Interop with the diffusers AnyFlow pipeline (load · convert · docs)#2
Closed
Enderfga wants to merge 6 commits into
Closed
Interop with the diffusers AnyFlow pipeline (load · convert · docs)#2Enderfga wants to merge 6 commits into
Enderfga wants to merge 6 commits into
Conversation
…layout
The nvidia/AnyFlow-*-Diffusers checkpoints reference the diffusers AnyFlow
class names from `model_index.json` and omit `init_far_model`,
`init_flowmap_model`, `chunk_partition`, `qk_norm`, and `added_kv_proj_dim`
from `transformer/config.json`.
- `FAR_Wan_Transformer3DModel`: when `init_*_model` is absent, derive the
flags from `_class_name`. For the FAR variant, additionally fall back to
`chunk_partition=[1, 3, 3, 3, 3, 3, 3, 2]` to match the 81-frame
inference schedule. Configs that already set these fields are passed
through unchanged; user kwargs always win.
- `WanAnyFlowPipeline` / `FARWanAnyFlowPipeline`: pre-instantiate the
transformer and scheduler with the classes defined here and pass them as
kwargs, so `DiffusionPipeline.from_pretrained` skips the module lookup
for those two entries. text_encoder / tokenizer / vae still load normally.
- Bind `diffusers.{AnyFlowTransformer3DModel, AnyFlowFARTransformer3DModel,
FlowMapEulerDiscreteScheduler}` at import time when not already defined;
idempotent.
- Export `FlowMapEulerDiscreteScheduler` as an alias for
`FlowMapDiscreteScheduler`.
Additive only — no class signature or behavior change for existing
checkpoints.
Document the standard diffusers usage for the nvidia/AnyFlow-*-Diffusers checkpoints — AnyFlowPipeline for bidirectional T2V and AnyFlowFARPipeline for the FAR causal variant (T2V / I2V / V2V via the `context_sequence` argument).
…s section
The diffusers AnyFlow pipelines renamed the conditioning kwarg from
``context_sequence={"raw"/"latent"}`` to ``video`` / ``video_latents`` in
huggingface/diffusers#13745 (review feedback from @dg845 — match
``WanVideoToVideoPipeline``'s API surface). Update the README to reflect the
new kwarg and add a short I2V example showing how to pass the single-frame
conditioning tensor.
Only docs change; the in-repo ``WanAnyFlowPipeline`` / ``FARWanAnyFlowPipeline``
keep their original ``context_sequence`` kwarg.
6 tasks
VideoProcessor.preprocess_video's 5D contract is (B, T, C, H, W) — the diffusers AnyFlow PR aligned its docstring + EXAMPLE_DOC_STRING with this in the third review pass (huggingface/diffusers#13745, commits ffdc969 and downstream). This README's I2V example still showed (B, C, T, H, W) and the matching unsqueeze(2); update both so users following the README verbatim get a tensor the diffusers pipeline accepts.
…Flow classes The training pipeline (far/main.py:save_checkpoint) emits .pt files keyed by 'ema' / 'model_state_dict_g'; the diffusers pipelines load from a structured directory written by pipeline.save_pretrained(). Until now this conversion script wrapped the .pt into a pipeline built from this repository's WanAnyFlowPipeline / FARWanAnyFlowPipeline / FAR_Wan_Transformer3DModel / FlowMapDiscreteScheduler — so the resulting model_index.json referenced far.* paths that diffusers.from_pretrained couldn't resolve. Switch the conversion to the diffusers AnyFlow classes (introduced in huggingface/diffusers#13745): - AnyFlowTransformer3DModel (bidirectional T2V variants) - AnyFlowFARTransformer3DModel (FAR causal variants) - AnyFlowPipeline / AnyFlowFARPipeline - FlowMapEulerDiscreteScheduler Output directories now load via AnyFlowPipeline.from_pretrained(...) with no compat shim. The CLI surface (OmegaConf model_type / model_path / model_save_dir keys + auto-append of model_type to the save dir) is preserved. Tensor keys are unchanged across FAR_Wan_Transformer3DModel and the diffusers AnyFlow classes (bit-exact L2=0 against the released NVlabs checkpoints), so load_state_dict(strict=False) handles the EMA bookkeeping fields without dropping any real weights.
Enderfga
changed the title
The scheduler hyperparameters were hardcoded (shift=5.0, num_train_timesteps=1000), matching the released AnyFlow distillation recipe but silently wrong for any future checkpoint trained with a different schedule (e.g. higher-resolution runs that re-tune shift). Make them OmegaConf CLI keys so the conversion stays correct without code edits. Also add a 'source' key so users can convert from the non-EMA generator weights (model_state_dict_g) instead of EMA — useful for ablation runs that want to diff EMA vs raw checkpoints without re-saving.
Contributor
Author
|
Superseded by #5, which takes the inverse approach: instead of shimming this repo's classes to load checkpoints written for the upstream diffusers AnyFlow layout, it realigns the in-repo scheduler / transformer / pipeline API with the upstream contract from huggingface/diffusers#13745. Closing this one. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
With huggingface/diffusers#13745 now merged, the
AnyFlowPipeline/AnyFlowFARPipeline/AnyFlowTransformer3DModel/AnyFlowFARTransformer3DModel/FlowMapEulerDiscreteSchedulerclasses ship indiffusers ≥ 0.36. This PR aligns this repository with that release on three axes — loading, conversion, and docs — so the same.ptweights and the same checkpoint URLs work through either entry point.What changes
1. Loading:
from_pretrainedoverrides + namespace bindingsThe
nvidia/AnyFlow-*-Diffuserscheckpoints currently use the metadata layout written by the pipelines in this repository —model_index.jsonreferencesfar.*modules andtransformer/config.jsoncarriesinit_far_model,init_flowmap_model,chunk_partition,qk_norm,added_kv_proj_dim. After the upcoming metadata flip on the Hub repos, those fields will reference the diffusers class names instead.FAR_Wan_Transformer3DModel.from_pretrained— when the config omitsinit_*_model, derive the flags from_class_nameand fall back tochunk_partition=[1, 3, 3, 3, 3, 3, 3, 2]for the FAR variant (matches the 81-frame inference schedule).WanAnyFlowPipeline.from_pretrained/FARWanAnyFlowPipeline.from_pretrained— pre-instantiate the transformer and scheduler with the classes defined here and pass them as kwargs, soDiffusionPipeline.from_pretrainedskips its class-name lookup for those entries.diffusers.{AnyFlowTransformer3DModel, AnyFlowFARTransformer3DModel, FlowMapEulerDiscreteScheduler}are bound to the local classes only whendiffusersdoesn't already provide those names; idempotent on modern diffusers.2. Conversion:
scripts/convert_model/convert_anyflow_to_diffusers.pyThe script no longer wraps the EMA
.ptintoWanAnyFlowPipeline/FARWanAnyFlowPipelinefrom this repository — it now importsAnyFlowPipeline/AnyFlowFARPipeline/AnyFlowTransformer3DModel/AnyFlowFARTransformer3DModel/FlowMapEulerDiscreteSchedulerdirectly fromdiffusersand runspipeline.save_pretrained(...). The output is the canonical diffusers directory layout —model_index.jsonreferencesAnyFlowPipelineetc., and the directory loads viaAnyFlowPipeline.from_pretrained(...)on a stockpip install diffuserswith no compat shim.Tensor keys are bit-exact between
FAR_Wan_Transformer3DModeland the diffusers AnyFlow classes (verified against the released NVlabs checkpoints on H200,L2 = 0), soload_state_dict(strict=False)carries the EMA bookkeeping fields without dropping any real weights.CLI surface (OmegaConf
model_type/model_path/model_save_dir) is preserved.3. Docs: README "Using with HF diffusers" section
End-to-end examples for the bidirectional and FAR-causal pipelines, including I2V via the
videokwarg ((B, T, C, H, W)in[0, 1], aligned withVideoProcessor.preprocess_video's 5D contract).Compatibility
Additive only:
diffusersships the AnyFlow classes.ImportError(with an upgrade hint) when run againstdiffusers < 0.36.No signature or behavior change for any existing code path in the repository;
demo.pyand the training entry points are untouched.