The configurable recipe had no page, no backend-model mapping and no sidebar entry, so the directory treated it as
example-only and the deployment link redirected to a missing path. The recipe task now groups under video upscale.
With dit_cpu_offload and no layerwise offload, the refine stage never staged the transformer like the stock H3 stage
does, so it could stay resident through the 2K VAE decode. It now moves it to the device for the refine and back to the
CPU in finally, and releases layerwise-offload buffers there too.
FASTVIDEO_H3_ADALN_CACHE and FASTVIDEO_H3_ADALN_TABLE compute the AdaLN modulation with F.linear instead of the
projection modules, so the Sharp curve hooks never ran and refinement silently lost all 50 branches. The pipeline now
refuses both before loading the transformer.
`fastvideo generate` takes a run config plus dotted overrides and never reads FastVideoArgs.add_cli_args, so the
documented `--h3-vae-nvfp4-decoder` flag failed before the model loaded. The flag is removed; the guide uses
`--generator.pipeline.experimental.h3_vae_nvfp4_decoder`, and a test drives that override through the generate parser
to FastVideoArgs. Also drops two narrating test docstrings.
The FSDP (DTensor) merge path still cast the adapter to the BF16 base before multiplying, so an FP32 adapter merged
differently with FSDP on. Both paths now share one helper, which also moves the factors to the weight's device.
attach() remembered the lazy proxy's id, so after the transformer was released and loaded again the new instance
never got its 50 curve hooks and later requests ran without them. The branches now attach to the materialized
module and mark that module, so each loaded instance gets its own hooks.
The pipeline already loads its base_model VAE through VAELoader, so the
existing h3_vae_nvfp4_decoder swap applies; this exposes it: a
--h3-vae-nvfp4-decoder CLI flag, a --vae-decoder example flag, the serve
option in the regen2k YAML, and the measured 2K decode/quality table in
the guide.
CPU tests pin the regen2k VAE load applying the exported decoder, the
serve-config field reaching FastVideoArgs, and the new CLI flag.
BaseLayerWithLoRA.forward was decorated with torch.compile, so wrapping a linear also compiled its base layer.
Compiling a packed NVFP4/FP8 linear on its own changes its numerics, so a wrapped layer at strength 0 could differ
from the unwrapped layer. forward now adds no compiled region: the base layer runs exactly as it would unwrapped
(eagerly, or traced with a caller's compiled block), and only the adapter delta is compiled.
The unit lane fails on main in two contract tests. All three failures are
bookkeeping that drifted from the code, not runtime bugs:
- `test_env_policy.py::test_env_doc_table_matches_registry`: the generated
table in docs/contributing/env_vars.md is stale for
FASTVIDEO_MLX_QUANT_GEMM. Regenerated with the documented command
(`python fastvideo/tests/contract/test_env_policy.py`).
- `test_env_policy.py::test_env_access_follows_policy`: two new violations.
- fastvideo/tests/mlx/test_mlx_affine_dq_gemm.py wrote
FASTVIDEO_MLX_QUANT_GEMM through monkeypatch.setenv; now uses the
registered field's override() through the env_overrides fixture.
- fastvideo/tests/train/models/test_load_kandinsky5.py read
FASTVIDEO_ATTENTION_BACKEND with `not in os.environ`; now uses
envs.FASTVIDEO_ATTENTION_BACKEND.is_set() and override(). Its entry in
KNOWN_VIOLATIONS is deleted, as the shrinking-list rule requires.
- `test_schema_parity_inventory.py::..._extension_fields_are_classified`:
MiniMaxH3PipelineConfig declares video_scheduler_shift and
audio_scheduler_shift (training-time noise clocks rejected by
check_pipeline_config at inference); both are added to
pipeline_config_extensions.internal_only.
Test plan (CPU, no GPU required):
python -m pytest fastvideo/tests/contract/test_env_policy.py \
fastvideo/tests/api/test_schema_parity_inventory.py \
fastvideo/tests/mlx/test_mlx_affine_dq_gemm.py -q
# 45 passed
python -m pytest fastvideo/tests/contract/ fastvideo/tests/api/ -q -m "not gpu"
# 501 passed
Aryan Kumar