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 12 fastvideo/training/*_pipeline.py entry points load a
TrainingRunConfig from --config <yaml> plus dotted overrides (mode
finetuning or distillation) instead of TrainingArgs argparse flags.
- The training pipelines read self.resolved_config.training.<section>.<field>
and typed paths; the training_args alias, the inference_mode toggles,
and the writes to the training config are gone. Teacher and critic
models load with a with_override config and the loading_teacher_critic_model
argument; validation pipelines are built from their own resolved config
(resolve_validation_config), and the tracker receives to_dict().
- Each of the 37 training launchers has a YAML file next to it with the
values its flags held; shell variables stay as dotted overrides. Against
the 3f6893a0 TrainingArgs snapshots every value and PipelineConfig
matches, except mode/inference_mode (the stage readers saw training mode
before and after), the dropped flags without a reader, lora_alpha (now
derived from the rank by a step), and the vae_tiling and compile-kwargs
defaults.
- Validation stages read the validation pipeline's config: modules that
validation loads itself use the inference offload defaults, and the
bf16-autocast VAE decode of two distillation launchers becomes fp32.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
- v1_preprocess.py and v1_preprocessing_new.py load a PreprocessRunConfig
from --config <yaml> plus dotted overrides; their argparse flags are
gone. The preprocessing pipelines, workflows, datasets, and stages read
resolved_config.preprocess.* and typed paths instead of
preprocess_config and an argparse namespace.
- The overfit preprocessors build resolved configs instead of FastVideoArgs.
- Each preprocessing launcher has a YAML file next to it with the values
its flags held; shell variables stay as dotted overrides. Against the
3f6893a0 snapshots of the 17 launchers, every value and PipelineConfig
matches, except mode (preprocess instead of inference), ltx2_vae_tiling,
and text_only's vae_config.load_encoder. Five launchers that the old
parser rejected now carry their intended values; the never-defined
--model_type and --validation_dataset_file flags are dropped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
- moduleloader builds the component-loading config and the inference
config as resolved configs (_build_training_resolved_config and
build_inference_resolved_config) instead of TrainingArgs objects mutated
after construction; the transformer override, weights path, and
teacher/critic flag are config values or load_module arguments.
keep_checkpoint_component_config keeps the trainer's PipelineConfig view
of the checkpoint-filled VAE and text-encoder configs.
- The validation callback builds its pipeline with
from_pretrained(path, resolved_config=...) and forwards with a resolved
config built per run; it no longer replaces or mutates the trainer's
pipeline_config, and generator_overrides takes nested GeneratorConfig
fields instead of flat names.
- Against the 3f6893a0 snapshots of the 67 trainer YAML files, every
PipelineConfig and value matches except ltx2_vae_tiling (the model's
vae_tiling) and the unread pretrained_model_name_or_path.
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.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
Every example calls VideoGenerator.generate with a nested request instead
of generate_video keywords, and passes settings outside the from_pretrained
convenience keywords through VideoGenerator.from_config at their typed
paths. Each request and generator config equals what the keyword call built.
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
The schema class MiniMaxH3Config had the same name as the MiniMax-H3 DiT
config class in fastvideo/configs/models/dits/minimax_h3.py. The family
partitions of PipelineSelection are renamed LTX2Options, LTX2RefineOptions,
MiniMaxH3Options, and LongCatOptions, so their names cannot be confused with
the model and pipeline config classes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
Every key that repository YAML configs, Python examples, the Dreamverse
backend, and tests passed through pipeline.experimental has a typed
GeneratorConfig field, so they use those fields instead:
engine.attention, engine.compile.regional, pipeline.flow_shift,
pipeline.embedded_cfg_scale, pipeline.dmd_denoising_steps,
pipeline.components.lora_nickname, pipeline.minimax_h3, and
pipeline.longcat. No repository config outside the tests uses
pipeline.experimental any more.
The flat FastVideoArgs keywords that each migrated YAML file and each
basic_fasth3.py option set produces match the previous ones. The golden
snapshots of the two Hunyuan scripts now record embedded_cfg_scale and
flow_shift as floats (6.0 instead of 6), because the typed fields are
floats. The guidance tensor is built as float32 either way. The golden case
untyped_experimental_keys is renamed attention_precision_and_flow_shift with
identical content, and keys_without_typed_fields covers keywords that still
go to pipeline.experimental.
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.
