video-I2V-Diffusers.json index edd469360..871d41159 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LongCat-Video-I2V-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LongCat-Video-I2V-Diffusers.json @@ -320,6 +320,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -335,6 +339,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -423,14 +428,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LongCat-Video-T2V-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LongCat-Video-T2V-Diffusers.json index fc705e946..566f4af18 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LongCat-Video-T2V-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LongCat-Video-T2V-Diffusers.json @@ -320,6 +320,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -335,6 +339,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -423,14 +428,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LongCat-Video-VC-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LongCat-Video-VC-Diffusers.json index aaa0c9b8f..422b13519 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LongCat-Video-VC-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__LongCat-Video-VC-Diffusers.json @@ -320,6 +320,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -335,6 +339,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -423,14 +428,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__MMAudio-large-44k-v2-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__MMAudio-large-44k-v2-Diffusers.json index 1d05b442f..bd6153dc6 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__MMAudio-large-44k-v2-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__MMAudio-large-44k-v2-Diffusers.json @@ -348,6 +348,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -363,6 +367,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -451,14 +456,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-Base-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-Base-Diffusers.json index 2ae3d53ac..85fec65cf 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-Base-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-Base-Diffusers.json @@ -415,6 +415,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -430,6 +434,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -522,14 +527,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-Base-Distilled-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-Base-Distilled-Diffusers.json index c5dbffcd5..4dfbd2343 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-Base-Distilled-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-Base-Distilled-Diffusers.json @@ -415,6 +415,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -430,6 +434,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -522,14 +527,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-GTA-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-GTA-Diffusers.json index 8e465cae2..4f7c18811 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-GTA-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-GTA-Diffusers.json @@ -415,6 +415,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -430,6 +434,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -522,14 +527,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-GTA-Distilled-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-GTA-Distilled-Diffusers.json index 8c01801e8..ea3c04958 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-GTA-Distilled-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-GTA-Distilled-Diffusers.json @@ -415,6 +415,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -430,6 +434,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -522,14 +527,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-TempleRun-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-TempleRun-Diffusers.json index e08371157..21ca3ffe4 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-TempleRun-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-TempleRun-Diffusers.json @@ -415,6 +415,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -430,6 +434,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -522,14 +527,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-TempleRun-Distilled-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-TempleRun-Distilled-Diffusers.json index 2470cdf5b..7be07eb3d 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-TempleRun-Distilled-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-2.0-TempleRun-Distilled-Diffusers.json @@ -415,6 +415,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -430,6 +434,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -522,14 +527,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-3.0-Base-Distilled-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-3.0-Base-Distilled-Diffusers.json index e84a971c9..46f85ff19 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-3.0-Base-Distilled-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Matrix-Game-3.0-Base-Distilled-Diffusers.json @@ -374,6 +374,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -389,6 +393,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -477,14 +482,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__SFWan2.2-I2V-A14B-Preview-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__SFWan2.2-I2V-A14B-Preview-Diffusers.json index d51f0b9af..fda2025e1 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__SFWan2.2-I2V-A14B-Preview-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__SFWan2.2-I2V-A14B-Preview-Diffusers.json @@ -341,6 +341,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -356,6 +360,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -453,14 +458,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Wan2.1-VSA-T2V-14B-720P-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Wan2.1-VSA-T2V-14B-720P-Diffusers.json index f58847010..f69f39243 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Wan2.1-VSA-T2V-14B-720P-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__Wan2.1-VSA-T2V-14B-720P-Diffusers.json @@ -321,6 +321,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -336,6 +340,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -424,14 +429,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__stable-audio-open-1.0-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__stable-audio-open-1.0-Diffusers.json index 0fa19ee71..040a27692 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__stable-audio-open-1.0-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__stable-audio-open-1.0-Diffusers.json @@ -176,6 +176,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -191,6 +195,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -277,14 +282,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__stable-audio-open-small-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__stable-audio-open-small-Diffusers.json index f6cb88cb1..a6a88b747 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__stable-audio-open-small-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/FastVideo__stable-audio-open-small-Diffusers.json @@ -176,6 +176,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -191,6 +195,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -277,14 +282,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/IRMChen__Wan2.1-Fun-1.3B-Control-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/IRMChen__Wan2.1-Fun-1.3B-Control-Diffusers.json index 066c69951..5ab398873 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/IRMChen__Wan2.1-Fun-1.3B-Control-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/IRMChen__Wan2.1-Fun-1.3B-Control-Diffusers.json @@ -355,6 +355,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -370,6 +374,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -458,14 +463,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/KyleShao__Cosmos-Predict2.5-2B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/KyleShao__Cosmos-Predict2.5-2B-Diffusers.json index 10f29036f..97bd7380c 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/KyleShao__Cosmos-Predict2.5-2B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/KyleShao__Cosmos-Predict2.5-2B-Diffusers.json @@ -317,6 +317,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -332,6 +336,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -420,14 +425,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/Lightricks__LTX-2.3.json b/fastvideo/tests/api/config_snapshot_goldens/registry/Lightricks__LTX-2.3.json index 1d44b0348..205d580ac 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/Lightricks__LTX-2.3.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/Lightricks__LTX-2.3.json @@ -300,6 +300,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -315,6 +319,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -403,14 +408,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/Lightricks__LTX-2.json b/fastvideo/tests/api/config_snapshot_goldens/registry/Lightricks__LTX-2.json index 5a15b4c6d..bcfb8ed90 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/Lightricks__LTX-2.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/Lightricks__LTX-2.json @@ -300,6 +300,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -315,6 +319,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -403,14 +408,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/MiniMaxAI__MiniMax-H3.json b/fastvideo/tests/api/config_snapshot_goldens/registry/MiniMaxAI__MiniMax-H3.json index ba0d99d61..12036c38b 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/MiniMaxAI__MiniMax-H3.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/MiniMaxAI__MiniMax-H3.json @@ -429,6 +429,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -444,6 +448,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -532,14 +537,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/Tongyi-MAI__Z-Image-Turbo.json b/fastvideo/tests/api/config_snapshot_goldens/registry/Tongyi-MAI__Z-Image-Turbo.json index 6b064fa6b..3562b819f 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/Tongyi-MAI__Z-Image-Turbo.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/Tongyi-MAI__Z-Image-Turbo.json @@ -281,6 +281,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -296,6 +300,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -384,14 +389,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-I2V-14B-480P-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-I2V-14B-480P-Diffusers.json index 6a7bb64fa..93dec675d 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-I2V-14B-480P-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-I2V-14B-480P-Diffusers.json @@ -355,6 +355,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -370,6 +374,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -458,14 +463,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-I2V-14B-720P-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-I2V-14B-720P-Diffusers.json index 22cddd381..c0df23ccd 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-I2V-14B-720P-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-I2V-14B-720P-Diffusers.json @@ -355,6 +355,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -370,6 +374,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -458,14 +463,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-T2V-1.3B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-T2V-1.3B-Diffusers.json index 208dc635c..7ec546182 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-T2V-1.3B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-T2V-1.3B-Diffusers.json @@ -321,6 +321,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -336,6 +340,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -424,14 +429,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-T2V-14B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-T2V-14B-Diffusers.json index 01ebf35ce..deab0d45f 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-T2V-14B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.1-T2V-14B-Diffusers.json @@ -321,6 +321,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -336,6 +340,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -424,14 +429,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.2-I2V-A14B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.2-I2V-A14B-Diffusers.json index 966fac4db..cc844423a 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.2-I2V-A14B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.2-I2V-A14B-Diffusers.json @@ -356,6 +356,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -371,6 +375,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -459,14 +464,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.2-T2V-A14B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.2-T2V-A14B-Diffusers.json index 64315414d..85dd28973 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.2-T2V-A14B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.2-T2V-A14B-Diffusers.json @@ -333,6 +333,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -348,6 +352,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -441,14 +446,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.2-TI2V-5B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.2-TI2V-5B-Diffusers.json index 80321a01a..465969dbc 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.2-TI2V-5B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/Wan-AI__Wan2.2-TI2V-5B-Diffusers.json @@ -322,6 +322,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -337,6 +341,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -425,14 +430,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.1-dev.json b/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.1-dev.json index 906166db6..86a9f5bdb 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.1-dev.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.1-dev.json @@ -341,6 +341,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -356,6 +360,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -445,14 +450,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.2-dev.json b/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.2-dev.json index 6c1b4d49f..9d40afc9a 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.2-dev.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.2-dev.json @@ -249,6 +249,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -264,6 +268,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -352,14 +357,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.2-klein-4B.json b/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.2-klein-4B.json index a589de4d0..6a76c2faf 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.2-klein-4B.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.2-klein-4B.json @@ -286,6 +286,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -301,6 +305,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -389,14 +394,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.2-klein-9B.json b/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.2-klein-9B.json index c6480c185..d5cbe5678 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.2-klein-9B.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/black-forest-labs__FLUX.2-klein-9B.json @@ -286,6 +286,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -301,6 +305,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -389,14 +394,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/decart-ai__Lucy-Edit-1.1-Dev.json b/fastvideo/tests/api/config_snapshot_goldens/registry/decart-ai__Lucy-Edit-1.1-Dev.json index 982a44096..3ffc26329 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/decart-ai__Lucy-Edit-1.1-Dev.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/decart-ai__Lucy-Edit-1.1-Dev.json @@ -386,6 +386,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -401,6 +405,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -489,14 +494,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/decart-ai__Lucy-Edit-Dev.json b/fastvideo/tests/api/config_snapshot_goldens/registry/decart-ai__Lucy-Edit-Dev.json index 5512b8d66..6be7b3131 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/decart-ai__Lucy-Edit-Dev.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/decart-ai__Lucy-Edit-Dev.json @@ -386,6 +386,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -401,6 +405,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -489,14 +494,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-480p_i2v_step_distilled.json b/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-480p_i2v_step_distilled.json index 403a8016d..c56c41ffd 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-480p_i2v_step_distilled.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-480p_i2v_step_distilled.json @@ -358,6 +358,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -373,6 +377,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -462,14 +467,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-480p_t2v.json b/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-480p_t2v.json index b9335f0b4..bbbd67877 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-480p_t2v.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-480p_t2v.json @@ -358,6 +358,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -373,6 +377,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -462,14 +467,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-720p_i2v_distilled.json b/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-720p_i2v_distilled.json index 8f4e5e998..aa1421801 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-720p_i2v_distilled.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-720p_i2v_distilled.json @@ -358,6 +358,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -373,6 +377,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -462,14 +467,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-720p_t2v.json b/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-720p_t2v.json index 8cdc905d1..319f4e025 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-720p_t2v.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo-1.5-Diffusers-720p_t2v.json @@ -358,6 +358,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -373,6 +377,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -462,14 +467,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo.json b/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo.json index 6252832a5..b89f5dc24 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/hunyuanvideo-community__HunyuanVideo.json @@ -349,6 +349,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -364,6 +368,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -453,14 +458,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-I2V-Lite-5s-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-I2V-Lite-5s-Diffusers.json index 24e59d4e5..7fb27721f 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-I2V-Lite-5s-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-I2V-Lite-5s-Diffusers.json @@ -354,6 +354,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -369,6 +373,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -458,14 +463,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-I2V-Pro-distilled-5s-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-I2V-Pro-distilled-5s-Diffusers.json index 7a745ceee..69d5ae039 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-I2V-Pro-distilled-5s-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-I2V-Pro-distilled-5s-Diffusers.json @@ -354,6 +354,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -369,6 +373,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -458,14 +463,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-I2V-Pro-sft-5s-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-I2V-Pro-sft-5s-Diffusers.json index 7d5abb656..1ff540278 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-I2V-Pro-sft-5s-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-I2V-Pro-sft-5s-Diffusers.json @@ -354,6 +354,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -369,6 +373,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -458,14 +463,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers.json index 63b65a3dd..94c7c87c9 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers.json @@ -354,6 +354,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -369,6 +373,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -458,14 +463,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers.json index 526910329..22ad48950 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers.json @@ -354,6 +354,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -369,6 +373,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -458,14 +463,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Pro-distilled-5s-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Pro-distilled-5s-Diffusers.json index fa39de173..de085a957 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Pro-distilled-5s-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Pro-distilled-5s-Diffusers.json @@ -354,6 +354,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -369,6 +373,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -458,14 +463,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Pro-sft-5s-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Pro-sft-5s-Diffusers.json index a002e881b..ba5e2dd9f 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Pro-sft-5s-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/kandinskylab__Kandinsky-5.0-T2V-Pro-sft-5s-Diffusers.json @@ -354,6 +354,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -369,6 +373,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -458,14 +463,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/loayrashid__TurboWan2.1-T2V-1.3B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/loayrashid__TurboWan2.1-T2V-1.3B-Diffusers.json index 70680d01a..71f57bbcb 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/loayrashid__TurboWan2.1-T2V-1.3B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/loayrashid__TurboWan2.1-T2V-1.3B-Diffusers.json @@ -320,6 +320,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -335,6 +339,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -423,14 +428,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/loayrashid__TurboWan2.1-T2V-14B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/loayrashid__TurboWan2.1-T2V-14B-Diffusers.json index 29dce7714..bcd9138de 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/loayrashid__TurboWan2.1-T2V-14B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/loayrashid__TurboWan2.1-T2V-14B-Diffusers.json @@ -320,6 +320,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -335,6 +339,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -423,14 +428,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/loayrashid__TurboWan2.2-I2V-A14B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/loayrashid__TurboWan2.2-I2V-A14B-Diffusers.json index 961666ad5..2b719bd82 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/loayrashid__TurboWan2.2-I2V-A14B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/loayrashid__TurboWan2.2-I2V-A14B-Diffusers.json @@ -355,6 +355,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -370,6 +374,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -458,14 +463,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_bidirectional_zelda.json b/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_bidirectional_zelda.json index 88b4c8302..b52fabf0a 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_bidirectional_zelda.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_bidirectional_zelda.json @@ -415,6 +415,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -430,6 +434,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -522,14 +527,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_causal_zelda.json b/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_causal_zelda.json index e5bf63cf3..f7c95eaa2 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_causal_zelda.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_causal_zelda.json @@ -415,6 +415,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -430,6 +434,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -522,14 +527,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_longtuning_distilled_zelda.json b/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_longtuning_distilled_zelda.json index 7bf39331c..1e515289a 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_longtuning_distilled_zelda.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_longtuning_distilled_zelda.json @@ -415,6 +415,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -430,6 +434,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -522,14 +527,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_sf_distilled_zelda.json b/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_sf_distilled_zelda.json index 312d82193..83a4b929c 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_sf_distilled_zelda.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_sf_distilled_zelda.json @@ -415,6 +415,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -430,6 +434,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -522,14 +527,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_sf_distilled_zelda_1k_steps.json b/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_sf_distilled_zelda_1k_steps.json index 637e6e81a..05b1886b3 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_sf_distilled_zelda_1k_steps.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/mignonjia__mg_sf_distilled_zelda_1k_steps.json @@ -415,6 +415,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -430,6 +434,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -522,14 +527,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/nvidia__Cosmos-Predict2-2B-Video2World.json b/fastvideo/tests/api/config_snapshot_goldens/registry/nvidia__Cosmos-Predict2-2B-Video2World.json index 3ac3dc945..929869c18 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/nvidia__Cosmos-Predict2-2B-Video2World.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/nvidia__Cosmos-Predict2-2B-Video2World.json @@ -324,6 +324,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -339,6 +343,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -427,14 +432,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/nvidia__Cosmos-Predict2.5-14B.json b/fastvideo/tests/api/config_snapshot_goldens/registry/nvidia__Cosmos-Predict2.5-14B.json index 9d23f6898..1fc254a83 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/nvidia__Cosmos-Predict2.5-14B.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/nvidia__Cosmos-Predict2.5-14B.json @@ -317,6 +317,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -332,6 +336,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -420,14 +425,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/rand0nmr__SFWan2.2-T2V-A14B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/rand0nmr__SFWan2.2-T2V-A14B-Diffusers.json index 94e567d11..161db9c7f 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/rand0nmr__SFWan2.2-T2V-A14B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/rand0nmr__SFWan2.2-T2V-A14B-Diffusers.json @@ -341,6 +341,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -356,6 +360,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -453,14 +458,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/robbyant__lingbot-world-v2-14b-causal-fast.json b/fastvideo/tests/api/config_snapshot_goldens/registry/robbyant__lingbot-world-v2-14b-causal-fast.json index 809676afe..46498e45a 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/robbyant__lingbot-world-v2-14b-causal-fast.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/robbyant__lingbot-world-v2-14b-causal-fast.json @@ -331,6 +331,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -346,6 +350,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -434,14 +439,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/stabilityai__stable-diffusion-3.5-medium.json b/fastvideo/tests/api/config_snapshot_goldens/registry/stabilityai__stable-diffusion-3.5-medium.json index e0e410858..056ad8573 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/stabilityai__stable-diffusion-3.5-medium.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/stabilityai__stable-diffusion-3.5-medium.json @@ -424,6 +424,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -439,6 +443,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -529,14 +534,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/weizhou03__HunyuanVideo-1.5-Diffusers-1080p-2SR.json b/fastvideo/tests/api/config_snapshot_goldens/registry/weizhou03__HunyuanVideo-1.5-Diffusers-1080p-2SR.json index 8e405d214..d916c6e97 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/weizhou03__HunyuanVideo-1.5-Diffusers-1080p-2SR.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/weizhou03__HunyuanVideo-1.5-Diffusers-1080p-2SR.json @@ -379,6 +379,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -394,6 +398,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -483,14 +488,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/weizhou03__HunyuanVideo-1.5-Diffusers-1080p.json b/fastvideo/tests/api/config_snapshot_goldens/registry/weizhou03__HunyuanVideo-1.5-Diffusers-1080p.json index b2d7b807e..3a4094931 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/weizhou03__HunyuanVideo-1.5-Diffusers-1080p.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/weizhou03__HunyuanVideo-1.5-Diffusers-1080p.json @@ -379,6 +379,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -394,6 +398,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -483,14 +488,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/weizhou03__Wan2.1-Fun-1.3B-InP-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/weizhou03__Wan2.1-Fun-1.3B-InP-Diffusers.json index 5edc39991..be88f6bff 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/weizhou03__Wan2.1-Fun-1.3B-InP-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/weizhou03__Wan2.1-Fun-1.3B-InP-Diffusers.json @@ -355,6 +355,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -370,6 +374,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -458,14 +463,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/wlsaidhi__SFWan2.1-T2V-1.3B-Diffusers.json b/fastvideo/tests/api/config_snapshot_goldens/registry/wlsaidhi__SFWan2.1-T2V-1.3B-Diffusers.json index adbc5c742..bc63b1bf6 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/wlsaidhi__SFWan2.1-T2V-1.3B-Diffusers.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/wlsaidhi__SFWan2.1-T2V-1.3B-Diffusers.json @@ -332,6 +332,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -347,6 +351,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -440,14 +445,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/registry/zai-org__GLM-Image.json b/fastvideo/tests/api/config_snapshot_goldens/registry/zai-org__GLM-Image.json index 33806d453..3463821c3 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/registry/zai-org__GLM-Image.json +++ b/fastvideo/tests/api/config_snapshot_goldens/registry/zai-org__GLM-Image.json @@ -310,6 +310,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -325,6 +329,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -413,14 +418,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/apps__dreamverse__serve_configs__streaming_demo.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/apps__dreamverse__serve_configs__streaming_demo.yaml.json index 50d8ec181..ffcc68dcf 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/apps__dreamverse__serve_configs__streaming_demo.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/apps__dreamverse__serve_configs__streaming_demo.yaml.json @@ -332,6 +332,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_1024x1536.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_1024x1536.yaml.json index d234fc488..c414cf164 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_1024x1536.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_1024x1536.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_1280x832.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_1280x832.yaml.json index 14687404c..6cf9e820b 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_1280x832.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_1280x832.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_512x768.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_512x768.yaml.json index 26b44f2fd..6b452ac62 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_512x768.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_512x768.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_768x1280.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_768x1280.yaml.json index 3e9a96c74..1aa0cfff6 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_768x1280.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_5s2_768x1280.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_1024x1536.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_1024x1536.yaml.json index e488c0900..a29edc88a 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_1024x1536.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_1024x1536.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_1280x832.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_1280x832.yaml.json index f95532959..35dce8891 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_1280x832.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_1280x832.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_512x768.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_512x768.yaml.json index bbeee04bd..3bdf1e62e 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_512x768.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_512x768.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_768x1280.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_768x1280.yaml.json index c33d20319..438db02c0 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_768x1280.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__i2v_8s3_768x1280.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_1024x1536.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_1024x1536.yaml.json index c2fe17b64..a2421e7ab 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_1024x1536.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_1024x1536.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_1280x832.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_1280x832.yaml.json index 4dfbd800a..79ebe7d9b 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_1280x832.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_1280x832.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_512x768.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_512x768.yaml.json index 2436c2b2d..ab6a29ad5 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_512x768.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_512x768.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_768x1280.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_768x1280.yaml.json index 0826a2c9c..3e099b4fc 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_768x1280.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_5s2_768x1280.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_1024x1536.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_1024x1536.yaml.json index d73d1384d..b326307d0 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_1024x1536.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_1024x1536.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_1280x832.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_1280x832.yaml.json index 8f66b0e9b..0c1f6145b 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_1280x832.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_1280x832.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_512x768.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_512x768.yaml.json index b9f2fccd1..412fc02bf 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_512x768.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_512x768.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_768x1280.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_768x1280.yaml.json index 365c199ce..18f1ff7bb 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_768x1280.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__inference__ltx2_3__t2v_8s3_768x1280.yaml.json @@ -334,6 +334,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fasth3.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fasth3.yaml.json index 13446a56b..e0bcba996 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fasth3.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fasth3.yaml.json @@ -427,6 +427,10 @@ "engine.compile.audio_vae_enabled": false, "engine.compile.text_encoder_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_encode": false, @@ -440,6 +444,7 @@ "backend": "VIDEO_SPARSE_ATTN_H3", "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.9, "vsa_tile_size": 64 }, @@ -528,14 +533,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fasth3_8step.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fasth3_8step.yaml.json index ee9de467c..8aa74e152 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fasth3_8step.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fasth3_8step.yaml.json @@ -427,6 +427,10 @@ "engine.compile.audio_vae_enabled": false, "engine.compile.text_encoder_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_encode": false, @@ -440,6 +444,7 @@ "backend": "VIDEO_SPARSE_ATTN_H3", "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.8, "vsa_tile_size": 64 }, @@ -528,14 +533,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fasth3_lora.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fasth3_lora.yaml.json index 31797eadb..dc4b42289 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fasth3_lora.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fasth3_lora.yaml.json @@ -429,6 +429,10 @@ "engine.compile.audio_vae_enabled": false, "engine.compile.text_encoder_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_encode": false, @@ -442,6 +446,7 @@ "backend": "FLASH_ATTN", "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -530,14 +535,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fastwan21_1_3b.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fastwan21_1_3b.yaml.json index fbab8de3b..fca1fb729 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fastwan21_1_3b.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_fastwan21_1_3b.yaml.json @@ -333,6 +333,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -348,6 +352,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.8, "vsa_tile_size": 256 }, @@ -440,14 +445,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_wan21_i2v_14b.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_wan21_i2v_14b.yaml.json index ee854032f..677d029cb 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_wan21_i2v_14b.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_wan21_i2v_14b.yaml.json @@ -362,6 +362,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -376,6 +380,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -464,14 +469,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_wan22_t2v_a14b.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_wan22_t2v_a14b.yaml.json index d69de2a97..5b9e0ddc7 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_wan22_t2v_a14b.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_wan22_t2v_a14b.yaml.json @@ -345,6 +345,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -360,6 +364,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -453,14 +458,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_wan22_ti2v_5b.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_wan22_ti2v_5b.yaml.json index 8e3bd0e1c..ae324b2ec 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_wan22_ti2v_5b.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__openai_wan22_ti2v_5b.yaml.json @@ -333,6 +333,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -348,6 +352,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -436,14 +441,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__streaming_demo.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__streaming_demo.yaml.json index 50d8ec181..ffcc68dcf 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__streaming_demo.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/examples__serving__streaming_demo.yaml.json @@ -332,6 +332,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_fasthunyuan.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_fasthunyuan.yaml.json index 1db667365..d7a95be4f 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_fasthunyuan.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_fasthunyuan.yaml.json @@ -355,6 +355,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -370,6 +374,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -459,14 +464,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_flux2_klein.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_flux2_klein.yaml.json index 1d4f9f46f..02e9c04d8 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_flux2_klein.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_flux2_klein.yaml.json @@ -297,6 +297,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -311,6 +315,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -399,14 +404,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_hunyuan.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_hunyuan.yaml.json index 83fae4dd2..89dbe1d70 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_hunyuan.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_hunyuan.yaml.json @@ -355,6 +355,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -370,6 +374,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -459,14 +464,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat.yaml.json index 5b7759b51..2a379d447 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat.yaml.json @@ -327,6 +327,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -342,6 +346,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -430,14 +435,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_distill.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_distill.yaml.json index 5e0e31a72..e037ff0da 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_distill.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_distill.yaml.json @@ -327,6 +327,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -342,6 +346,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -430,14 +435,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_i2v.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_i2v.yaml.json index b92b7033a..48a585a33 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_i2v.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_i2v.yaml.json @@ -327,6 +327,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -341,6 +345,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -429,14 +434,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_refine_fromvideo.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_refine_fromvideo.yaml.json index c9829f3ff..58f2618db 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_refine_fromvideo.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_refine_fromvideo.yaml.json @@ -336,6 +336,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -351,6 +355,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -447,14 +452,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_vc.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_vc.yaml.json index 4c86455e4..c66037170 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_vc.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_longcat_vc.yaml.json @@ -327,6 +327,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -342,6 +346,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -430,14 +435,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan.yaml.json index 4c29997fc..9e680c3f8 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan.yaml.json @@ -328,6 +328,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -343,6 +347,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -431,14 +436,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_1.3B_VMoba.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_1.3B_VMoba.yaml.json index 0b1c1bbdf..ec8622dc4 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_1.3B_VMoba.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_1.3B_VMoba.yaml.json @@ -355,6 +355,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -392,6 +396,7 @@ "temporal_topk": 3 }, "moba_config_path": "fastvideo/configs/backend/vmoba/wan_1.3B_77_480_832.json", + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -480,14 +485,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA.yaml.json index dc2b1d906..e7b302bba 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA.yaml.json @@ -327,6 +327,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -342,6 +346,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.9, "vsa_tile_size": 256 }, @@ -430,14 +435,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_14B_480P.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_14B_480P.yaml.json index 8b3b30303..9d7bcbcf8 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_14B_480P.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_14B_480P.yaml.json @@ -333,6 +333,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -348,6 +352,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.9, "vsa_tile_size": 256 }, @@ -440,14 +445,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_14B_720P.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_14B_720P.yaml.json index 4ca024395..b9d13e5ce 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_14B_720P.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_14B_720P.yaml.json @@ -333,6 +333,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -348,6 +352,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.9, "vsa_tile_size": 256 }, @@ -440,14 +445,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_1_3B.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_1_3B.yaml.json index 6606d5dfa..6edbb77a5 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_1_3B.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_1_3B.yaml.json @@ -333,6 +333,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -348,6 +352,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.8, "vsa_tile_size": 256 }, @@ -440,14 +445,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_5B_720P.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_5B_720P.yaml.json index 68a5e8e1b..13974826e 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_5B_720P.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_VSA_DMD_5B_720P.yaml.json @@ -335,6 +335,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -350,6 +354,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -442,14 +447,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_i2v.yaml.json b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_i2v.yaml.json index b835a35f4..97c5a1a32 100644 --- a/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_i2v.yaml.json +++ b/fastvideo/tests/api/config_snapshot_goldens/yaml/scripts__inference__inference_wan_i2v.yaml.json @@ -362,6 +362,10 @@ "engine.compile.text_encoder_enabled": false, "engine.compile.vae_enabled": false, "pipeline.ltx2.legacy_native_noise_order": false, + "pipeline.ltx2.refine.add_noise": true, + "pipeline.ltx2.refine.enabled": false, + "pipeline.ltx2.refine.guidance_scale": 1.0, + "pipeline.ltx2.refine.num_inference_steps": 3, "pipeline.ltx2.use_distilled_sigmas": true, "pipeline.minimax_h3.taeh3_chunk_size": 5, "pipeline.minimax_h3.vae_parallel_decode": false, @@ -376,6 +380,7 @@ "backend": null, "moba_config": {}, "moba_config_path": null, + "nvfp4_fa4": false, "vsa_sparsity": 0.0, "vsa_tile_size": 256 }, @@ -464,14 +469,14 @@ "initial_latent_path": null, "legacy_native_noise_order": false, "refine": { - "add_noise": null, + "add_noise": true, "audio_noise_path": null, - "enabled": null, - "guidance_scale": null, + "enabled": false, + "guidance_scale": 1.0, "image_crf": null, "lora_path": null, "noise_path": null, - "num_inference_steps": null, + "num_inference_steps": 3, "transformer_path": null, "video_position_offset_sec": null }, diff --git a/fastvideo/tests/api/test_compat_translation.py b/fastvideo/tests/api/test_compat_translation.py index 5eafffb38..4e66cbe06 100644 --- a/fastvideo/tests/api/test_compat_translation.py +++ b/fastvideo/tests/api/test_compat_translation.py @@ -1,52 +1,16 @@ # SPDX-License-Identifier: Apache-2.0 -"""Tests for the ``fastvideo.api.compat`` keyword and request helpers, and -for how resolution carries the typed CompileConfig and -PipelineSelection.vae_tiling surfaces promoted in PR 6. +"""Tests for the ``fastvideo.api.compat`` request helpers, and for how resolution carries the typed CompileConfig +and PipelineSelection.vae_tiling surfaces. """ from __future__ import annotations -import pytest - -from fastvideo.api.compat import ( - FROM_PRETRAINED_KWARGS, - _FROM_PRETRAINED_FIELDS, - from_pretrained_kwargs_to_config, - request_to_sampling_param, -) +from fastvideo.api.compat import request_to_sampling_param from fastvideo.api.inference_resolution import resolve_inference_config, torch_compile_kwargs from fastvideo.api.parser import parse_config from fastvideo.api.schema import CompileConfig, GenerationRequest, GeneratorConfig from fastvideo.api.sampling_param import SamplingParam -class TestFromPretrainedKwargsTranslation: - """The ``from_pretrained`` keyword ``torch_compile_kwargs={...}`` gets - split across the four first-class :class:`CompileConfig` fields and - anything unknown falls into ``extras``. Keywords outside the - ``from_pretrained`` set are rejected.""" - - def test_empty_kwargs_produces_empty_extras(self) -> None: - config = from_pretrained_kwargs_to_config( - "/models/ltx2", - {"torch_compile_kwargs": {}}, - ) - compile_config = config.engine.compile - assert compile_config.extras == {} - assert compile_config.backend is None - - def test_other_keyword_points_to_from_config(self) -> None: - with pytest.raises(TypeError, match=r"does not accept ltx2_vae_tiling\. It accepts only these keywords: " - r"(.|\n)*VideoGenerator\.from_config"): - from_pretrained_kwargs_to_config("/models/ltx2", {"ltx2_vae_tiling": True}) - - def test_keyword_fields_are_the_from_pretrained_keywords(self) -> None: - """Every keyword with a schema field sets one field; ``torch_compile_kwargs`` and ``nvfp4_fa4`` have none.""" - assert set(_FROM_PRETRAINED_FIELDS) == FROM_PRETRAINED_KWARGS - {"torch_compile_kwargs", "nvfp4_fa4"} - assert len(set(_FROM_PRETRAINED_FIELDS.values())) == len(_FROM_PRETRAINED_FIELDS) - config = from_pretrained_kwargs_to_config("/models/wan", {"dit_cpu_offload": False, "sp_size": 2}) - assert config.engine.offload.dit is False and config.engine.parallelism.sp_size == 2 - - class TestCompileConfigRoundTrip: """typed CompileConfig -> ``torch_compile_kwargs(resolved_config)`` reconstruction drops ``None`` typed fields and merges ``extras``.""" diff --git a/fastvideo/tests/api/test_inference_resolution_steps.py b/fastvideo/tests/api/test_inference_resolution_steps.py index a3e108a0e..b7f6a7893 100644 --- a/fastvideo/tests/api/test_inference_resolution_steps.py +++ b/fastvideo/tests/api/test_inference_resolution_steps.py @@ -110,16 +110,23 @@ def test_environment_takes_precedence_over_model_defaults(): "fill_vae_parallel_from_env", "fill_pipeline_config_defaults[WanT2V480PConfig]", "copy_refine_preset_overrides", + "fill_ltx2_refine_from_checkpoint", + "fill_dmd_schedule_from_checkpoint", "derive_parallel_sizes", "derive_vae_tiling_from_ltx2_tile_sizes", "fill_vae_tiling_default[WanT2V480PConfig]", "load_moba_config", + "apply_unified_memory_offload_policy", + "fill_lazy_module_load", + "apply_mps_offload_policy", + "apply_layerwise_offload_conflicts", "validate_unsupported_fields", "validate_experimental_keys", "validate_lora_strength", "validate_attention_backend", "validate_vae_parallel_decode_strategy", "warn_deprecated_environment_variables", + "apply_nvfp4_fa4_env", "validate_parallel_sizes", "derive_num_gpus_from_parallel_sizes", "fill_runtime_defaults", diff --git a/fastvideo/tests/api/test_ltx2_gpu_pool_translation.py b/fastvideo/tests/api/test_ltx2_gpu_pool_translation.py index 15007acd0..20ea7ee53 100644 --- a/fastvideo/tests/api/test_ltx2_gpu_pool_translation.py +++ b/fastvideo/tests/api/test_ltx2_gpu_pool_translation.py @@ -9,7 +9,7 @@ from __future__ import annotations import pytest -from fastvideo.api.compat import from_pretrained_kwargs_to_config +from fastvideo.api.compat import normalize_generator_config from fastvideo.api.inference_resolution import resolve_inference_config, torch_compile_kwargs from fastvideo.tests.api.config_snapshot import isolated_environment @@ -144,17 +144,21 @@ class TestCompileExtrasPreserved: round-trip through ``CompileConfig.extras``.""" def test_extras_preserved(self) -> None: - kwargs = { - "enable_torch_compile": True, - "torch_compile_kwargs": { - "backend": "inductor", - "options": { - "triton.cudagraphs": False - }, - "disable": False, + config = normalize_generator_config({ + "model_path": "FastVideo/LTX2-Distilled-Diffusers", + "engine": { + "compile": { + "enabled": True, + "backend": "inductor", + "extras": { + "options": { + "triton.cudagraphs": False + }, + "disable": False, + }, + } }, - } - config = from_pretrained_kwargs_to_config("FastVideo/LTX2-Distilled-Diffusers", kwargs) + }) assert config.engine.compile.backend == "inductor" assert config.engine.compile.extras == { "options": { diff --git a/fastvideo/tests/api/test_parser.py b/fastvideo/tests/api/test_parser.py index 854df3743..be34ae502 100644 --- a/fastvideo/tests/api/test_parser.py +++ b/fastvideo/tests/api/test_parser.py @@ -140,6 +140,7 @@ def test_load_run_config_supports_yaml_roundtrip(tmp_path) -> None: "vsa_tile_size": None, "moba_config_path": None, "moba_config": None, + "nvfp4_fa4": False, }, "precision": { "dit": None, diff --git a/fastvideo/tests/api/test_resolved_runtime_config.py b/fastvideo/tests/api/test_resolved_runtime_config.py index bc5fe59a8..68d0438cf 100644 --- a/fastvideo/tests/api/test_resolved_runtime_config.py +++ b/fastvideo/tests/api/test_resolved_runtime_config.py @@ -6,6 +6,7 @@ import json import pickle from contextlib import nullcontext from types import SimpleNamespace +from unittest.mock import patch import pytest import torch @@ -20,8 +21,6 @@ from fastvideo.api.training_schema import (PreprocessRunConfig, TrainingRunConfi from fastvideo.tests.api.config_snapshot import isolated_environment WAN_T2V = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers" -LTX2 = "FastVideo/LTX2-Distilled-Diffusers" -FASTH3 = "FastVideo/FastVideo-FastH3-8-Step-V2" def _resolve(raw, env_values=None): @@ -29,6 +28,12 @@ def _resolve(raw, env_values=None): return resolve_inference_config(raw) +def _resolve_with_device_policy(raw): + """Resolve inside the isolation with the device-policy steps on, which the isolation turns off like downloads.""" + with isolated_environment(), patch.object(device_policy, "APPLY_DEVICE_POLICY", True): + return resolve_inference_config(raw) + + def test_resolution_builds_and_freezes_the_pipeline_config_once(monkeypatch): from fastvideo.configs.pipelines.base import PipelineConfig @@ -100,46 +105,56 @@ def test_override_of_a_typed_home_leaves_the_pipeline_config_unchanged(): assert overridden.provenance("pipeline.dmd_denoising_steps").source == "checkpoint:test" -def test_unified_memory_policy_returns_recorded_overrides_once(monkeypatch): - resolved = _resolve({"model_path": WAN_T2V, "engine": {"offload": {"dit_layerwise": False}}}) - monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", lambda device_id: device_id == 1) +def test_unified_memory_policy_is_decided_during_resolution(monkeypatch): + monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", lambda device_id: True) monkeypatch.setattr("fastvideo.platforms.current_platform.get_device_name", lambda device_id: "NVIDIA GB10") monkeypatch.setattr("fastvideo.platforms.current_platform.is_mps", lambda: False) - decided = device_policy.finalize_device_offload_policy(resolved, 1) - again = device_policy.finalize_device_offload_policy(decided, 1) + resolved = _resolve_with_device_policy({"model_path": WAN_T2V, "engine": {"offload": {"dit_layerwise": False}}}) - assert [source for source, _ in decided.override_log] == [ - "device_policy:unified_memory", - "device_policy:lazy_module_load", + device_decisions = [(source, values) for source, values in resolved.decisions + if source in ("apply_unified_memory_offload_policy", "fill_lazy_module_load")] + assert device_decisions == [ + ("apply_unified_memory_offload_policy", { + "engine.offload.dit": False, + "engine.offload.text_encoder": False, + "engine.offload.image_encoder": False, + "engine.offload.vae": False, + }), + ("fill_lazy_module_load", { + "engine.offload.lazy_module_load": True + }), ] - assert decided.override_log[0][1] == { + assert resolved.engine.offload.lazy_module_load is True and resolved.engine.offload.dit is False + assert resolved.override_log == () + assert resolved.provenance("engine.offload.dit").raw_value is True + + +def test_discrete_device_leaves_offload_requests_and_turns_lazy_module_load_off(monkeypatch): + monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", lambda device_id: False) + monkeypatch.setattr("fastvideo.platforms.current_platform.is_mps", lambda: False) + + resolved = _resolve_with_device_policy({"model_path": WAN_T2V, "engine": {"offload": {"dit_layerwise": False}}}) + + assert resolved.engine.offload.dit is True and resolved.engine.offload.text_encoder is True + assert resolved.engine.offload.lazy_module_load is False + assert resolved.provenance("engine.offload.lazy_module_load").source == "fill_lazy_module_load" + + +def test_layerwise_offload_turns_off_conflicting_modes(monkeypatch): + monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", lambda device_id: False) + monkeypatch.setattr("fastvideo.platforms.current_platform.is_mps", lambda: False) + + resolved = _resolve_with_device_policy({"model_path": WAN_T2V, "engine": {"use_fsdp_inference": True}}) + + assert [values for source, values in resolved.decisions if source == "apply_layerwise_offload_conflicts"] == [{ + "engine.use_fsdp_inference": False, "engine.offload.dit": False, - "engine.offload.text_encoder": False, - "engine.offload.image_encoder": False, - "engine.offload.vae": False, - } - assert decided.engine.offload.lazy_module_load is True and decided.engine.offload.dit is False - assert again.override_log == decided.override_log - assert resolved.override_log == () and resolved.engine.offload.dit is True - assert device_policy.offload_disabled_on_unified_memory(1, "text_encoder_cpu_offload") - assert not device_policy.offload_disabled_on_unified_memory(0, "text_encoder_cpu_offload") - assert device_policy.finalize_device_offload_policy(resolved, 0).engine.offload.lazy_module_load is False - - -def test_layerwise_offload_turns_off_conflicting_modes(): - resolved = _resolve({"model_path": WAN_T2V, "engine": {"use_fsdp_inference": True}}) - - decided = device_policy.resolve_device_offload_conflicts(resolved) - - assert [values for _, values in decided.override_log] == [{ - "engine.use_fsdp_inference": False - }, { - "engine.offload.dit": False }] + assert resolved.engine.use_fsdp_inference is False and resolved.engine.offload.dit is False -def test_direct_pipeline_keeps_the_policy_result_of_a_resolved_config(monkeypatch): +def test_direct_pipeline_keeps_the_resolved_config(monkeypatch): import fastvideo.pipelines.composed_pipeline_base as composed_pipeline_base from fastvideo.pipelines.composed_pipeline_base import ComposedPipelineBase @@ -159,14 +174,14 @@ def test_direct_pipeline_keeps_the_policy_result_of_a_resolved_config(monkeypatc monkeypatch.setattr(composed_pipeline_base, "get_world_group", lambda: SimpleNamespace(local_rank=2)) monkeypatch.setattr(composed_pipeline_base, "get_or_create_profiler", lambda trace_dir: profiler) monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", lambda device_id: True) + monkeypatch.setattr("fastvideo.platforms.current_platform.get_device_name", lambda device_id: "NVIDIA GB10") monkeypatch.setattr("fastvideo.platforms.current_platform.is_mps", lambda: False) - resolved = _resolve({"model_path": WAN_T2V}) + resolved = _resolve_with_device_policy({"model_path": WAN_T2V}) pipeline = _Pipeline("unused", resolved, required_config_modules=[]) - assert pipeline.resolved_config is pipeline.loaded_with - assert pipeline.resolved_config.engine.offload.lazy_module_load is True - assert resolved.engine.offload.lazy_module_load is None + assert pipeline.resolved_config is resolved and pipeline.loaded_with is resolved + assert resolved.engine.offload.lazy_module_load is True def test_pipeline_records_component_paths_and_shares_its_component_state(monkeypatch, tmp_path): @@ -255,40 +270,41 @@ def test_training_resolution_keeps_the_dit_on_the_device(): assert built["resolved_config"] is resolved -def test_ltx2_checkpoint_refine_defaults_rebind_the_pipeline_config(tmp_path): - from fastvideo.pipelines.basic.ltx2.ltx2_pipeline import LTX2Pipeline +def test_ltx2_checkpoint_refine_defaults_are_resolved_from_a_local_checkpoint(tmp_path): + (tmp_path / "transformer").mkdir() + (tmp_path / "model_index.json").write_text( + json.dumps({ + "_class_name": "LTX2Pipeline", + "_diffusers_version": "0", + "transformer": ["diffusers", "Model"], + "fastvideo_refine_lora_path": "FastVideo/LTX2-Distilled-LoRA", + "fastvideo_refine_num_inference_steps": 2, + })) - resolved = _resolve({"model_path": LTX2}) - pipeline = object.__new__(LTX2Pipeline) - pipeline.model_path = str(tmp_path) - pipeline.resolved_config = resolved - pipeline._required_config_modules = list(LTX2Pipeline._required_config_modules) - model_index = {name: ["diffusers", "Model"] for name in pipeline._required_config_modules} - pipeline._load_config = lambda model_path: { - "_class_name": "LTX2Pipeline", - "_diffusers_version": "0", - "fastvideo_refine_lora_path": "FastVideo/LTX2-Distilled-LoRA", - "fastvideo_refine_num_inference_steps": 2, - **model_index, - } + from fastvideo.pipelines.basic.ltx2.pipeline_configs import LTX2T2VConfig - pipeline.load_modules(resolved, {name: object() for name in model_index}) + resolved = _resolve({ + "model_path": str(tmp_path), + "pipeline": { + "experimental": { + "pipeline_config": LTX2T2VConfig() + } + } + }) - rebound = pipeline.resolved_config - assert rebound is not resolved - assert rebound.override_log == (("checkpoint:model_index.json", { + refine_decisions = [values for source, values in resolved.decisions if source == "fill_ltx2_refine_from_checkpoint"] + assert refine_decisions == [{ "pipeline.ltx2.refine.lora_path": "FastVideo/LTX2-Distilled-LoRA", "pipeline.ltx2.refine.num_inference_steps": 2, - "pipeline.ltx2.refine.enabled": False, - "pipeline.ltx2.refine.add_noise": True, - "pipeline.ltx2.refine.guidance_scale": 1.0, - }), ) - assert rebound.pipeline.ltx2.refine.lora_path == "FastVideo/LTX2-Distilled-LoRA" - assert rebound.pipeline.ltx2.refine.num_inference_steps == 2 - assert resolved.pipeline.ltx2.refine.num_inference_steps is None + }] + refine = resolved.pipeline.ltx2.refine + assert (refine.lora_path, refine.num_inference_steps) == ("FastVideo/LTX2-Distilled-LoRA", 2) + assert (refine.enabled, refine.add_noise, refine.guidance_scale) == (False, True, 1.0) + assert resolved.provenance("pipeline.ltx2.refine.enabled").source == "fill_runtime_defaults" -def test_minimax_h3_checkpoint_schedule_rebinds_and_reaches_the_first_forward(tmp_path, monkeypatch): +def test_minimax_h3_checkpoint_schedule_is_resolved_and_reaches_the_first_forward(tmp_path, monkeypatch): + from fastvideo.configs.pipelines.minimax_h3 import MiniMaxH3PipelineConfig from fastvideo.models.schedulers.scheduling_minimax_h3 import MiniMaxH3Scheduler from fastvideo.pipelines.basic.minimax_h3.minimax_h3_pipeline import MiniMaxH3ModularPipeline from fastvideo.pipelines.composed_pipeline_base import ComposedPipelineBase @@ -301,7 +317,24 @@ def test_minimax_h3_checkpoint_schedule_rebinds_and_reaches_the_first_forward(tm "num_inference_steps": 9, "transformer_forwards": 8, })) - resolved = _resolve({"model_path": FASTH3}) + (tmp_path / "transformer").mkdir() + (tmp_path / "model_index.json").write_text( + json.dumps({ + "_class_name": "MiniMaxH3ModularPipeline", + "_diffusers_version": "0", + "transformer": ["diffusers", "Model"], + })) + resolved = _resolve({ + "model_path": str(tmp_path), + "pipeline": { + "experimental": { + "pipeline_config": MiniMaxH3PipelineConfig() + } + } + }) + assert resolved.pipeline.dmd_denoising_steps == tuple(steps) + assert resolved.provenance("pipeline.dmd_denoising_steps").source == "fill_dmd_schedule_from_checkpoint" + pipeline = object.__new__(MiniMaxH3ModularPipeline) pipeline.model_path = str(tmp_path) pipeline.resolved_config = resolved @@ -309,7 +342,7 @@ def test_minimax_h3_checkpoint_schedule_rebinds_and_reaches_the_first_forward(tm pipeline.post_init_called = False seen = [] monkeypatch.setattr(ComposedPipelineBase, "post_init", - lambda self: (setattr(self, "post_init_called", True), self._load_checkpoint_schedule( + lambda self: (setattr(self, "post_init_called", True), self._validate_checkpoint_schedule( self.resolved_config))) monkeypatch.setattr(MiniMaxH3ModularPipeline, "_defer_denoise_modules", lambda self, args: False) monkeypatch.setattr(ComposedPipelineBase, "_stages", [], raising=False) @@ -317,12 +350,9 @@ def test_minimax_h3_checkpoint_schedule_rebinds_and_reaches_the_first_forward(tm monkeypatch.setattr(ComposedPipelineBase, "stages", property(lambda self: [lambda batch, args: seen.append(args) or batch])) - pipeline.forward(SimpleNamespace(), pipeline.resolved_config) + pipeline.forward(SimpleNamespace(), resolved) - assert seen == [pipeline.resolved_config] and pipeline.resolved_config is not resolved - assert pipeline.resolved_config.pipeline.dmd_denoising_steps == tuple(steps) - assert pipeline.resolved_config.provenance("pipeline.dmd_denoising_steps").source == ( - "checkpoint:fastvideo_inference.json") + assert seen == [resolved] and pipeline.resolved_config is resolved def test_training_root_uses_command_line_defaults_and_typed_values(tmp_path): diff --git a/fastvideo/tests/contract/test_env_policy.py b/fastvideo/tests/contract/test_env_policy.py index c84eb44f2..f66e79eba 100644 --- a/fastvideo/tests/contract/test_env_policy.py +++ b/fastvideo/tests/contract/test_env_policy.py @@ -153,7 +153,6 @@ KNOWN_VIOLATIONS: dict[str, int] = { 'fastvideo/distributed/device_communicators/cpu_communicator.py: read VLLM_DIST_IDENT': 1, 'fastvideo/entrypoints/cli/utils.py: whole-environ': 1, 'fastvideo/entrypoints/openai/api_server.py: write FASTVIDEO_STAGE_LOGGING': 1, - 'fastvideo/entrypoints/video_generator.py: write FASTVIDEO_NVFP4_FA4': 1, 'fastvideo/mlx_runtime/memory.py: write ': 1, 'fastvideo/performance/hf_store.py: read HF_REPO_ID': 1, 'fastvideo/performance/hf_store.py: read PERFORMANCE_TRACKING_SYNC_REUSE_TTL_SECONDS': 1, diff --git a/fastvideo/tests/entrypoints/test_video_generator.py b/fastvideo/tests/entrypoints/test_video_generator.py index 192a4bc50..a4570451d 100644 --- a/fastvideo/tests/entrypoints/test_video_generator.py +++ b/fastvideo/tests/entrypoints/test_video_generator.py @@ -631,7 +631,7 @@ def test_from_file_loads_generator_from_run_config(tmp_path, monkeypatch): assert captured["resolved_config"].engine.num_gpus == 3 -def test_from_pretrained_convenience_kwargs_do_not_warn(monkeypatch): +def test_from_pretrained_nested_config_does_not_warn(monkeypatch): captured = _patch_from_resolved_config(monkeypatch) _skip_model_definition(monkeypatch) @@ -639,12 +639,18 @@ def test_from_pretrained_convenience_kwargs_do_not_warn(monkeypatch): warnings.simplefilter("always") generator = VideoGenerator.from_pretrained( "test-model", - num_gpus=4, - use_fsdp_inference=False, - text_encoder_cpu_offload=True, - pin_cpu_memory=True, - dit_cpu_offload=False, - vae_cpu_offload=False, + { + "engine": { + "num_gpus": 4, + "use_fsdp_inference": False, + "offload": { + "text_encoder": True, + "pin_cpu_memory": True, + "dit": False, + "vae": False, + }, + }, + }, ) assert not caught @@ -661,11 +667,11 @@ def test_from_pretrained_convenience_kwargs_do_not_warn(monkeypatch): assert generator.config.engine.num_gpus == 4 -def test_from_pretrained_rejects_other_kwargs_and_points_to_from_config(monkeypatch): +def test_from_pretrained_rejects_flat_keywords_and_points_to_a_nested_config(monkeypatch): captured = _patch_from_resolved_config(monkeypatch) _skip_model_definition(monkeypatch) - with pytest.raises(TypeError, match=r"does not accept workload_type\.(.|\n)*VideoGenerator\.from_config"): + with pytest.raises(TypeError, match=r"does not accept num_gpus, workload_type\.(.|\n)*nested config"): VideoGenerator.from_pretrained( "test-model", num_gpus=4, @@ -675,6 +681,16 @@ def test_from_pretrained_rejects_other_kwargs_and_points_to_from_config(monkeypa assert "resolved_config" not in captured +def test_from_pretrained_rejects_a_config_with_another_model_path(monkeypatch): + captured = _patch_from_resolved_config(monkeypatch) + _skip_model_definition(monkeypatch) + + with pytest.raises(ValueError, match="differs from model_path"): + VideoGenerator.from_pretrained("test-model", {"model_path": "other-model"}) + + assert "resolved_config" not in captured + + def test_generate_uses_typed_request_path(monkeypatch): generator = _new_runtime_video_generator() _patch_sampling_param_from_pretrained(monkeypatch) diff --git a/fastvideo/tests/loader/test_text_encoder_unified_memory_offload.py b/fastvideo/tests/loader/test_text_encoder_unified_memory_offload.py index a96a2c275..47966a7b6 100644 --- a/fastvideo/tests/loader/test_text_encoder_unified_memory_offload.py +++ b/fastvideo/tests/loader/test_text_encoder_unified_memory_offload.py @@ -1,14 +1,15 @@ # SPDX-License-Identifier: Apache-2.0 -"""Regression tests for encoder placement on unified memory.""" +"""Regression tests for encoder placement: the loader reads the offload decision that resolution made.""" from __future__ import annotations from types import SimpleNamespace -from unittest.mock import Mock +from unittest.mock import Mock, patch import pytest import torch import torch.nn as nn +from fastvideo.api import device_policy from fastvideo.api.inference_resolution import resolve_inference_config from fastvideo.models.loader.component_loader import ImageEncoderLoader, TextEncoderLoader from fastvideo.tests.api.config_snapshot import isolated_environment @@ -31,58 +32,57 @@ def _model_config(): return SimpleNamespace(architectures=["PassthroughEncoder"], _fsdp_shard_conditions=[], quant_config=None) -def _resolved_config(**offload): - """A resolved inference config for a registered model with the given ``engine.offload`` values.""" - with isolated_environment(): +def _resolved_config(monkeypatch, *, unified: bool, **offload): + """A resolved inference config for a registered model, resolved on a unified or discrete device.""" + monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", lambda device_id: unified) + monkeypatch.setattr("fastvideo.platforms.current_platform.get_device_name", lambda device_id: "NVIDIA GB10") + monkeypatch.setattr("fastvideo.platforms.current_platform.is_mps", lambda: False) + with isolated_environment(), patch.object(device_policy, "APPLY_DEVICE_POLICY", True): return resolve_inference_config({"model_path": WAN_T2V, "engine": {"offload": offload}}) -@pytest.mark.parametrize( - ("cpu_offload", "requested_target"), - [ - (None, torch.device("cuda:5")), - (None, torch.device("cpu")), - (True, torch.device("cpu")), - ], -) -def test_unified_memory_uses_worker_device_before_model_construction(monkeypatch, tmp_path, cpu_offload, - requested_target) -> None: - probe = Mock(return_value=True) +@pytest.fixture +def passthrough_registry(monkeypatch): + monkeypatch.setattr( + "fastvideo.models.loader.component_loader.ModelRegistry.resolve_model_cls", + lambda architectures: (_PassthroughEncoder, None), + ) + + +@pytest.mark.parametrize("requested_target", [torch.device("cuda:5"), torch.device("cpu")]) +def test_unified_memory_decision_keeps_the_encoder_on_the_worker_device(monkeypatch, tmp_path, passthrough_registry, + requested_target) -> None: + resolved_config = _resolved_config(monkeypatch, unified=True, text_encoder=True) + assert resolved_config.engine.offload.text_encoder is False + probe = Mock(side_effect=AssertionError("device probe ran in the loader")) + monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", probe) monkeypatch.setattr("fastvideo.models.loader.component_loader.get_local_torch_device", lambda: torch.device("cuda:5")) - monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", probe) - monkeypatch.setattr("fastvideo.platforms.current_platform.get_device_name", lambda device_id: "NVIDIA GB10") - monkeypatch.setattr( - "fastvideo.models.loader.component_loader.ModelRegistry.resolve_model_cls", - lambda architectures: (_PassthroughEncoder, None), - ) - resolved_config = _resolved_config(text_encoder=True) - model = TextEncoderLoader().load_model( - str(tmp_path), - _model_config(), - requested_target, - resolved_config, - cpu_offload=cpu_offload, - ) + model = TextEncoderLoader().load_model(str(tmp_path), _model_config(), requested_target, resolved_config) assert isinstance(model, _PassthroughEncoder) - assert _PassthroughEncoder.loaded_device == torch.device("cuda:5") - # The loader only queries the policy; the worker records the offload decisions on the config. - assert resolved_config.engine.offload.text_encoder is True + # cpu_offload is off, so the loader keeps the target that the caller chose. + assert _PassthroughEncoder.loaded_device == requested_target assert resolved_config.override_log == () - probe.assert_called_once_with(5) + probe.assert_not_called() -def test_discrete_memory_preserves_explicit_cpu_target(monkeypatch, tmp_path) -> None: +def test_discrete_memory_offloads_the_encoder_to_the_host(monkeypatch, tmp_path, passthrough_registry) -> None: + resolved_config = _resolved_config(monkeypatch, unified=False, text_encoder=True) + assert resolved_config.engine.offload.text_encoder is True + monkeypatch.setattr("fastvideo.models.loader.component_loader.get_local_torch_device", + lambda: torch.device("cuda:2")) + + TextEncoderLoader().load_model(str(tmp_path), _model_config(), torch.device("cuda:2"), resolved_config) + + assert _PassthroughEncoder.loaded_device == torch.device("cpu") + + +def test_discrete_memory_preserves_explicit_cpu_target(monkeypatch, tmp_path, passthrough_registry) -> None: + resolved_config = _resolved_config(monkeypatch, unified=False, text_encoder=False) monkeypatch.setattr("fastvideo.models.loader.component_loader.get_local_torch_device", lambda: torch.device("cuda:2")) - monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", lambda device_id: False) - monkeypatch.setattr( - "fastvideo.models.loader.component_loader.ModelRegistry.resolve_model_cls", - lambda architectures: (_PassthroughEncoder, None), - ) - resolved_config = _resolved_config(text_encoder=False) TextEncoderLoader().load_model( str(tmp_path), @@ -95,27 +95,20 @@ def test_discrete_memory_preserves_explicit_cpu_target(monkeypatch, tmp_path) -> assert _PassthroughEncoder.loaded_device == torch.device("cpu") -def test_image_encoder_explicit_offload_resets_cpu_target(monkeypatch, tmp_path) -> None: - probe = Mock(return_value=True) +def test_image_encoder_reads_its_own_offload_decision(monkeypatch, tmp_path, passthrough_registry) -> None: + resolved_config = _resolved_config(monkeypatch, unified=True, text_encoder=True, image_encoder=True) + assert resolved_config.engine.offload.image_encoder is False monkeypatch.setattr("fastvideo.models.loader.component_loader.get_local_torch_device", lambda: torch.device("cuda:4")) - monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", probe) - monkeypatch.setattr("fastvideo.platforms.current_platform.get_device_name", lambda device_id: "NVIDIA GB10") - monkeypatch.setattr( - "fastvideo.models.loader.component_loader.ModelRegistry.resolve_model_cls", - lambda architectures: (_PassthroughEncoder, None), - ) - resolved_config = _resolved_config(text_encoder=True, image_encoder=True) ImageEncoderLoader().load_model( str(tmp_path), _model_config(), - torch.device("cpu"), + torch.device("cuda:4"), resolved_config, - cpu_offload=True, + cpu_offload=resolved_config.engine.offload.image_encoder, offload_flag="image_encoder_cpu_offload", ) assert _PassthroughEncoder.loaded_device == torch.device("cuda:4") assert resolved_config.override_log == () - probe.assert_called_once_with(4) diff --git a/fastvideo/tests/loader/test_transformer_loader_config.py b/fastvideo/tests/loader/test_transformer_loader_config.py index d501e4fd7..83fa9f8fc 100644 --- a/fastvideo/tests/loader/test_transformer_loader_config.py +++ b/fastvideo/tests/loader/test_transformer_loader_config.py @@ -57,6 +57,7 @@ def test_transformer_loader_passes_the_typed_engine_settings(transformer_dir, fs "use_fsdp_inference": True, "offload": { "dit": False, + "dit_layerwise": False, "pin_cpu_memory": False }, "compile": { @@ -120,3 +121,22 @@ def test_teacher_critic_flag_loads_checkpoint_weights_without_quantization(monke assert teacher["weight_dir_list"] == [str(transformer_dir / "model.safetensors")] assert teacher["init_params"]["config"].quant_config is None assert resolved_config.pipeline_config.dit_config.quant_config is quant_config + + +def test_load_argument_overrides_the_transformer_class_name(monkeypatch, transformer_dir, fsdp_calls) -> None: + """``override_transformer_cls_name`` of the load call wins over the typed component setting.""" + resolved_config = _resolve({"pipeline": {"components": {"override_transformer_cls_name": "StudentModel"}}}) + requested = [] + monkeypatch.setattr(component_loader.ModelRegistry, "resolve_model_cls", + lambda name: requested.append(name) or (_FakeTransformer, None)) + + PipelineComponentLoader.load_module("transformer", str(transformer_dir), "diffusers", resolved_config) + PipelineComponentLoader.load_module("transformer", + str(transformer_dir), + "diffusers", + resolved_config, + loading_teacher_critic_model=True, + override_transformer_cls_name="TeacherModel") + + assert requested == ["StudentModel", "TeacherModel"] + assert resolved_config.pipeline.components.override_transformer_cls_name == "StudentModel" diff --git a/fastvideo/tests/pipelines/test_direct_pipeline_unified_memory.py b/fastvideo/tests/pipelines/test_direct_pipeline_unified_memory.py index 7cd7f0436..a0c93cae3 100644 --- a/fastvideo/tests/pipelines/test_direct_pipeline_unified_memory.py +++ b/fastvideo/tests/pipelines/test_direct_pipeline_unified_memory.py @@ -1,13 +1,15 @@ # SPDX-License-Identifier: Apache-2.0 -"""Direct pipeline construction applies offload policy after device setup.""" +"""Direct pipeline construction runs on the resolved config, whose offload settings resolution already settled.""" from __future__ import annotations from contextlib import nullcontext from types import SimpleNamespace +from unittest.mock import Mock, patch import torch import fastvideo.pipelines.composed_pipeline_base as composed_pipeline_base +from fastvideo.api import device_policy from fastvideo.api.device_policy import UNIFIED_MEMORY_OFFLOAD_PATHS from fastvideo.api.inference_resolution import resolve_inference_config from fastvideo.pipelines.composed_pipeline_base import ComposedPipelineBase @@ -35,10 +37,13 @@ class _Pipeline(ComposedPipelineBase): del resolved_config -def test_direct_pipeline_applies_policy_after_device_initialization(monkeypatch) -> None: +def test_direct_pipeline_loads_with_the_resolved_offload_policy(monkeypatch) -> None: events = [] monkeypatch.setattr(_Pipeline, "events", events) - with isolated_environment(): + monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", lambda device_id: True) + monkeypatch.setattr("fastvideo.platforms.current_platform.get_device_name", lambda device_id: "NVIDIA GB10") + monkeypatch.setattr("fastvideo.platforms.current_platform.is_mps", lambda: False) + with isolated_environment(), patch.object(device_policy, "APPLY_DEVICE_POLICY", True): resolved_config = resolve_inference_config({ "model_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", "engine": { @@ -46,28 +51,22 @@ def test_direct_pipeline_applies_policy_after_device_initialization(monkeypatch) }, }) - def classify_device(device_id): - events.append(("offload_policy", device_id)) - return True - + probe = Mock(side_effect=AssertionError("device probe ran while building the pipeline")) + monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", probe) monkeypatch.setattr( composed_pipeline_base, "maybe_init_distributed_environment_and_model_parallel", lambda *args: events.append(("distributed", None)), ) - monkeypatch.setattr(composed_pipeline_base, "get_local_torch_device", lambda: torch.device("cuda:4")) monkeypatch.setattr(composed_pipeline_base, "get_world_group", lambda: SimpleNamespace(local_rank=4)) monkeypatch.setattr(composed_pipeline_base, "get_or_create_profiler", lambda trace_dir: _Profiler()) - monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", classify_device) - monkeypatch.setattr("fastvideo.platforms.current_platform.get_device_name", lambda device_id: "NVIDIA GB10") - monkeypatch.setattr("fastvideo.platforms.current_platform.is_mps", lambda: False) pipeline = _Pipeline("unused", resolved_config, required_config_modules=[]) assert pipeline.modules == {} + assert pipeline.resolved_config is resolved_config assert events == [ ("distributed", None), - ("offload_policy", 4), ( "load_modules", { @@ -76,3 +75,4 @@ def test_direct_pipeline_applies_policy_after_device_initialization(monkeypatch) }, ), ] + probe.assert_not_called() diff --git a/fastvideo/tests/pipelines/test_ltx2_refine_checkpoint_defaults.py b/fastvideo/tests/pipelines/test_ltx2_refine_checkpoint_defaults.py index 3d6d668b7..a69871800 100644 --- a/fastvideo/tests/pipelines/test_ltx2_refine_checkpoint_defaults.py +++ b/fastvideo/tests/pipelines/test_ltx2_refine_checkpoint_defaults.py @@ -1,9 +1,13 @@ # SPDX-License-Identifier: Apache-2.0 -"""LTX-2 refine settings after the checkpoint's model_index.json defaults are applied at load time.""" +"""LTX-2 refine settings after resolution fills them from the checkpoint's model_index.json.""" from __future__ import annotations -from fastvideo.api.inference_resolution import resolve_inference_config -from fastvideo.pipelines.basic.ltx2.ltx2_pipeline import LTX2Pipeline, _checkpoint_refine_values +import json + +from fastvideo.api.checkpoint_defaults import ltx2_refine_checkpoint_step +from fastvideo.api.inference_resolution import fill_runtime_defaults, resolve_inference_config +from fastvideo.api.resolution import resolve_generator_config +from fastvideo.pipelines.basic.ltx2.pipeline_configs import LTX2T2VConfig from fastvideo.tests.api.config_snapshot import isolated_environment LTX2 = "FastVideo/LTX2-Distilled-Diffusers" @@ -14,46 +18,90 @@ CHECKPOINT_REFINE = { "fastvideo_refine_add_noise": False, "fastvideo_refine_lora_path": "FastVideo/LTX2-Distilled-LoRA", } +REFINE_SWITCHES = ( + "pipeline.ltx2.refine.enabled", + "pipeline.ltx2.refine.num_inference_steps", + "pipeline.ltx2.refine.guidance_scale", + "pipeline.ltx2.refine.add_noise", +) -def _resolve(refine: dict | None = None): - raw = {"model_path": LTX2} +def _checkpoint(tmp_path, **manifest_entries) -> str: + """A local LTX-2 checkpoint directory whose model_index.json holds ``manifest_entries``.""" + (tmp_path / "transformer").mkdir() + (tmp_path / "model_index.json").write_text( + json.dumps({ + "_class_name": "LTX2Pipeline", + "_diffusers_version": "0", + "transformer": ["diffusers", "Model"], + **manifest_entries + })) + return str(tmp_path) + + +def _resolve_local(model_path: str, refine: dict | None = None, upsampler_weights: str | None = None): + """Resolve a local LTX-2 checkpoint through the checkpoint step and the runtime defaults only.""" + raw: dict = {"model_path": model_path, "pipeline": {"components": {"upsampler_weights": upsampler_weights}}} if refine is not None: - raw["pipeline"] = {"ltx2": {"refine": refine}} - with isolated_environment(): - return resolve_inference_config(raw) + raw["pipeline"]["ltx2"] = {"refine": refine} + return resolve_generator_config(raw, (ltx2_refine_checkpoint_step(LTX2T2VConfig()), fill_runtime_defaults)) -def test_load_fills_unset_refine_settings_from_the_checkpoint_then_the_defaults(tmp_path): - resolved = _resolve() - pipeline = object.__new__(LTX2Pipeline) - pipeline.model_path = str(tmp_path) - pipeline.resolved_config = resolved - pipeline._required_config_modules = list(LTX2Pipeline._required_config_modules) - model_index = {name: ["diffusers", "Model"] for name in pipeline._required_config_modules} - pipeline._load_config = lambda model_path: { - "_class_name": "LTX2Pipeline", - "_diffusers_version": "0", - "fastvideo_refine_num_inference_steps": 2, - **model_index, - } +def test_checkpoint_fills_unset_refine_settings_then_the_defaults_fill_the_rest(tmp_path): + model_path = _checkpoint(tmp_path, fastvideo_refine_num_inference_steps=2) + (tmp_path / "spatial_upscaler").mkdir() - pipeline.load_modules(resolved, {name: object() for name in model_index}) + resolved = _resolve_local(model_path) - refine = pipeline.resolved_config.pipeline.ltx2.refine + refine = resolved.pipeline.ltx2.refine switches = (refine.enabled, refine.num_inference_steps, refine.guidance_scale, refine.add_noise) assert switches == (False, 2, 1.0, True) - assert pipeline.resolved_config.provenance("pipeline.ltx2.refine.add_noise").source == "checkpoint:model_index.json" - assert resolved.pipeline.ltx2.refine.num_inference_steps is None + assert resolved.provenance("pipeline.ltx2.refine.num_inference_steps").source == "fill_ltx2_refine_from_checkpoint" + assert resolved.provenance("pipeline.ltx2.refine.add_noise").source == "fill_runtime_defaults" + assert resolved.pipeline.components.upsampler_weights == str(tmp_path / "spatial_upscaler") -def test_input_refine_settings_win_over_the_checkpoint(): - resolved = _resolve({"enabled": False, "num_inference_steps": 3, "guidance_scale": 2.0}) +def test_input_refine_settings_win_over_the_checkpoint(tmp_path): + model_path = _checkpoint(tmp_path, **CHECKPOINT_REFINE) - values = _checkpoint_refine_values(resolved, "/unused", dict(CHECKPOINT_REFINE)) + resolved = _resolve_local(model_path, {"enabled": False, "num_inference_steps": 3, "guidance_scale": 2.0}) - assert values == { + checkpoint_decisions = [ + values for source, values in resolved.decisions if source == "fill_ltx2_refine_from_checkpoint" + ] + assert checkpoint_decisions == [{ "pipeline.ltx2.refine.add_noise": False, "pipeline.ltx2.refine.lora_path": "FastVideo/LTX2-Distilled-LoRA", - } - assert _checkpoint_refine_values(_resolve(), "/unused", dict(CHECKPOINT_REFINE))["pipeline.ltx2.refine.enabled"] + }] + assert resolved.pipeline.ltx2.refine.enabled is False + assert _resolve_local(model_path).pipeline.ltx2.refine.enabled is True + + +def test_relative_refine_paths_resolve_inside_the_checkpoint(tmp_path): + (tmp_path / "transformer_refine").mkdir() + model_path = _checkpoint(tmp_path, + fastvideo_refine_transformer_path="transformer_refine", + fastvideo_refine_noise_path="/weights/noise.safetensors", + spatial_upsampler=["diffusers", "LTX2LatentUpsampler"]) + (tmp_path / "spatial_upsampler").mkdir() + + resolved = _resolve_local(model_path) + + assert resolved.pipeline.ltx2.refine.transformer_path == str(tmp_path / "transformer_refine") + assert resolved.pipeline.ltx2.refine.noise_path == "/weights/noise.safetensors" + assert resolved.pipeline.components.upsampler_weights == str(tmp_path / "spatial_upsampler") + + +def test_step_skips_other_models_and_blocked_downloads(tmp_path): + model_path = _checkpoint(tmp_path, **CHECKPOINT_REFINE) + other_model = resolve_generator_config({"model_path": model_path}, + (ltx2_refine_checkpoint_step(object()), fill_runtime_defaults)) + assert other_model.pipeline.ltx2.refine.enabled is False + assert other_model.pipeline.ltx2.refine.lora_path is None + + with isolated_environment(): + hub_model = resolve_inference_config({"model_path": LTX2}) + + assert [source for source, _ in hub_model.decisions if source == "fill_ltx2_refine_from_checkpoint"] == [] + assert tuple(hub_model.provenance(path).value for path in REFINE_SWITCHES) == (False, 3, 1.0, True) + assert hub_model.provenance("pipeline.ltx2.refine.enabled").source == "fill_runtime_defaults" diff --git a/fastvideo/tests/platforms/test_unified_memory_offload.py b/fastvideo/tests/platforms/test_unified_memory_offload.py index 959624e07..ab0223fd7 100644 --- a/fastvideo/tests/platforms/test_unified_memory_offload.py +++ b/fastvideo/tests/platforms/test_unified_memory_offload.py @@ -1,14 +1,14 @@ # SPDX-License-Identifier: Apache-2.0 -"""CPU tests for worker-local offload policy on unified-memory devices.""" +"""CPU tests for the device-policy resolution steps, which settle host offload for the local device's memory class.""" from __future__ import annotations import dataclasses -from unittest.mock import Mock +from unittest.mock import Mock, patch import pytest -from fastvideo.api.device_policy import (UNIFIED_MEMORY_OFFLOAD_PATHS, disable_offload_on_unified_memory, - finalize_device_offload_policy) +from fastvideo.api import device_policy +from fastvideo.api.device_policy import LOCAL_DEVICE_ID, UNIFIED_MEMORY_OFFLOAD_PATHS from fastvideo.api.inference_resolution import resolve_inference_config from fastvideo.api.schema import OffloadConfig from fastvideo.api.training_schema import resolve_training_config @@ -23,9 +23,14 @@ def _offload_field(flag: str) -> str: return UNIFIED_MEMORY_OFFLOAD_PATHS[flag].rsplit(".", 1)[1] +def _device_policy_enabled(): + """Turn the device-policy steps on inside the test isolation, which turns them off like downloads.""" + return patch.object(device_policy, "APPLY_DEVICE_POLICY", True) + + def _resolve(engine: dict | None = None): """A resolved inference config for a registered model with the given ``engine`` section.""" - with isolated_environment(): + with isolated_environment(), _device_policy_enabled(): return resolve_inference_config({"model_path": WAN_T2V, "engine": engine or {}}) @@ -42,6 +47,11 @@ def _enabled_flags(resolved_config) -> list[str]: ] +def _decisions_of(resolved_config, source: str) -> list[dict]: + """The values that the step ``source`` decided, in order.""" + return [values for decided_by, values in resolved_config.decisions if decided_by == source] + + @pytest.fixture def as_unified_cuda(monkeypatch): probe = Mock(return_value=True) @@ -51,22 +61,23 @@ def as_unified_cuda(monkeypatch): return probe -def test_resolving_an_inference_config_defers_device_policy(monkeypatch) -> None: - probe = Mock(side_effect=AssertionError("device probe ran in the driver")) +@pytest.fixture +def as_discrete_cuda(monkeypatch): + probe = Mock(return_value=False) monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", probe) - - resolved_config = _resolve({"use_fsdp_inference": True}) - - assert resolved_config.engine.use_fsdp_inference is True - assert resolved_config.engine.offload.dit_layerwise is True - assert resolved_config.engine.offload.dit is True - probe.assert_not_called() - - -def test_training_resolution_retains_offload_conflict_normalization(monkeypatch) -> None: monkeypatch.setattr("fastvideo.platforms.current_platform.is_mps", lambda: False) + return probe - with isolated_environment(): + +def test_resolution_queries_the_local_device(as_unified_cuda) -> None: + _resolve() + + assert as_unified_cuda.call_args_list + assert {call.args for call in as_unified_cuda.call_args_list} == {(LOCAL_DEVICE_ID, )} + + +def test_training_resolution_settles_offload_conflicts(as_discrete_cuda) -> None: + with isolated_environment(), _device_policy_enabled(): resolved_config = resolve_training_config({ "model_path": WAN_T2V, "engine": { @@ -85,6 +96,8 @@ def test_training_resolution_retains_offload_conflict_normalization(monkeypatch) assert resolved_config.engine.offload.dit_layerwise is True assert resolved_config.engine.offload.dit is False assert resolved_config.engine.use_fsdp_inference is False + assert resolved_config.provenance("engine.use_fsdp_inference").source == "apply_layerwise_offload_conflicts" + assert resolved_config.engine.offload.lazy_module_load is False def test_policy_list_covers_every_host_offload_field() -> None: @@ -96,145 +109,139 @@ def test_policy_list_covers_every_host_offload_field() -> None: def test_unified_device_disables_every_offload_flag(as_unified_cuda) -> None: resolved_config = _resolve() - decided = finalize_device_offload_policy(resolved_config, device_id=6) - - as_unified_cuda.assert_called_once_with(6) - assert not _enabled_flags(decided) - assert _enabled_flags(resolved_config) == list(UNIFIED_MEMORY_OFFLOAD_FLAGS) + assert not _enabled_flags(resolved_config) + assert _decisions_of(resolved_config, "apply_unified_memory_offload_policy") == [{ + path: False + for path in UNIFIED_MEMORY_OFFLOAD_PATHS.values() + }] + for flag in UNIFIED_MEMORY_OFFLOAD_FLAGS: + provenance = resolved_config.provenance(UNIFIED_MEMORY_OFFLOAD_PATHS[flag]) + assert (provenance.raw_value, provenance.source) == (True, "apply_unified_memory_offload_policy") @pytest.mark.parametrize("flag", UNIFIED_MEMORY_OFFLOAD_FLAGS) def test_each_offload_flag_is_independently_disabled(as_unified_cuda, flag: str) -> None: resolved_config = _with_offloads(flag) - assert _enabled_flags(resolved_config) == [flag] - decided = disable_offload_on_unified_memory(resolved_config, device_id=2) - - assert not _enabled_flags(decided) - assert decided.override_log == (("device_policy:unified_memory", {UNIFIED_MEMORY_OFFLOAD_PATHS[flag]: False}), ) + assert not _enabled_flags(resolved_config) + assert _decisions_of(resolved_config, "apply_unified_memory_offload_policy") == [{ + UNIFIED_MEMORY_OFFLOAD_PATHS[flag]: False + }] -def test_discrete_device_classification_preserves_offload_requests(monkeypatch) -> None: - probe = Mock(return_value=False) - monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", probe) +def test_discrete_device_classification_preserves_offload_requests(as_discrete_cuda) -> None: resolved_config = _resolve() - assert disable_offload_on_unified_memory(resolved_config, device_id=3) is resolved_config - - probe.assert_called_once_with(3) - assert _enabled_flags(resolved_config) == list(UNIFIED_MEMORY_OFFLOAD_FLAGS) + assert _enabled_flags(resolved_config) == [ + flag for flag in UNIFIED_MEMORY_OFFLOAD_FLAGS if flag != "dit_cpu_offload" + ] + assert _decisions_of(resolved_config, "apply_unified_memory_offload_policy") == [] + assert resolved_config.provenance("engine.offload.dit").source == "apply_layerwise_offload_conflicts" -def test_discrete_device_finalization_retains_layerwise_precedence(monkeypatch) -> None: - monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", lambda device_id: False) - monkeypatch.setattr("fastvideo.platforms.current_platform.is_mps", lambda: False) +def test_discrete_device_retains_layerwise_precedence(as_discrete_cuda) -> None: resolved_config = _resolve({"use_fsdp_inference": True}) - decided = finalize_device_offload_policy(resolved_config, device_id=3) - - offload = decided.engine.offload + offload = resolved_config.engine.offload assert offload.dit_layerwise is True assert offload.dit is False - assert decided.engine.use_fsdp_inference is False + assert resolved_config.engine.use_fsdp_inference is False assert offload.text_encoder is True assert offload.image_encoder is True assert offload.vae is True assert offload.lazy_module_load is False - - -def test_workers_classify_their_own_device(monkeypatch) -> None: - seen_device_ids = [] - - def has_unified_memory(device_id): - seen_device_ids.append(device_id) - return device_id == 1 - - monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", has_unified_memory) - monkeypatch.setattr("fastvideo.platforms.current_platform.get_device_name", lambda device_id: "NVIDIA GB10") - monkeypatch.setattr("fastvideo.platforms.current_platform.is_mps", lambda: False) - resolved_config = _resolve() - - device_zero = finalize_device_offload_policy(resolved_config, device_id=0) - device_one = finalize_device_offload_policy(resolved_config, device_id=1) - - assert seen_device_ids == [0, 1] - assert device_zero.engine.offload.dit_layerwise is True - assert device_zero.engine.offload.text_encoder is True - assert not _enabled_flags(device_one) - - -def test_policy_applied_to_its_own_result_changes_nothing(as_unified_cuda) -> None: - decided = finalize_device_offload_policy(_resolve(), device_id=1) - - again = finalize_device_offload_policy(decided, device_id=1) - - assert again.override_log == decided.override_log - assert again.to_dict() == decided.to_dict() + assert _decisions_of(resolved_config, "apply_layerwise_offload_conflicts") == [{ + "engine.use_fsdp_inference": False, + "engine.offload.dit": False, + }] def test_mps_clears_offload_and_keeps_its_fsdp_rule(monkeypatch) -> None: monkeypatch.setattr("fastvideo.platforms.current_platform.is_mps", lambda: True) monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", lambda device_id: True) monkeypatch.setattr("fastvideo.platforms.current_platform.get_device_name", lambda device_id: "mps") + resolved_config = _resolve({"use_fsdp_inference": True}) - decided = finalize_device_offload_policy(resolved_config) - - assert decided.engine.use_fsdp_inference is False - assert not _enabled_flags(decided) + assert resolved_config.engine.use_fsdp_inference is False + assert not _enabled_flags(resolved_config) + assert _decisions_of(resolved_config, "apply_mps_offload_policy") == [{ + "engine.use_fsdp_inference": False, + "engine.offload.dit_layerwise": False, + }] def test_cuda_unified_memory_preserves_realistic_fsdp_request(as_unified_cuda) -> None: resolved_config = _resolve({"use_fsdp_inference": True}) - assert resolved_config.engine.offload.dit_layerwise is True + assert resolved_config.engine.use_fsdp_inference is True - - decided = finalize_device_offload_policy(resolved_config) - - assert decided.engine.use_fsdp_inference is True - assert not _enabled_flags(decided) + assert not _enabled_flags(resolved_config) + assert _decisions_of(resolved_config, "apply_layerwise_offload_conflicts") == [] def test_pin_cpu_memory_is_not_a_host_offload_mode(as_unified_cuda) -> None: resolved_config = _resolve({"offload": {"pin_cpu_memory": True}}) - decided = finalize_device_offload_policy(resolved_config) - assert "pin_cpu_memory" not in UNIFIED_MEMORY_OFFLOAD_FLAGS - assert decided.engine.offload.pin_cpu_memory is True + assert resolved_config.engine.offload.pin_cpu_memory is True def test_already_disabled_flags_stay_disabled(as_unified_cuda) -> None: - decided = finalize_device_offload_policy(_with_offloads()) + resolved_config = _with_offloads() - assert not _enabled_flags(decided) + assert not _enabled_flags(resolved_config) + assert _decisions_of(resolved_config, "apply_unified_memory_offload_policy") == [] @pytest.mark.parametrize("name_error", [NotImplementedError, ValueError, RuntimeError]) def test_platform_without_device_name_uses_generic_name(monkeypatch, name_error: type[Exception]) -> None: monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", lambda device_id: True) + monkeypatch.setattr("fastvideo.platforms.current_platform.is_mps", lambda: False) def unsupported_name(device_id): raise name_error("device name unavailable") monkeypatch.setattr("fastvideo.platforms.current_platform.get_device_name", unsupported_name) + resolved_config = _with_offloads("text_encoder_cpu_offload") - decided = disable_offload_on_unified_memory(resolved_config, device_id=0) + assert resolved_config.engine.offload.text_encoder is False - assert decided.engine.offload.text_encoder is False + +def test_failed_device_query_counts_as_discrete(monkeypatch) -> None: + + def unavailable(device_id): + raise RuntimeError("no CUDA driver") + + monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", unavailable) + monkeypatch.setattr("fastvideo.platforms.current_platform.is_mps", lambda: False) + + resolved_config = _with_offloads("text_encoder_cpu_offload") + + assert resolved_config.engine.offload.text_encoder is True + assert resolved_config.engine.offload.lazy_module_load is False def test_unified_device_auto_enables_lazy_module_load(as_unified_cuda) -> None: resolved_config = _resolve() - assert resolved_config.engine.offload.lazy_module_load is None - assert finalize_device_offload_policy(resolved_config, device_id=6).engine.offload.lazy_module_load is True + provenance = resolved_config.provenance("engine.offload.lazy_module_load") + assert (provenance.raw_value, provenance.value, provenance.source) == (None, True, "fill_lazy_module_load") def test_explicit_false_lazy_module_load_stays_off_on_unified(as_unified_cuda) -> None: resolved_config = _resolve({"offload": {"lazy_module_load": False}}) - decided = finalize_device_offload_policy(resolved_config, device_id=6) + assert resolved_config.engine.offload.lazy_module_load is False + assert resolved_config.provenance("engine.offload.lazy_module_load").source == "input" - assert decided.engine.offload.lazy_module_load is False + +def test_disabled_device_policy_decides_nothing(as_unified_cuda) -> None: + with isolated_environment(), patch.object(device_policy, "APPLY_DEVICE_POLICY", False): + resolved_config = resolve_inference_config({"model_path": WAN_T2V, "engine": {"use_fsdp_inference": True}}) + + assert _enabled_flags(resolved_config) == list(UNIFIED_MEMORY_OFFLOAD_FLAGS) + assert resolved_config.engine.use_fsdp_inference is True + assert resolved_config.engine.offload.lazy_module_load is None + assert [source for source, _ in resolved_config.decisions if source.startswith(("apply_", "fill_lazy"))] == [] + as_unified_cuda.assert_not_called() diff --git a/fastvideo/tests/ssim/test_gamecraft_similarity.py b/fastvideo/tests/ssim/test_gamecraft_similarity.py index 809a53375..5fc5c4fb3 100644 --- a/fastvideo/tests/ssim/test_gamecraft_similarity.py +++ b/fastvideo/tests/ssim/test_gamecraft_similarity.py @@ -178,13 +178,17 @@ def test_gamecraft_t2v_similarity(prompt, ATTENTION_BACKEND, model_id, env_overr dtype=torch.bfloat16, ) - init_kwargs = { - "num_gpus": BASE_PARAMS["num_gpus"], - "use_fsdp_inference": True, - "dit_cpu_offload": True, - "vae_cpu_offload": True, - "text_encoder_cpu_offload": True, - "pin_cpu_memory": True, + generator_config = { + "engine": { + "num_gpus": BASE_PARAMS["num_gpus"], + "use_fsdp_inference": True, + "offload": { + "dit": True, + "vae": True, + "text_encoder": True, + "pin_cpu_memory": True, + }, + }, } request = { @@ -210,7 +214,7 @@ def test_gamecraft_t2v_similarity(prompt, ATTENTION_BACKEND, model_id, env_overr generator: VideoGenerator | None = None try: - generator = VideoGenerator.from_pretrained(model_path=BASE_PARAMS["model_path"], **init_kwargs) + generator = VideoGenerator.from_pretrained(BASE_PARAMS["model_path"], generator_config) generator.generate(request) finally: _shutdown_executor(generator) @@ -308,13 +312,17 @@ def test_gamecraft_i2v_similarity(prompt, ATTENTION_BACKEND, model_id, env_overr dtype=torch.bfloat16, ) - init_kwargs = { - "num_gpus": BASE_PARAMS["num_gpus"], - "use_fsdp_inference": True, - "dit_cpu_offload": True, - "vae_cpu_offload": True, - "text_encoder_cpu_offload": True, - "pin_cpu_memory": True, + generator_config = { + "engine": { + "num_gpus": BASE_PARAMS["num_gpus"], + "use_fsdp_inference": True, + "offload": { + "dit": True, + "vae": True, + "text_encoder": True, + "pin_cpu_memory": True, + }, + }, } request = { @@ -343,7 +351,7 @@ def test_gamecraft_i2v_similarity(prompt, ATTENTION_BACKEND, model_id, env_overr generator: VideoGenerator | None = None try: - generator = VideoGenerator.from_pretrained(model_path=BASE_PARAMS["model_path"], **init_kwargs) + generator = VideoGenerator.from_pretrained(BASE_PARAMS["model_path"], generator_config) generator.generate(request) finally: _shutdown_executor(generator) diff --git a/fastvideo/tests/ssim/test_matrixgame2_similarity.py b/fastvideo/tests/ssim/test_matrixgame2_similarity.py index 6ae963152..9f8cfee91 100644 --- a/fastvideo/tests/ssim/test_matrixgame2_similarity.py +++ b/fastvideo/tests/ssim/test_matrixgame2_similarity.py @@ -117,14 +117,18 @@ def test_matrixgame2_similarity(prompt, ATTENTION_BACKEND, model_id, env_overrid latent_frames = (BASE_PARAMS["num_frames"] - 1) // 4 + 1 grid_sizes = torch.tensor([latent_frames, 44, 80]) - init_kwargs = { - "num_gpus": BASE_PARAMS["num_gpus"], - "use_fsdp_inference": True, - "dit_layerwise_offload": False, - "dit_cpu_offload": False, - "vae_cpu_offload": False, - "text_encoder_cpu_offload": True, - "pin_cpu_memory": True, + generator_config = { + "engine": { + "num_gpus": BASE_PARAMS["num_gpus"], + "use_fsdp_inference": True, + "offload": { + "dit_layerwise": False, + "dit": False, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, + }, + }, } request = { @@ -148,7 +152,7 @@ def test_matrixgame2_similarity(prompt, ATTENTION_BACKEND, model_id, env_overrid }, } - generator = VideoGenerator.from_pretrained(model_path=BASE_PARAMS["model_path"], **init_kwargs) + generator = VideoGenerator.from_pretrained(BASE_PARAMS["model_path"], generator_config) generator.generate(request) if isinstance(generator.executor, MultiprocExecutor): diff --git a/fastvideo/tests/ssim/test_matrixgame3_similarity.py b/fastvideo/tests/ssim/test_matrixgame3_similarity.py index a829570d5..6f489d5b8 100644 --- a/fastvideo/tests/ssim/test_matrixgame3_similarity.py +++ b/fastvideo/tests/ssim/test_matrixgame3_similarity.py @@ -104,14 +104,18 @@ def test_matrixgame3_similarity(prompt, ATTENTION_BACKEND, model_id, env_overrid BASE_PARAMS = params_map[model_id] num_inference_steps = BASE_PARAMS["num_inference_steps"] - init_kwargs = { - "num_gpus": BASE_PARAMS["num_gpus"], - "use_fsdp_inference": True, - "dit_layerwise_offload": False, - "dit_cpu_offload": False, - "vae_cpu_offload": False, - "text_encoder_cpu_offload": True, - "pin_cpu_memory": True, + generator_config = { + "engine": { + "num_gpus": BASE_PARAMS["num_gpus"], + "use_fsdp_inference": True, + "offload": { + "dit_layerwise": False, + "dit": False, + "vae": False, + "text_encoder": True, + "pin_cpu_memory": True, + }, + }, } request = { @@ -133,7 +137,7 @@ def test_matrixgame3_similarity(prompt, ATTENTION_BACKEND, model_id, env_overrid }, } - generator = VideoGenerator.from_pretrained(model_path=BASE_PARAMS["model_path"], **init_kwargs) + generator = VideoGenerator.from_pretrained(BASE_PARAMS["model_path"], generator_config) generator.generate(request) if isinstance(generator.executor, MultiprocExecutor): diff --git a/fastvideo/tests/stages/test_ltx2_refine_upsampler_resolve.py b/fastvideo/tests/stages/test_ltx2_refine_upsampler_resolve.py index a9289c3b0..2ab9efd78 100644 --- a/fastvideo/tests/stages/test_ltx2_refine_upsampler_resolve.py +++ b/fastvideo/tests/stages/test_ltx2_refine_upsampler_resolve.py @@ -4,7 +4,7 @@ Pure tmp_path tests: no GPU, no model downloads. """ -from fastvideo.pipelines.basic.ltx2.ltx2_pipeline import _resolve_refine_upsampler_path +from fastvideo.api.checkpoint_defaults import _resolve_refine_upsampler_path def test_model_index_key_wins_over_directory_probe(tmp_path): diff --git a/fastvideo/tests/stages/test_minimax_h3_distilled_schedule.py b/fastvideo/tests/stages/test_minimax_h3_distilled_schedule.py index adb0308b4..510b38e1c 100644 --- a/fastvideo/tests/stages/test_minimax_h3_distilled_schedule.py +++ b/fastvideo/tests/stages/test_minimax_h3_distilled_schedule.py @@ -9,6 +9,9 @@ import pytest import torch from torch.testing import assert_close +from fastvideo.api.checkpoint_defaults import dmd_schedule_checkpoint_step +from fastvideo.api.inference_resolution import fill_runtime_defaults +from fastvideo.api.resolution import resolve_generator_config from fastvideo.configs.pipelines.minimax_h3 import MiniMaxH3PipelineConfig from fastvideo.models.schedulers.scheduling_minimax_h3 import MiniMaxH3Scheduler from fastvideo.pipelines.basic.minimax_h3.minimax_h3_pipeline import MiniMaxH3ModularPipeline @@ -38,6 +41,19 @@ def _pipeline(tmp_path, *, video_shift=10.0): return pipeline +def _resolve_checkpoint(tmp_path, raw=None): + """Resolve the local checkpoint ``tmp_path`` through the checkpoint schedule step and the runtime defaults.""" + return resolve_generator_config({"model_path": str(tmp_path), **(raw or {})}, + (dmd_schedule_checkpoint_step(MiniMaxH3PipelineConfig()), fill_runtime_defaults), + materialize=lambda _: MiniMaxH3PipelineConfig()) + + +def _with_steps(steps): + """A resolved config whose input sets ``pipeline.dmd_denoising_steps``.""" + return make_resolved_config(MiniMaxH3PipelineConfig(dmd_denoising_steps=steps), + raw={"pipeline": {"dmd_denoising_steps": steps}}) + + def test_pipeline_preserves_shift10_from_checkpoint(tmp_path): # Exercise real initialization with weight loading omitted. The schedulers # represent scheduler_config.json from the eight-forward T2AV checkpoint. @@ -85,8 +101,8 @@ def _run_tiny_stage(monkeypatch, steps, grid_points, *, video_shift=10.0, offloa resolved_config = make_resolved_config(MiniMaxH3PipelineConfig(), raw={"engine": {"offload": offload, "use_fsdp_inference": False}}) if steps is not None: - # A schedule decided after resolution, such as the checkpoint's, arrives through with_override, which does - # not type-check values, so the stage's own ladder validation is what rejects a malformed one. + # The checkpoint step copies the schedule as written and an override skips the parser's type check, so the + # stage's own ladder validation is what rejects a malformed one. resolved_config = resolved_config.with_override("test:dmd_schedule", {"pipeline.dmd_denoising_steps": steps}) batch = ForwardBatch( data_type="video", prompt_embeds=[torch.zeros(1, 2, 8)], latents=torch.ones(2, 96), audio_latents=torch.ones(4, 32), @@ -158,17 +174,34 @@ def test_eight_forward_schedule_uses_trained_rungs_for_both_modalities(monkeypat assert_close(result.audio_latents, torch.full_like(result.audio_latents, 1.0 - 3.0 * audio_sigmas[0])) -def test_pipeline_loads_the_exported_ladder_and_rejects_a_conflicting_one(tmp_path): +def test_resolution_fills_the_exported_ladder_and_the_pipeline_accepts_it(tmp_path): (tmp_path / "fastvideo_inference.json").write_text(json.dumps(CONTRACT)) - pipeline = _pipeline(tmp_path) - pipeline.initialize_pipeline(make_resolved_config(MiniMaxH3PipelineConfig())) - assert list(pipeline.resolved_config.pipeline.dmd_denoising_steps) == DMD_STEPS + resolved_config = _resolve_checkpoint(tmp_path) + assert list(resolved_config.pipeline.dmd_denoising_steps) == DMD_STEPS + provenance = resolved_config.provenance("pipeline.dmd_denoising_steps") + assert provenance.source == "fill_dmd_schedule_from_checkpoint" + pipeline = _pipeline(tmp_path) + pipeline.initialize_pipeline(resolved_config) + + # An input schedule wins during resolution; the pipeline then rejects the disagreement with the checkpoint. + explicit = _resolve_checkpoint(tmp_path, {"pipeline": {"dmd_denoising_steps": DMD_STEPS}}) + assert explicit.provenance("pipeline.dmd_denoising_steps").source == "input" wrong_steps = [1000, 750, 500, 250] - wrong = make_resolved_config(MiniMaxH3PipelineConfig(dmd_denoising_steps=wrong_steps), - raw={"pipeline": {"dmd_denoising_steps": wrong_steps}}) with pytest.raises(ValueError, match="checkpoint.*DMD"): - pipeline.initialize_pipeline(wrong) + pipeline.initialize_pipeline(_with_steps(wrong_steps)) + with pytest.raises(ValueError, match="dmd_denoising_steps is unset"): + pipeline.initialize_pipeline(make_resolved_config(MiniMaxH3PipelineConfig())) + + +def test_checkpoint_step_skips_other_models_and_checkpoints_without_a_schedule(tmp_path): + (tmp_path / "fastvideo_inference.json").write_text(json.dumps(CONTRACT)) + other_model = resolve_generator_config({"model_path": str(tmp_path)}, + (dmd_schedule_checkpoint_step(object()), fill_runtime_defaults)) + assert other_model.pipeline.dmd_denoising_steps is None + + (tmp_path / "fastvideo_inference.json").unlink() + assert _resolve_checkpoint(tmp_path).pipeline.dmd_denoising_steps is None @pytest.mark.parametrize("grid_points", [5, 50]) @@ -179,8 +212,8 @@ def test_base_and_four_forward_schedules_are_unchanged(monkeypatch, tmp_path, gr pipeline.initialize_pipeline(resolved_config) assert pipeline.modules["scheduler"].shift == 12.0 assert pipeline.modules["audio_scheduler"].shift == 3.0 - # Without a sidecar the pipeline keeps the config it was given; a checkpoint schedule would rebind it. - assert (getattr(pipeline, "resolved_config", None) or resolved_config).pipeline.dmd_denoising_steps is None + # Without a sidecar, resolution leaves the schedule unset and the pipeline accepts that. + assert _resolve_checkpoint(tmp_path).pipeline.dmd_denoising_steps is None assert config.text_encoder_precisions == ("bf16",) stage, calls, _ = _run_tiny_stage(monkeypatch, None, grid_points, video_shift=12.0) assert len(calls) == grid_points - 1 @@ -210,7 +243,7 @@ def test_explicit_eight_forward_ladder_requires_nine_grid_points(monkeypatch, gr def test_checkpoint_schedule_metadata_must_be_self_consistent(tmp_path, field, value): (tmp_path / "fastvideo_inference.json").write_text(json.dumps({**CONTRACT, field: value})) with pytest.raises(ValueError): - _pipeline(tmp_path).initialize_pipeline(make_resolved_config(MiniMaxH3PipelineConfig())) + _pipeline(tmp_path).initialize_pipeline(_with_steps(DMD_STEPS)) @pytest.mark.parametrize("module_name", ["scheduler", "audio_scheduler"]) @@ -231,12 +264,13 @@ def test_sidecar_without_shift_fields_uses_the_checkpoint_schedulers(tmp_path): four_step.pop("video_scheduler_shift") four_step.pop("audio_scheduler_shift") (tmp_path / "fastvideo_inference.json").write_text(json.dumps(four_step)) + resolved_config = _resolve_checkpoint(tmp_path) + assert list(resolved_config.pipeline.dmd_denoising_steps) == [999, 749, 500, 250] pipeline = _pipeline(tmp_path, video_shift=12.0) - pipeline.initialize_pipeline(make_resolved_config(MiniMaxH3PipelineConfig())) - assert list(pipeline.resolved_config.pipeline.dmd_denoising_steps) == [999, 749, 500, 250] + pipeline.initialize_pipeline(resolved_config) assert pipeline.modules["scheduler"].shift == 12.0 assert pipeline.modules["audio_scheduler"].shift == 3.0 # Explicit disagreement is still rejected. (tmp_path / "fastvideo_inference.json").write_text(json.dumps({**four_step, "video_scheduler_shift": 10.0})) with pytest.raises(ValueError, match="disagrees"): - _pipeline(tmp_path, video_shift=12.0).initialize_pipeline(make_resolved_config(MiniMaxH3PipelineConfig())) + _pipeline(tmp_path, video_shift=12.0).initialize_pipeline(resolved_config) diff --git a/fastvideo/tests/stages/test_minimax_h3_sequential_start.py b/fastvideo/tests/stages/test_minimax_h3_sequential_start.py index 16467a61b..e7c7978b0 100644 --- a/fastvideo/tests/stages/test_minimax_h3_sequential_start.py +++ b/fastvideo/tests/stages/test_minimax_h3_sequential_start.py @@ -199,7 +199,8 @@ def test_lazy_module_load_owns_deferral_when_both_would_arm(monkeypatch) -> None monkeypatch.setattr(ComposedPipelineBase, "load_modules", fake_load) monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", lambda device_id: True) - MiniMaxH3Pipeline("unused/for-this-test", _h3_config(sequential_load=True)) + # Resolution turns lazy_module_load on for unified memory; the pipeline reads that decision. + MiniMaxH3Pipeline("unused/for-this-test", _h3_config(lazy_module_load=True, sequential_load=True)) assert loads == [list(MiniMaxH3Pipeline._required_config_modules)] assert all(name in loads[0] for name in _DENOISE_MODULE_NAMES) diff --git a/fastvideo/tests/worker/test_gpu_worker.py b/fastvideo/tests/worker/test_gpu_worker.py index d407e7d7f..a8910fec8 100644 --- a/fastvideo/tests/worker/test_gpu_worker.py +++ b/fastvideo/tests/worker/test_gpu_worker.py @@ -41,9 +41,9 @@ def test_cuda_device_uuid_receipt_identifies_profiled_worker(monkeypatch, env_ov @pytest.mark.parametrize("executor_backend", ["mp", "ray"]) -def test_init_device_applies_offload_policy_after_binding_worker_device(monkeypatch, env_overrides, - executor_backend: str) -> None: - """The runtime probe must see this worker's device, never driver device 0, and the worker keeps its result.""" +def test_init_device_binds_the_worker_device_and_keeps_the_resolved_config(monkeypatch, env_overrides, + executor_backend: str) -> None: + """The worker binds its own device and builds the pipeline on the config it received, without a device query.""" events = [] with isolated_environment(): args = resolve_inference_config({ @@ -53,13 +53,8 @@ def test_init_device_applies_offload_policy_after_binding_worker_device(monkeypa "execution_backend": executor_backend }, }) - decided = args.with_override("test:device_policy", {"engine.offload.lazy_module_load": False}) - - def fake_policy(resolved_config, device_id): - events.append(("policy", device_id)) - return decided - - monkeypatch.setattr("fastvideo.worker.gpu_worker.finalize_device_offload_policy", fake_policy) + probe = Mock(side_effect=AssertionError("device probe ran in the worker")) + monkeypatch.setattr("fastvideo.platforms.current_platform.has_unified_memory", probe) worker = Worker(args, local_rank=3, rank=3, distributed_init_method="env://") env_overrides.enter_context(envs.override_external("LOCAL_RANK", "0")) @@ -77,23 +72,34 @@ def test_init_device_applies_offload_policy_after_binding_worker_device(monkeypa assert events == [ ("set_device", 3), - ("policy", 3), ("distributed", None), ("pipeline", None), ] assert os.environ["LOCAL_RANK"] == "3" assert worker.device == torch.device("cuda:3") assert worker.init_gpu_memory == 123 - assert worker.resolved_config is decided + assert worker.resolved_config is args + probe.assert_not_called() def _worker_returning(output_batch: ForwardBatch) -> Worker: worker = Worker.__new__(Worker) worker.resolved_config = SimpleNamespace() - worker.pipeline = SimpleNamespace(resolved_config=SimpleNamespace(), forward=lambda batch, args: output_batch) + worker.pipeline = SimpleNamespace(forward=lambda batch, args: output_batch) return worker +def test_execute_forward_runs_the_pipeline_on_the_worker_config(): + worker = Worker.__new__(Worker) + worker.resolved_config = SimpleNamespace() + seen = [] + worker.pipeline = SimpleNamespace(forward=lambda batch, args: seen.append(args) or batch) + + worker.execute_forward(ForwardBatch(data_type="video", return_frames=True), _request_config()) + + assert seen == [worker.resolved_config] + + def _request_config(output_type: str = "pil") -> SimpleNamespace: """Stand-in for the request's resolved config with the typed path that ``Worker.execute_forward`` reads.""" return SimpleNamespace(pipeline=SimpleNamespace(output_type=output_type)) diff --git a/fastvideo/train/callbacks/validation.py b/fastvideo/train/callbacks/validation.py index d0466a30c..bee45673d 100644 --- a/fastvideo/train/callbacks/validation.py +++ b/fastvideo/train/callbacks/validation.py @@ -1381,19 +1381,20 @@ class ValidationCallback(Callback): stages read. """ tc = self.training_config + validation_pipeline_config = self._validation_pipeline_config( + transformer, + pipeline.resolved_config.pipeline_config, + ) + dmd_denoising_steps = None + if (self.sampling_timesteps is not None + and getattr(validation_pipeline_config, "dmd_denoising_steps", None) is None): + dmd_denoising_steps = [int(s) for s in self.sampling_timesteps] forward_config = build_inference_resolved_config( tc, model_path=tc.model_path, - pipeline_config=self._validation_pipeline_config( - transformer, - pipeline.resolved_config.pipeline_config, - ), + pipeline_config=validation_pipeline_config, + dmd_denoising_steps=dmd_denoising_steps, ) - if (self.sampling_timesteps is not None and forward_config.pipeline.dmd_denoising_steps is None): - forward_config = forward_config.with_override( - "validation_callback:sampling_timesteps", - {"pipeline.dmd_denoising_steps": [int(s) for s in self.sampling_timesteps]}, - ) dit_config = forward_config.pipeline_config.dit_config logger.info( "Validation forward config: local_attn_size=%s sink_size=%s boundary_ratio=%s", diff --git a/fastvideo/train/utils/moduleloader.py b/fastvideo/train/utils/moduleloader.py index 729377e5a..7415bab4b 100644 --- a/fastvideo/train/utils/moduleloader.py +++ b/fastvideo/train/utils/moduleloader.py @@ -15,10 +15,7 @@ from fastvideo.api.inference_resolution import ( ) from fastvideo.api.resolution import ResolutionStep, ResolvedGeneratorConfig from fastvideo.api.schema import ExecutionMode, GeneratorConfig -from fastvideo.api.training_schema import ( - resolve_training_offload_conflicts, - validate_training_parallel_sizes, -) +from fastvideo.api.training_schema import validate_training_parallel_sizes from fastvideo.attention.selector import ( _component_attention_backend_scope, coerce_attn_backend, @@ -90,11 +87,11 @@ def _generator_config_mapping( def _training_resolution_steps(config: GeneratorConfig, defaults: Any) -> tuple[ResolutionStep, ...]: - """Generator resolution steps plus the training offload-conflict and parallel-size checks.""" + """Generator resolution steps plus the training parallel-size check.""" return generator_resolution_steps( config, defaults, - before_placeholders=(resolve_training_offload_conflicts, validate_training_parallel_sizes), + before_placeholders=(validate_training_parallel_sizes, ), ) @@ -132,13 +129,15 @@ def build_inference_resolved_config( *, model_path: str, pipeline_config: PipelineConfig | None = None, + dmd_denoising_steps: list[int] | None = None, ) -> ResolvedGeneratorConfig: """Build the inference-mode resolved config for validation forwards and standalone encoder loads. It has the training parallel layout, the DiT offloaded to the CPU, every other offload off, and ``tc.vsa_sparsity`` as the VSA sparsity. ``pipeline_config`` replaces ``tc.pipeline_config`` as the model - definition. + definition. ``dmd_denoising_steps`` sets ``pipeline.dmd_denoising_steps``, + the DMD schedule that the causal and DMD denoising stages read. """ if pipeline_config is None: pipeline_config = tc.pipeline_config if tc.pipeline_config is not None else PipelineConfig() @@ -150,6 +149,8 @@ def build_inference_resolved_config( ) raw["engine"]["offload"]["dit"] = True raw["engine"]["attention"] = {"vsa_sparsity": tc.vsa_sparsity} + if dmd_denoising_steps is not None: + raw["pipeline"]["dmd_denoising_steps"] = list(dmd_denoising_steps) return resolve_inference_config(raw) diff --git a/fastvideo/training/distillation_pipeline.py b/fastvideo/training/distillation_pipeline.py index c15b170ca..a2ae915c7 100644 --- a/fastvideo/training/distillation_pipeline.py +++ b/fastvideo/training/distillation_pipeline.py @@ -97,19 +97,23 @@ class DistillationPipeline(TrainingPipeline): distillation = resolved_config.training.distillation # The teacher (real score) and the critic (fake score) load as the base Wan transformer, whatever class the # student uses. - teacher_critic_config = resolved_config.with_override( - "DistillationPipeline.load_modules", - {"pipeline.components.override_transformer_cls_name": "WanTransformer3DModel"}) + teacher_critic_cls_name = "WanTransformer3DModel" if distillation.real_score_model_path: logger.info("Loading real score transformer from: %s", distillation.real_score_model_path) # TODO(will): can use deepcopy instead if the model is the same - self.real_score_transformer = self.load_module_from_path(distillation.real_score_model_path, "transformer", - teacher_critic_config) + self.real_score_transformer = self.load_module_from_path( + distillation.real_score_model_path, + "transformer", + resolved_config, + override_transformer_cls_name=teacher_critic_cls_name) modules["real_score_transformer"] = self.real_score_transformer try: - self.real_score_transformer_2 = self.load_module_from_path(distillation.real_score_model_path, - "transformer_2", teacher_critic_config) + self.real_score_transformer_2 = self.load_module_from_path( + distillation.real_score_model_path, + "transformer_2", + resolved_config, + override_transformer_cls_name=teacher_critic_cls_name) logger.info("Loaded real score transformer_2 for MoE support") modules["real_score_transformer_2"] = self.real_score_transformer_2 except Exception: @@ -120,12 +124,18 @@ class DistillationPipeline(TrainingPipeline): if distillation.fake_score_model_path: logger.info("Loading fake score transformer from: %s", distillation.fake_score_model_path) - self.fake_score_transformer = self.load_module_from_path(distillation.fake_score_model_path, "transformer", - teacher_critic_config) + self.fake_score_transformer = self.load_module_from_path( + distillation.fake_score_model_path, + "transformer", + resolved_config, + override_transformer_cls_name=teacher_critic_cls_name) modules["fake_score_transformer"] = self.fake_score_transformer try: - self.fake_score_transformer_2 = self.load_module_from_path(distillation.fake_score_model_path, - "transformer_2", teacher_critic_config) + self.fake_score_transformer_2 = self.load_module_from_path( + distillation.fake_score_model_path, + "transformer_2", + resolved_config, + override_transformer_cls_name=teacher_critic_cls_name) logger.info("Loaded fake score transformer_2 for MoE support") modules["fake_score_transformer_2"] = self.fake_score_transformer_2 except Exception: @@ -265,7 +275,11 @@ class DistillationPipeline(TrainingPipeline): else: logger.info("Generator EMA disabled (ema_decay <= 0.0)") - def load_module_from_path(self, model_path: str, module_type: str, resolved_config: ResolvedGeneratorConfig): + def load_module_from_path(self, + model_path: str, + module_type: str, + resolved_config: ResolvedGeneratorConfig, + override_transformer_cls_name: str | None = None): """ Load a teacher (real score) or critic (fake score) module from a specific path using the same loading logic as the pipeline. @@ -276,6 +290,8 @@ class DistillationPipeline(TrainingPipeline): model_path: Path to the model module_type: Type of module to load (e.g., "transformer") resolved_config: The config to load the module with + override_transformer_cls_name: The transformer class to build, over + ``pipeline.components.override_transformer_cls_name`` Returns: The loaded module @@ -311,6 +327,7 @@ class DistillationPipeline(TrainingPipeline): transformers_or_diffusers=transformers_or_diffusers, resolved_config=resolved_config, loading_teacher_critic_model=True, + override_transformer_cls_name=override_transformer_cls_name, ) logger.info("Successfully loaded %s from %s", module_type, component_path) diff --git a/fastvideo/training/matrixgame2_self_forcing_distillation_pipeline.py b/fastvideo/training/matrixgame2_self_forcing_distillation_pipeline.py index 2e40e21be..34b1f66fb 100644 --- a/fastvideo/training/matrixgame2_self_forcing_distillation_pipeline.py +++ b/fastvideo/training/matrixgame2_self_forcing_distillation_pipeline.py @@ -57,10 +57,7 @@ class MatrixGame2SelfForcingDistillationPipeline(SelfForcingDistillationPipeline loaded_modules, ) # The teacher (real score) and critic (fake score) load as the bidirectional MatrixGame2WanModel class. - teacher_config = resolved_config.with_override( - "MatrixGame2SelfForcingDistillationPipeline.load_modules", - {"pipeline.components.override_transformer_cls_name": "MatrixGame2WanModel"}, - ) + teacher_cls_name = "MatrixGame2WanModel" if loaded_modules is not None and "real_score_transformer" in loaded_modules: self.real_score_transformer = loaded_modules["real_score_transformer"] elif resolved_config.training.distillation.real_score_model_path: @@ -71,7 +68,8 @@ class MatrixGame2SelfForcingDistillationPipeline(SelfForcingDistillationPipeline self.real_score_transformer = self.load_module_from_path( resolved_config.training.distillation.real_score_model_path, "transformer", - teacher_config, + resolved_config, + override_transformer_cls_name=teacher_cls_name, ) else: raise ValueError("real_score_model_path is required for DMD distillation pipeline") @@ -87,7 +85,8 @@ class MatrixGame2SelfForcingDistillationPipeline(SelfForcingDistillationPipeline self.fake_score_transformer = self.load_module_from_path( resolved_config.training.distillation.fake_score_model_path, "transformer", - teacher_config, + resolved_config, + override_transformer_cls_name=teacher_cls_name, ) else: raise ValueError("fake_score_model_path is required for DMD distillation pipeline") diff --git a/fastvideo/worker/gpu_worker.py b/fastvideo/worker/gpu_worker.py index f4a1f7bb3..ab9a9e59b 100644 --- a/fastvideo/worker/gpu_worker.py +++ b/fastvideo/worker/gpu_worker.py @@ -4,7 +4,6 @@ from typing import Any, cast import torch import fastvideo.envs as envs -from fastvideo.api.device_policy import finalize_device_offload_policy from fastvideo.api.resolution import ResolvedGeneratorConfig from fastvideo.distributed import (cleanup_dist_env_and_memory, maybe_init_distributed_environment_and_model_parallel) from fastvideo.distributed.parallel_state import get_local_torch_device @@ -80,13 +79,6 @@ class Worker: # For MPS, we can't get memory info the same way self.init_gpu_memory = 0 - # CUDA's unified-memory classification reads runtime device - # properties, so make this decision only after this worker has bound - # its own device. The worker keeps the config that the policy returns, - # and every loader and pipeline stage below consumes it. - device_id = self.device.index if self.device.index is not None else 0 - self.resolved_config = finalize_device_offload_policy(self.resolved_config, device_id) - # Initialize the distributed environment. maybe_init_distributed_environment_and_model_parallel(self.resolved_config.engine.parallelism.tp_size, self.resolved_config.engine.parallelism.sp_size, @@ -95,8 +87,7 @@ class Worker: self.pipeline = build_pipeline(self.resolved_config) def execute_forward(self, forward_batch: ForwardBatch, resolved_config: ResolvedGeneratorConfig) -> ForwardBatch: - # The pipeline's config carries the overrides that the pipeline made from its checkpoint. - output_batch = self.pipeline.forward(forward_batch, self.pipeline.resolved_config) + output_batch = self.pipeline.forward(forward_batch, self.resolved_config) needs_output = forward_batch.return_frames or (forward_batch.save_video and resolved_config.pipeline.output_type != "latent" and not output_batch.extra.get("audio_only")) diff --git a/tests/local_tests/flux2/README.md b/tests/local_tests/flux2/README.md index 7ee014d06..aad931e50 100644 --- a/tests/local_tests/flux2/README.md +++ b/tests/local_tests/flux2/README.md @@ -83,7 +83,7 @@ pytest tests/local_tests/flux2/test_flux2_component_parity.py::test_flux2_mistra | Full pipeline smoke | [`../pipelines/test_flux2_pipeline_smoke.py`](../pipelines/test_flux2_pipeline_smoke.py) | Full Flux2 Mistral3/AutoProcessor wiring; short latent generate | `PASSED on Modal L40S:2` | | Pipeline parity | [`../pipelines/test_flux2_pipeline_parity.py`](../pipelines/test_flux2_pipeline_parity.py) | Four-step denoised latent parity vs `diffusers.Flux2KleinPipeline` | `PASSED on Modal L40S` | | Full pipeline parity | [`../pipelines/test_flux2_pipeline_parity.py`](../pipelines/test_flux2_pipeline_parity.py) | Short full Flux2 latent parity vs `diffusers.Flux2Pipeline` | `PASSED on Modal L40S:2, L40S:4, and H100:1` | -| Pipeline TP2 parity | [`../pipelines/test_flux2_pipeline_parity.py`](../pipelines/test_flux2_pipeline_parity.py) | Two-worker tensor-parallel load/generate with `num_gpus=2`, `tp_size=2`, `sp_size=1` | `PASSED on Modal L40S:2` | +| Pipeline TP2 parity | [`../pipelines/test_flux2_pipeline_parity.py`](../pipelines/test_flux2_pipeline_parity.py) | Two-worker tensor-parallel load/generate with `engine.num_gpus: 2`, `tp_size=2`, `sp_size=1` | `PASSED on Modal L40S:2` | | Pixel image comparison | Modal image runner | Same-prompt Diffusers vs FastVideo PNG generation and pixel metrics | `PASSED on Modal L40S` | ## Latest Remote Evidence @@ -116,7 +116,7 @@ Flux2 modal L40S:2 statuses: component=0 smoke=0 pipeline=0 ``` The `L40S:2` run verifies the same parity suite in a two-GPU Modal allocation. -These tests currently instantiate FastVideo with `num_gpus=1`, so this is not a +These tests currently instantiate FastVideo with `engine.num_gpus: 1`, so this is not a tensor-parallel two-GPU parity test. Additional Modal `L40S:4` allocation run `ap-tQEUnFr00uOvpMZdOngebQ` applied @@ -132,7 +132,7 @@ Flux2 modal L40S:4 statuses: component=0 smoke=0 pipeline=0 Modal `L40S:2` tensor-parallel run `ap-szNgcJRiUv11lmmNFvjjPT` applied patch `patches/flux2-local-tp2-c850fcad.patch`, confirmed two visible L40S devices, -and instantiated FastVideo with `num_gpus=2`, `tp_size=2`, `sp_size=1`, +and instantiated FastVideo with `engine.num_gpus: 2`, `tp_size=2`, `sp_size=1`, `executor_world_size=2`. ```text @@ -172,7 +172,7 @@ tests/local_tests/pipelines/test_flux2_pipeline_parity.py -v -s: 2 passed The pipeline parity file covered both strict single-GPU latent parity and true TP2 latent parity. The strict path reported zero max/mean/median diff for all four trajectory steps and final latents. The TP2 path instantiated FastVideo -with `num_gpus=2`, `tp_size=2`, `sp_size=1`, and `executor_world_size=2`. +with `engine.num_gpus: 2`, `tp_size=2`, `sp_size=1`, and `executor_world_size=2`. ```text TP2 final diff max=2.107544 mean=0.020110 median=0.015625 @@ -231,7 +231,7 @@ Current-session Modal evidence: `0.062500`, mean `0.008581`, median `0.007812`; only `9/8192` latent entries hit the max bucket. - Modal `L40S:4` run `ap-KKOzv9THDmTYt3gevhQkVR` passed full pipeline latent - parity with true TP4 (`num_gpus=4`, `tp_size=4`, `sp_size=1`). Final latent + parity with true TP4 (`engine.num_gpus: 4`, `tp_size=4`, `sp_size=1`). Final latent diff max stayed `0.062500`, mean was `0.007946`, and median was `0.007812`. - Modal `L40S:4` diagnostic run `ap-SlhykTTLxzsYxnsCUbCH7S` reproduced the TP4 result with max `0.062500`, mean `0.007946`, median `0.007812`; only `6/8192` diff --git a/tests/local_tests/gamecraft/test_gamecraft_pipeline_parity.py b/tests/local_tests/gamecraft/test_gamecraft_pipeline_parity.py index fbe9cb4a7..7955306b6 100644 --- a/tests/local_tests/gamecraft/test_gamecraft_pipeline_parity.py +++ b/tests/local_tests/gamecraft/test_gamecraft_pipeline_parity.py @@ -330,12 +330,18 @@ def test_gamecraft_pipeline_latent_parity(): # Create generator generator = VideoGenerator.from_pretrained( diffusers_path, - num_gpus=1, - use_fsdp_inference=False, - dit_cpu_offload=False, - vae_cpu_offload=True, # Save memory - text_encoder_cpu_offload=True, - pin_cpu_memory=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "dit": False, + "vae": True, # Save memory + "text_encoder": True, + "pin_cpu_memory": False, + }, + }, + }, ) # Run generation @@ -481,12 +487,18 @@ def test_gamecraft_pipeline_smoke(): generator = VideoGenerator.from_pretrained( diffusers_path, - num_gpus=1, - use_fsdp_inference=False, - dit_cpu_offload=True, - vae_cpu_offload=True, - text_encoder_cpu_offload=True, - pin_cpu_memory=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "dit": True, + "vae": True, + "text_encoder": True, + "pin_cpu_memory": False, + }, + }, + }, ) result = generator.generate({ diff --git a/tests/local_tests/gen3c/test_gen3c_pipeline_smoke.py b/tests/local_tests/gen3c/test_gen3c_pipeline_smoke.py index 23c77a1ec..a692597a0 100644 --- a/tests/local_tests/gen3c/test_gen3c_pipeline_smoke.py +++ b/tests/local_tests/gen3c/test_gen3c_pipeline_smoke.py @@ -214,11 +214,17 @@ def test_gen3c_full_pipeline_smoke(): generator = VideoGenerator.from_pretrained( diffusers_path, - num_gpus=1, - use_fsdp_inference=False, - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "dit": False, + "vae": False, + "text_encoder": False, + }, + }, + }, ) result = generator.generate({ "prompt": prompt, diff --git a/tests/local_tests/glm_image/test_edit_pipeline_parity.py b/tests/local_tests/glm_image/test_edit_pipeline_parity.py index df8a4fbb6..fbec083f9 100644 --- a/tests/local_tests/glm_image/test_edit_pipeline_parity.py +++ b/tests/local_tests/glm_image/test_edit_pipeline_parity.py @@ -98,7 +98,15 @@ def _fastvideo_edit_image(device) -> np.ndarray: except ImportError as e: pytest.skip(f"FastVideo VideoGenerator unavailable: {e}") condition = Image.open(CONDITION_IMAGE).convert("RGB") - gen = VideoGenerator.from_pretrained(str(LOCAL_WEIGHTS_DIR), num_gpus=1, trust_remote_code=True) + gen = VideoGenerator.from_pretrained( + str(LOCAL_WEIGHTS_DIR), + { + "engine": { + "num_gpus": 1, + }, + "trust_remote_code": True, + }, + ) result = gen.generate({ "prompt": EDIT_PROMPT, "inputs": {"pil_image": condition}, diff --git a/tests/local_tests/glm_image/test_pipeline_parity.py b/tests/local_tests/glm_image/test_pipeline_parity.py index bd83d7625..9be0406ff 100644 --- a/tests/local_tests/glm_image/test_pipeline_parity.py +++ b/tests/local_tests/glm_image/test_pipeline_parity.py @@ -200,7 +200,15 @@ def _fastvideo_image(device) -> np.ndarray: from fastvideo import VideoGenerator except ImportError as e: pytest.skip(f"FastVideo VideoGenerator unavailable: {e}") - gen = VideoGenerator.from_pretrained(str(LOCAL_WEIGHTS_DIR), num_gpus=1, trust_remote_code=True) + gen = VideoGenerator.from_pretrained( + str(LOCAL_WEIGHTS_DIR), + { + "engine": { + "num_gpus": 1, + }, + "trust_remote_code": True, + }, + ) result = gen.generate({ "prompt": SAMPLE_PROMPT, "sampling": { diff --git a/tests/local_tests/magi_human/test_magi_human_pipeline_smoke.py b/tests/local_tests/magi_human/test_magi_human_pipeline_smoke.py index c4a8715e4..d1ef79e5d 100644 --- a/tests/local_tests/magi_human/test_magi_human_pipeline_smoke.py +++ b/tests/local_tests/magi_human/test_magi_human_pipeline_smoke.py @@ -212,11 +212,17 @@ def test_magi_human_pipeline_smoke() -> None: generator = VideoGenerator.from_pretrained( diffusers_path, - num_gpus=1, - use_fsdp_inference=False, - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "dit": False, + "vae": False, + "text_encoder": False, + }, + }, + }, ) try: result = generator.generate({ diff --git a/tests/local_tests/pipelines/test_lingbot_video_pipeline_smoke.py b/tests/local_tests/pipelines/test_lingbot_video_pipeline_smoke.py index 7325f060d..e0d8696e1 100644 --- a/tests/local_tests/pipelines/test_lingbot_video_pipeline_smoke.py +++ b/tests/local_tests/pipelines/test_lingbot_video_pipeline_smoke.py @@ -146,15 +146,25 @@ def test_lingbot_video_pipeline_load_generate_smoke(tmp_path: Path) -> None: ) generator = VideoGenerator.from_pretrained( str(model_dir), - num_gpus=num_gpus, - sp_size=sp_size, - use_fsdp_inference=use_fsdp_inference, - dit_cpu_offload=False, - dit_layerwise_offload=False, - vae_cpu_offload=True, - text_encoder_cpu_offload=True, - pin_cpu_memory=False, - output_type="latent", + { + "engine": { + "num_gpus": num_gpus, + "parallelism": { + "sp_size": sp_size, + }, + "use_fsdp_inference": use_fsdp_inference, + "offload": { + "dit": False, + "dit_layerwise": False, + "vae": True, + "text_encoder": True, + "pin_cpu_memory": False, + }, + }, + "pipeline": { + "output_type": "latent", + }, + }, ) try: result = generator.generate({ diff --git a/tests/local_tests/pipelines/test_zimage_pipeline_parity.py b/tests/local_tests/pipelines/test_zimage_pipeline_parity.py index ecf61b175..851d55d27 100644 --- a/tests/local_tests/pipelines/test_zimage_pipeline_parity.py +++ b/tests/local_tests/pipelines/test_zimage_pipeline_parity.py @@ -290,16 +290,26 @@ def test_zimage_pipeline_latents_match_pinned_native_repo(tmp_path: Path) -> Non generator = VideoGenerator.from_pretrained( str(MODEL_DIR), - num_gpus=1, - tp_size=1, - sp_size=1, - use_fsdp_inference=False, - dit_cpu_offload=False, - dit_layerwise_offload=False, - text_encoder_cpu_offload=False, - vae_cpu_offload=False, - pin_cpu_memory=False, - output_type="latent", + { + "engine": { + "num_gpus": 1, + "parallelism": { + "tp_size": 1, + "sp_size": 1, + }, + "use_fsdp_inference": False, + "offload": { + "dit": False, + "dit_layerwise": False, + "text_encoder": False, + "vae": False, + "pin_cpu_memory": False, + }, + }, + "pipeline": { + "output_type": "latent", + }, + }, ) try: result = generator.generate({ diff --git a/tests/local_tests/pipelines/test_zimage_pipeline_smoke.py b/tests/local_tests/pipelines/test_zimage_pipeline_smoke.py index 50546c072..3d0dd1fda 100644 --- a/tests/local_tests/pipelines/test_zimage_pipeline_smoke.py +++ b/tests/local_tests/pipelines/test_zimage_pipeline_smoke.py @@ -115,16 +115,26 @@ def test_zimage_pipeline_load_generate_smoke() -> None: generator = VideoGenerator.from_pretrained( str(model_dir), - num_gpus=1, - tp_size=1, - sp_size=1, - use_fsdp_inference=False, - dit_cpu_offload=False, - dit_layerwise_offload=False, - text_encoder_cpu_offload=False, - vae_cpu_offload=False, - pin_cpu_memory=False, - output_type="latent", + { + "engine": { + "num_gpus": 1, + "parallelism": { + "tp_size": 1, + "sp_size": 1, + }, + "use_fsdp_inference": False, + "offload": { + "dit": False, + "dit_layerwise": False, + "text_encoder": False, + "vae": False, + "pin_cpu_memory": False, + }, + }, + "pipeline": { + "output_type": "latent", + }, + }, ) try: result = generator.generate({ diff --git a/tests/local_tests/stable_audio/test_stable_audio_a2a_parity.py b/tests/local_tests/stable_audio/test_stable_audio_a2a_parity.py index 920b4a9cb..a2f2cbf8b 100644 --- a/tests/local_tests/stable_audio/test_stable_audio_a2a_parity.py +++ b/tests/local_tests/stable_audio/test_stable_audio_a2a_parity.py @@ -113,11 +113,17 @@ def test_stable_audio_a2a_official_parity(): from fastvideo import VideoGenerator generator = VideoGenerator.from_pretrained( _FV_REPO_ID, - num_gpus=1, - use_fsdp_inference=False, - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "dit": False, + "vae": False, + "text_encoder": False, + }, + }, + }, ) try: result = generator.generate({ diff --git a/tests/local_tests/stable_audio/test_stable_audio_inpaint_parity.py b/tests/local_tests/stable_audio/test_stable_audio_inpaint_parity.py index 6753fbeda..ac9146ee2 100644 --- a/tests/local_tests/stable_audio/test_stable_audio_inpaint_parity.py +++ b/tests/local_tests/stable_audio/test_stable_audio_inpaint_parity.py @@ -69,11 +69,17 @@ def test_stable_audio_inpaint_kept_region_preserved(): from fastvideo import VideoGenerator generator = VideoGenerator.from_pretrained( _FV_REPO_ID, - num_gpus=1, - use_fsdp_inference=False, - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "dit": False, + "vae": False, + "text_encoder": False, + }, + }, + }, ) try: result = generator.generate({ diff --git a/tests/local_tests/stable_audio/test_stable_audio_pipeline_parity.py b/tests/local_tests/stable_audio/test_stable_audio_pipeline_parity.py index 2ef2675f1..8a91b64c3 100644 --- a/tests/local_tests/stable_audio/test_stable_audio_pipeline_parity.py +++ b/tests/local_tests/stable_audio/test_stable_audio_pipeline_parity.py @@ -161,11 +161,17 @@ def test_stable_audio_pipeline_official_parity(): from fastvideo import VideoGenerator generator = VideoGenerator.from_pretrained( _FV_REPO_ID, - num_gpus=1, - use_fsdp_inference=False, - dit_cpu_offload=False, - vae_cpu_offload=False, - text_encoder_cpu_offload=False, + { + "engine": { + "num_gpus": 1, + "use_fsdp_inference": False, + "offload": { + "dit": False, + "vae": False, + "text_encoder": False, + }, + }, + }, ) try: result = generator.generate({