[AnyFlow] FAR: move chunk_partition default into transformer config

- AnyFlowFARTransformer3DModel.__init__ now accepts chunk_partition via
  @register_to_config (default (1, 3, 3, 3, 3, 3, 3, 2) for the released
  81-frame checkpoints, matching the field on Hub).
- AnyFlowFARPipeline.__call__ no longer requires chunk_partition; defaults
  to self.transformer.config.chunk_partition. Per-call override still
  supported for V2V / non-default num_frames.
- Drop the AnyFlowFARPipeline.default_chunk_partition class attribute.
- Update docs (en pipelines/models, zh using-diffusers) and the conversion
  script to match.

docs/source/en/api/models/anyflow_far_transformer3d.md

@@ -25,8 +25,10 @@ the v0.35.1 Wan2.1 backbone with three additions:
    [`AnyFlowTransformer3DModel`](anyflow_transformer3d)) — every forward call conditions on both the source
    timestep ``t`` and the target timestep ``r``.
 
-The chunk schedule (`chunk_partition`) is **not** baked into the model config. It is a per-call argument to
-`forward`, so the same checkpoint handles different `num_frames` configurations without retraining.
+The default chunk schedule (`chunk_partition`) is stored in the model config; the released NVIDIA AnyFlow-FAR
+checkpoints use `[1, 3, 3, 3, 3, 3, 3, 2]` for the canonical 81-frame setting. `forward` accepts a per-call
+`chunk_partition` override, so the same checkpoint also handles other `num_frames` configurations without
+retraining.
 
 ```python
 from diffusers import AnyFlowFARTransformer3DModel

docs/source/en/api/pipelines/anyflow.md

@@ -70,10 +70,10 @@ clip for V2V continuation. If you already have pre-encoded latents in the model
 ``video_latents=<tensor>`` to skip VAE encoding. ``video`` and ``video_latents`` are mutually exclusive.
 
 > [!IMPORTANT]
-> `AnyFlowFARPipeline.default_chunk_partition = [1, 3, 3, 3, 3, 3, 3, 2]` (sum 21) is matched to the
-> released checkpoints' canonical 81 raw frames (21 latent frames at the VAE temporal stride of 4). When
-> you change `num_frames`, you must also pass a matching `chunk_partition` summing to
-> `(num_frames - 1) // 4 + 1`, otherwise the pipeline raises an `AssertionError`.
+> The released checkpoints bake `chunk_partition=[1, 3, 3, 3, 3, 3, 3, 2]` (sum 21) into the transformer
+> config, matched to the canonical 81 raw frames (21 latent frames at the VAE temporal stride of 4). When
+> you change `num_frames`, pass a matching `chunk_partition` summing to `(num_frames - 1) // 4 + 1`,
+> otherwise the pipeline raises a `ValueError`.
 
 <hfoptions id="anyflow-far">
 <hfoption id="t2v">

docs/source/zh/using-diffusers/anyflow.md

@@ -125,11 +125,11 @@ Causal pipeline 用同一个蒸馏模型支持三种任务模式，**通过 `vid
 Context tensor 的帧数必须满足 `T = 4n + 1`，跟 VAE 时间步长对齐。
 
 > [!IMPORTANT]
-> FAR pipeline 是分块 (chunk) rollout，`num_frames` 必须配合 chunk 调度。默认
-> `chunk_partition=[1, 3, 3, 3, 3, 3, 3, 2]`（求和 21）对应发布 checkpoint 的标准 `num_frames=81`
-> （21 = (81 − 1) // 4 + 1）。改 `num_frames` 时**必须**显式传匹配的 `chunk_partition`，使其求和等于
-> `(num_frames - 1) // 4 + 1`，否则 pipeline 会抛 `AssertionError`。比如 `num_frames=33` 对应 9 个 latent
-> 帧，可用 `chunk_partition=[1, 4, 4]`。
+> FAR pipeline 是分块 (chunk) rollout，`num_frames` 必须配合 chunk 调度。发布的 checkpoint 在
+> transformer config 里写入 `chunk_partition=[1, 3, 3, 3, 3, 3, 3, 2]`（求和 21），对应标准
+> `num_frames=81`（21 = (81 − 1) // 4 + 1）。改 `num_frames` 时**必须**显式传匹配的 `chunk_partition`，
+> 使其求和等于 `(num_frames - 1) // 4 + 1`，否则 pipeline 会抛 `ValueError`。比如 `num_frames=33` 对应
+> 9 个 latent 帧，可用 `chunk_partition=[1, 4, 4]`。
 
 ```py
 import numpy as np

scripts/convert_anyflow_to_diffusers.py

@@ -57,13 +57,21 @@ VARIANTS = {
     "AnyFlow-FAR-Wan2.1-1.3B-Diffusers": {
         "base_model": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
         "transformer_cls": AnyFlowFARTransformer3DModel,
-        "transformer_kwargs": {"full_chunk_limit": 3, "compressed_patch_size": [1, 4, 4]},
+        "transformer_kwargs": {
+            "full_chunk_limit": 3,
+            "compressed_patch_size": [1, 4, 4],
+            "chunk_partition": [1, 3, 3, 3, 3, 3, 3, 2],
+        },
         "pipeline_cls": AnyFlowFARPipeline,
     },
     "AnyFlow-FAR-Wan2.1-14B-Diffusers": {
         "base_model": "Wan-AI/Wan2.1-T2V-14B-Diffusers",
         "transformer_cls": AnyFlowFARTransformer3DModel,
-        "transformer_kwargs": {"full_chunk_limit": 3, "compressed_patch_size": [1, 4, 4]},
+        "transformer_kwargs": {
+            "full_chunk_limit": 3,
+            "compressed_patch_size": [1, 4, 4],
+            "chunk_partition": [1, 3, 3, 3, 3, 3, 3, 2],
+        },
         "pipeline_cls": AnyFlowFARPipeline,
     },
     "AnyFlow-Wan2.1-T2V-1.3B-Diffusers": {

src/diffusers/models/transformers/transformer_anyflow_far.py

@@ -1007,11 +1007,11 @@ class AnyFlowFARTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, Fr
             Mixing gate between source-timestep and delta-timestep embeddings.
         deltatime_type (`str`, defaults to `'r'`):
             Either ``"r"`` (delta is the target timestep) or ``"t-r"`` (delta is the absolute interval).
-
-    .. note::
-        ``chunk_partition`` is **not** a model config field — it is a per-call argument passed to :meth:`forward`.
-        Different inference setups (varying ``num_frames`` or full-vs-compressed schedules) therefore do not require
-        separate checkpoints.
+        chunk_partition (`Tuple[int, ...]`, defaults to `(1, 3, 3, 3, 3, 3, 3, 2)`):
+            Default per-chunk frame counts used by the pipeline. The released NVIDIA AnyFlow-FAR checkpoints
+            target ``num_frames=81`` (21 latent frames at VAE temporal stride 4) split as
+            ``1 + 3*6 + 2``. A different ``num_frames`` requires a matching ``chunk_partition`` override
+            passed to :meth:`AnyFlowFARPipeline.__call__` (and likewise to :meth:`forward`).
     """
 
     _supports_gradient_checkpointing = True
@@ -1040,6 +1040,7 @@ class AnyFlowFARTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, Fr
         rope_max_seq_len: int = 1024,
         gate_value: float = 0.25,
         deltatime_type: str = "r",
+        chunk_partition: Tuple[int, ...] = (1, 3, 3, 3, 3, 3, 3, 2),
     ) -> None:
         super().__init__()
 

src/diffusers/pipelines/anyflow/pipeline_anyflow_far.py

@@ -129,11 +129,6 @@ class AnyFlowFARPipeline(DiffusionPipeline, WanLoraLoaderMixin):
     model_cpu_offload_seq = "text_encoder->transformer->vae"
     _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
 
-    # Default chunk partition for the released NVIDIA AnyFlow-FAR checkpoints (81 frames at the diffusers
-    # VAE temporal stride of 4 → 21 latent frames split into 1 + 3*6 + 2 = [1, 3, 3, 3, 3, 3, 3, 2]). Override
-    # via the ``chunk_partition`` argument to ``__call__`` for other frame counts.
-    default_chunk_partition: List[int] = [1, 3, 3, 3, 3, 3, 3, 2]
-
     def __init__(
         self,
         tokenizer: AutoTokenizer,
@@ -547,9 +542,9 @@ class AnyFlowFARPipeline(DiffusionPipeline, WanLoraLoaderMixin):
             use_kv_cache (`bool`, defaults to `True`):
                 Reuse the FAR attention KV cache across causal chunks. Disable only for debugging.
             chunk_partition (`List[int]`, *optional*):
-                Per-chunk frame counts. Defaults to `default_chunk_partition` (matched to the released 81-frame
-                checkpoints). When you change `num_frames`, supply a `chunk_partition` that sums to `(num_frames - 1)
-                // vae_scale_factor_temporal + 1`.
+                Per-chunk frame counts. Defaults to `self.transformer.config.chunk_partition` (matched to the
+                released 81-frame checkpoints). When you change `num_frames`, supply a `chunk_partition` that
+                sums to `(num_frames - 1) // vae_scale_factor_temporal + 1`.
 
         Examples:
 
@@ -638,7 +633,7 @@ class AnyFlowFARPipeline(DiffusionPipeline, WanLoraLoaderMixin):
             video_latents = self.encode_video(video, height=height, width=width)
 
         if chunk_partition is None:
-            chunk_partition = list(self.default_chunk_partition)
+            chunk_partition = list(self.transformer.config.chunk_partition)
         if init_latents.shape[1] != sum(chunk_partition):
             raise ValueError(
                 f"chunk_partition={chunk_partition} sums to {sum(chunk_partition)}, but the input latent "

tests/pipelines/anyflow/test_anyflow_far.py

@@ -90,6 +90,7 @@ class AnyFlowFARPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             rope_max_seq_len=32,
             gate_value=0.25,
             deltatime_type="r",
+            chunk_partition=(1, 1, 1),
         )
 
         components = {
@@ -106,8 +107,8 @@ class AnyFlowFARPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             generator = torch.manual_seed(seed)
         else:
             generator = torch.Generator(device=device).manual_seed(seed)
-        # num_frames=9 -> 3 latent frames (VAE temporal stride 4); use a matching
-        # chunk_partition so the FAR pipeline's pre-flight assertion passes.
+        # num_frames=9 -> 3 latent frames (VAE temporal stride 4); the transformer config above
+        # has chunk_partition=(1, 1, 1) (sum 3) baked in, so __call__ picks it up automatically.
         inputs = {
             "prompt": "dance monkey",
             "negative_prompt": "negative",
@@ -119,7 +120,6 @@ class AnyFlowFARPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             "num_frames": 9,
             "max_sequence_length": 16,
             "output_type": "pt",
-            "chunk_partition": [1, 1, 1],
         }
         return inputs