pipeline.model is a tagged union: a mapping with exactly one key, which
names the model family (ltx2, minimax_h3, longcat, or generic) and selects
that family's options class. The former sibling sections pipeline.ltx2,
pipeline.minimax_h3, and pipeline.longcat, and the DiT/VAE architecture
override dicts pipeline.dit and pipeline.vae, live inside the family block.
The parser implements tagged unions as a general mechanism (a dataclass
with a TAGS table); the family is readable as pipeline.model.family.
Resolution validates the tag against the family that the registry chose
for model_path (validate_model_family raises and names the registry class
and the right block) and fills the family's empty block when the input has
none (fill_model_family), so readers never see None. Family-only fills and
derivations (the MiniMax-H3 parallel-VAE environment variables, the LTX-2
tile-size derivation and refine preset copy, the LongCat BSA mirrors) run
only for their family.
Against the 3f6893a0 goldens every PipelineConfig is identical and no
other field differs beyond the allowed categories; family-only settings
are no longer decided on models of another family. LingBot-Video reads
its refiner switch from pipeline.preset_overrides.refine.enabled and loads
the refiner by default when the checkpoint has one, as at 3f6893a0.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CbSuTKA9oX4YsECdcBPAsX
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 <noreply@anthropic.com>
ResolvedGeneratorConfig is the only runtime config. It answers only its
typed fields, pipeline_config, mode, training_mode, and inference_mode;
the transitional flat-name fallback is gone.
Deleted: fastvideo/fastvideo_args.py (FastVideoArgs, TrainingArgs, their
argparse, prepare/set/get_current_fastvideo_args), the flat-name fallback,
generator_config_to_fastvideo_args and the other flattening helpers,
PipelineConfig/model-config/PreprocessConfig/SamplingParam add_cli_args,
configs/utils.py, and the tests of the deleted code.
Materialization builds the model's PipelineConfig directly from the typed
values through an explicit table of typed path to PipelineConfig mirror
attribute (PIPELINE_CONFIG_MIRRORS); PipelineConfig.from_source replaces
from_kwargs. flat_field metadata stays only on the fields behind the 24
from_pretrained keywords; a pipeline.experimental key that names a mirror
attribute, and pipeline.preset / preset_version / components.vae_weights,
are rejected during resolution. The training root keeps the DiT on the
device through a resolution step instead of a from_pretrained override.
The goldens snapshot the end-state objects (resolved config, PipelineConfig,
decisions, request); the cli category is gone. Against the 3f6893a0
goldens, every field matches except disable_autocast, boundary_ratio,
ltx2_vae_tiling (DESIGN section 9), and the fields that no longer exist.
The schema parity inventory classifies the typed GeneratorConfig paths.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
- The component loaders read engine.offload.*, engine.precision.*,
engine.parallelism.hsdp_*, engine.use_fsdp_inference, engine.compile.*,
engine.attention.*, pipeline.components.*, and pipeline.flow_shift (the
scheduler shift) from the resolved config, and compile settings through
torch_compile_kwargs(). They no longer write model_paths (the pipeline's
ComponentState records paths) or apply the unified-memory policy to the
config; the text-encoder loader only queries the policy for its target
device.
- PipelineComponentLoader.load_module and TransformerLoader.load take
loading_teacher_critic_model instead of reading an attribute that
training code set on the config.
- Loader, encoder, transformer, VAE, and device-policy tests build resolved
configs; the env-var docs name typed paths.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
- Docs, AGENTS.md files, and skills describe ResolvedGeneratorConfig,
resolve_inference_config, typed YAML, and --config plus dotted
overrides for training and preprocessing instead of FastVideoArgs and
flat flags.
- The local tests build resolved configs (or small fakes that expose typed
paths) instead of FastVideoArgs; the LoRA extraction scripts build
pipelines with from_pretrained(model_path, resolved_config=...).
- The mixkit README's NVFP4 snippet uses from_config with
engine.quantization.transformer_quant (the from_pretrained keyword it
used is rejected).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
VideoGenerator.from_pretrained accepts the 24 common engine and offload
keywords in FROM_PRETRAINED_KWARGS. Any other keyword raises TypeError
with the config path to use with VideoGenerator.from_config. The
conversion of the removed flat keywords (LTX-2 refine names, component
config prefixes, pipeline_config objects, empty-string paths) is removed
from from_pretrained_kwargs_to_config.
The golden kwargs cases that used removed keywords move to a config
category of typed configs; their golden files are byte-identical. The
DreamVerse contract test and the GPU pool forward-translation tests
checked only the removed keyword mapping; the GPU pool tests keep the
typed-config flattening checks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
VideoGenerator.generate(request) is the one inference entry point. The
generate_video keyword call, its mapping from flat keywords to a request
(legacy_generate_call_to_request and its helpers), and the tests of that
mapping are removed. Comments and docs that named generate_video describe
generate and request extensions instead.
scripts/ltx2_sr_alignment.py keeps its generate_video call, because its
reference run imports the separate FastVideo-internal package.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
DreamVerse, FastVideo Studio, the ComfyUI node, the README, and the
inference docs call VideoGenerator.generate with nested requests and
VideoGenerator.from_config with typed configs, and read GenerationResult
attributes instead of result dict keys.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
docs/inference/configuration.md describes resolved_config.provenance(path)
on a VideoGenerator and resolved_request.provenance(path) on a
GenerationResult, with examples checked against the Wan2.1 T2V 1.3B preset.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
A request's embedded_cfg_scale (request extension or legacy generate_video
keyword) was applied to a deep copy of FastVideoArgs.pipeline_config in
VideoGenerator. The worker runs the pipeline with its own FastVideoArgs
(GpuWorker.execute_forward passes self.fastvideo_args), so the copy never
reached the denoising stages and the request value was ignored.
embedded_cfg_scale is now a per-request sampling value:
- the request field request.sampling.embedded_cfg_scale (the extension key
still works);
- SamplingParam.embedded_cfg_scale and ForwardBatch.embedded_cfg_scale;
- embedded_cfg_scale_for_batch(batch, fastvideo_args), which the denoising,
SR denoising, Wan DMD, and Flux2 text-encoding stages read. It returns the
batch value, else pipeline_config.embedded_cfg_scale, so training and
requests without the field are unchanged.
VideoGenerator no longer copies or changes FastVideoArgs per request, and
request_to_pipeline_overrides is removed. The golden snapshots record the new
SamplingParam field, and the request with extensions.embedded_cfg_scale now
carries 3.0 on SamplingParam instead of on a pipeline override.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
An unsupported FASTVIDEO_ATTENTION_BACKEND name, such as a typo, was ignored
and the run fell back to automatic backend selection without telling the
user. An explicit attention_backend argument with the same name raised.
get_env_variable_attn_backend now raises ValueError for an unsupported name,
using the same parse as an explicit request (coerce_attn_backend). Every
reader goes through it: the fill_attention_backend_from_env resolution
step, FastVideoArgs.__post_init__, and the attention selector's
environment fallbacks.
The golden case environment/attention_backend_unknown_name now records the
error, and the registry description and env_vars.md table say that an
unsupported name raises.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
The from_pretrained keyword ltx2_refine_lora_path set
pipeline.components.lora_path, which reaches FastVideoArgs.lora_path and loads
as the main LoRA on the stage-1 transformer. The refine stage reads
FastVideoArgs.ltx2_refine_lora_path, which stayed unset. The conversion also
turned an empty string into None. An empty string is how callers such as the
Dreamverse GPU pool disable the refine LoRA, and None makes LTX2Pipeline load
the checkpoint default instead: fastvideo_refine_lora_path in model_index.json
is FastVideo/LTX2-Distilled-LoRA for the LTX-2 Diffusers checkpoints.
The refine LoRA gets its own field, pipeline.ltx2.refine.lora_path, whose
flat name is ltx2_refine_lora_path. None keeps the checkpoint default, and an
empty string disables the refine LoRA. The Dreamverse contract test, the
Dynamo contract example, and the parity inventory point at the field.
Golden changes: kwargs/ltx2_refine_lora_path now records the path on
ltx2_refine_lora_path instead of lora_path, and the added case
kwargs/ltx2_refine_lora_disabled pins the empty string.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
Parameters that belong to one model family, or to one component config,
could be set only through argparse flags or pipeline.experimental. They gain
typed GeneratorConfig fields, each with the flat name that FastVideoArgs or
PipelineConfig uses:
- pipeline.dit and pipeline.vae: dicts of DiT and VAE config field
overrides, passed as dit_config.<key> and vae_config.<key>
- pipeline.ltx2: VAE tile sizes, initial video and audio latent paths, noise
order, distilled sigmas, and refine.{transformer,noise,audio_noise}_path
- pipeline.minimax_h3: sequential load, video decode backend, TAEH3
checkpoint and chunk size, and the sequence-parallel VAE switches
- pipeline.longcat: the block sparse attention (BSA) settings
legacy_from_pretrained_to_config routes dit_config.* and vae_config.*
keywords into the new dicts. Every new scalar field defaults to None and is
not passed to FastVideoArgs while unset. Behavior is unchanged: the golden
config snapshots match.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
These inference parameters could be set only through argparse flags or the
untyped pipeline.experimental dict. Each one gains a typed GeneratorConfig
field whose flat name is its FastVideoArgs or PipelineConfig field, so YAML,
dotted CLI overrides, and from_pretrained keywords all reach it:
- engine.attention: backend, vsa_sparsity, vsa_tile_size, moba_config_path
- engine.precision: dit, vae, vae_decode, image_encoder, text_encoders
- engine.compile.regional (inference_torch_compile)
- pipeline: vae_sp, flow_shift, embedded_cfg_scale, dmd_denoising_steps
- pipeline.components.lora_target_modules
Every field defaults to None, which keeps the runtime or model default; a
None field is not passed to FastVideoArgs. pipeline.experimental still
accepts the flat names and still wins over the typed fields. The parity
inventory lists the typed paths, and the docs show the typed forms.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
Accept the causal-VAE-aligned 362-frame bucket (15.083 s) for 15-second H3 requests.
- Hoist MINIMAX_H3_MIN/MAX_ALIGNED_FRAMES next to align_num_frames and reuse them in the CUDA stage, the MLX runtime, and the LoRA example so the bound is consistent across entry points.
- State the accepted frame range in the error messages.
- Cover seconds="15" -> 362 at OpenAI admission and the MLX resolve_geometry bound.
- Update the Spark/openai cookbook frame caps from 345 to 362.
Known gaps: no GPU-lane coverage for the 362 bucket (SSIM runs 124 frames; the golden gate is a fixed-geometry fingerprint), and explicit num_frames=360 still hits the pre-existing grid gate.
