1. mode/options conflict: _compile_model_regions now rejects
torch_compile_kwargs['mode'] with an actionable error instead of letting
torch raise about an 'options' key the user never wrote (the CLI help's
own example uses mode). CLI help updated.
2. Legacy stack scoping: new regional_compile flag (FastVideoArgs, default
False). The modular fastvideo/train moduleloader opts in; legacy
--enable-torch-compile recipes keep their pre-PR whole-model
torch.compile semantics (partial graphs, kwargs passed through).
3. AC-prefix normalization in the weight loader: checkpoint_wrapper strips
its prefix from state_dict() keys but not named_parameters()/
named_buffers(); with pre_fsdp_transform applying AC before load, a
loaded buffer inside a wrapped block was silently converted into a
TRAINABLE parameter by load_state_dict(assign=True) (verified on torch
2.12). All name-keyed lookups (buffer membership, fused weight_loader,
_get_parameter_dtype at load and shard time) now compare canonical names
via _strip_checkpoint_wrapper_prefix.
4. FA3: flash-attn 3's grad-enabled attention path graph-breaks, which
fullgraph regional compile turns into a hard error at step 1 on Hopper
hosts. The loader now detects FLASH_ATTN+fa3 and warns-and-disables
compile instead of crashing, and dmd2_t2v_fa3.yaml ships the recipe with
compile explicitly off (main config annotated).
Tests: mode-conflict rejection; AC-prefix canonicalization incl. the
buffer-stays-buffer regression under assign=True.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
