Merge branch 'main' into refactor-hub-attn-kernels

Author: sayakpaul

docs/source/en/_toctree.yml

@@ -529,8 +529,6 @@
         title: Kandinsky 2.2
       - local: api/pipelines/kandinsky3
         title: Kandinsky 3
-      - local: api/pipelines/kandinsky5
-        title: Kandinsky 5
       - local: api/pipelines/kolors
         title: Kolors
       - local: api/pipelines/latent_consistency_models
@@ -656,6 +654,8 @@
         title: Text2Video-Zero
       - local: api/pipelines/wan
         title: Wan
+      - local: api/pipelines/kandinsky5_video
+        title: Kandinsky 5.0 Video
       title: Video
     title: Pipelines
   - sections:

…cs/source/en/api/pipelines/kandinsky5.md …rce/en/api/pipelines/kandinsky5_video.mddocs/source/en/api/pipelines/kandinsky5.md renamed to docs/source/en/api/pipelines/kandinsky5_video.md docs/source/en/api/pipelines/kandinsky5.md renamed to docs/source/en/api/pipelines/kandinsky5_video.md

@@ -7,9 +7,9 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# Kandinsky 5.0
+# Kandinsky 5.0 Video
 
-Kandinsky 5.0 is created by the Kandinsky team: Alexey Letunovskiy, Maria Kovaleva, Ivan Kirillov, Lev Novitskiy, Denis Koposov, Dmitrii Mikhailov, Anna Averchenkova, Andrey Shutkin, Julia Agafonova, Olga Kim, Anastasiia Kargapoltseva, Nikita Kiselev, Anna Dmitrienko,  Anastasia Maltseva, Kirill Chernyshev, Ilia Vasiliev, Viacheslav Vasilev, Vladimir Polovnikov, Yury Kolabushin, Alexander Belykh, Mikhail Mamaev, Anastasia Aliaskina, Tatiana Nikulina, Polina Gavrilova, Vladimir Arkhipkin, Vladimir Korviakov, Nikolai Gerasimenko, Denis Parkhomenko, Denis Dimitrov
+Kandinsky 5.0 Video is created by the Kandinsky team: Alexey Letunovskiy, Maria Kovaleva, Ivan Kirillov, Lev Novitskiy, Denis Koposov, Dmitrii Mikhailov, Anna Averchenkova, Andrey Shutkin, Julia Agafonova, Olga Kim, Anastasiia Kargapoltseva, Nikita Kiselev, Anna Dmitrienko,  Anastasia Maltseva, Kirill Chernyshev, Ilia Vasiliev, Viacheslav Vasilev, Vladimir Polovnikov, Yury Kolabushin, Alexander Belykh, Mikhail Mamaev, Anastasia Aliaskina, Tatiana Nikulina, Polina Gavrilova, Vladimir Arkhipkin, Vladimir Korviakov, Nikolai Gerasimenko, Denis Parkhomenko, Denis Dimitrov
 
 
 Kandinsky 5.0 is a family of diffusion models for Video & Image generation. Kandinsky 5.0 T2V Lite is a lightweight video generation model (2B parameters) that ranks #1 among open-source models in its class. It outperforms larger models and offers the best understanding of Russian concepts in the open-source ecosystem.
@@ -92,7 +92,7 @@ pipe = pipe.to("cuda")
 
 pipe.transformer.set_attention_backend(
     "flex"
-)                                       # <--- Set attention backend to Flex
+)                                       # <--- Sett attention bakend to Flex
 pipe.transformer.compile(
     mode="max-autotune-no-cudagraphs", 
     dynamic=True
@@ -115,7 +115,7 @@ export_to_video(output, "output.mp4", fps=24, quality=9)
 ```
 
 ### Diffusion Distilled model
-**⚠️ Warning!** all nocfg and diffusion distilled models should be inferred without CFG (```guidance_scale=1.0```):
+**⚠️ Warning!** all nocfg and diffusion distilled models should be infered wothout CFG (```guidance_scale=1.0```):
 
 ```python
 model_id = "ai-forever/Kandinsky-5.0-T2V-Lite-distilled16steps-5s-Diffusers"

src/diffusers/models/attention_dispatch.py

@@ -675,6 +675,86 @@ def _(
 # ===== Helper functions to use attention backends with templated CP autograd functions =====
 
 
+def _native_attention_forward_op(
+    ctx: torch.autograd.function.FunctionCtx,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attn_mask: Optional[torch.Tensor] = None,
+    dropout_p: float = 0.0,
+    is_causal: bool = False,
+    scale: Optional[float] = None,
+    enable_gqa: bool = False,
+    return_lse: bool = False,
+    _save_ctx: bool = True,
+    _parallel_config: Optional["ParallelConfig"] = None,
+):
+    # Native attention does not return_lse
+    if return_lse:
+        raise ValueError("Native attention does not support return_lse=True")
+
+    # used for backward pass
+    if _save_ctx:
+        ctx.save_for_backward(query, key, value)
+        ctx.attn_mask = attn_mask
+        ctx.dropout_p = dropout_p
+        ctx.is_causal = is_causal
+        ctx.scale = scale
+        ctx.enable_gqa = enable_gqa
+
+    query, key, value = (x.permute(0, 2, 1, 3) for x in (query, key, value))
+    out = torch.nn.functional.scaled_dot_product_attention(
+        query=query,
+        key=key,
+        value=value,
+        attn_mask=attn_mask,
+        dropout_p=dropout_p,
+        is_causal=is_causal,
+        scale=scale,
+        enable_gqa=enable_gqa,
+    )
+    out = out.permute(0, 2, 1, 3)
+
+    return out
+
+
+def _native_attention_backward_op(
+    ctx: torch.autograd.function.FunctionCtx,
+    grad_out: torch.Tensor,
+    *args,
+    **kwargs,
+):
+    query, key, value = ctx.saved_tensors
+
+    query.requires_grad_(True)
+    key.requires_grad_(True)
+    value.requires_grad_(True)
+
+    query_t, key_t, value_t = (x.permute(0, 2, 1, 3) for x in (query, key, value))
+    out = torch.nn.functional.scaled_dot_product_attention(
+        query=query_t,
+        key=key_t,
+        value=value_t,
+        attn_mask=ctx.attn_mask,
+        dropout_p=ctx.dropout_p,
+        is_causal=ctx.is_causal,
+        scale=ctx.scale,
+        enable_gqa=ctx.enable_gqa,
+    )
+    out = out.permute(0, 2, 1, 3)
+
+    grad_out_t = grad_out.permute(0, 2, 1, 3)
+    grad_query_t, grad_key_t, grad_value_t = torch.autograd.grad(
+        outputs=out, inputs=[query_t, key_t, value_t], grad_outputs=grad_out_t, retain_graph=False
+    )
+
+    grad_query = grad_query_t.permute(0, 2, 1, 3)
+    grad_key = grad_key_t.permute(0, 2, 1, 3)
+    grad_value = grad_value_t.permute(0, 2, 1, 3)
+
+    return grad_query, grad_key, grad_value
+
+
 # https://github.com/pytorch/pytorch/blob/8904ba638726f8c9a5aff5977c4aa76c9d2edfa6/aten/src/ATen/native/native_functions.yaml#L14958
 # forward declaration:
 #   aten::_scaled_dot_product_cudnn_attention(Tensor query, Tensor key, Tensor value, Tensor? attn_bias, bool compute_log_sumexp, float dropout_p=0., bool is_causal=False, bool return_debug_mask=False, *, float? scale=None) -> (Tensor output, Tensor logsumexp, Tensor cum_seq_q, Tensor cum_seq_k, SymInt max_q, SymInt max_k, Tensor philox_seed, Tensor philox_offset, Tensor debug_attn_mask)
@@ -1550,6 +1630,7 @@ def score_mod(score, batch_idx, head_idx, q_idx, kv_idx):
 @_AttentionBackendRegistry.register(
     AttentionBackendName.NATIVE,
     constraints=[_check_device, _check_shape],
+    supports_context_parallel=True,
 )
 def _native_attention(
     query: torch.Tensor,
@@ -1565,18 +1646,35 @@ def _native_attention(
 ) -> torch.Tensor:
     if return_lse:
         raise ValueError("Native attention backend does not support setting `return_lse=True`.")
-    query, key, value = (x.permute(0, 2, 1, 3) for x in (query, key, value))
-    out = torch.nn.functional.scaled_dot_product_attention(
-        query=query,
-        key=key,
-        value=value,
-        attn_mask=attn_mask,
-        dropout_p=dropout_p,
-        is_causal=is_causal,
-        scale=scale,
-        enable_gqa=enable_gqa,
-    )
-    out = out.permute(0, 2, 1, 3)
+    if _parallel_config is None:
+        query, key, value = (x.permute(0, 2, 1, 3) for x in (query, key, value))
+        out = torch.nn.functional.scaled_dot_product_attention(
+            query=query,
+            key=key,
+            value=value,
+            attn_mask=attn_mask,
+            dropout_p=dropout_p,
+            is_causal=is_causal,
+            scale=scale,
+            enable_gqa=enable_gqa,
+        )
+        out = out.permute(0, 2, 1, 3)
+    else:
+        out = _templated_context_parallel_attention(
+            query,
+            key,
+            value,
+            attn_mask,
+            dropout_p,
+            is_causal,
+            scale,
+            enable_gqa,
+            return_lse,
+            forward_op=_native_attention_forward_op,
+            backward_op=_native_attention_backward_op,
+            _parallel_config=_parallel_config,
+        )
+
     return out
 
 

src/diffusers/modular_pipelines/modular_pipeline.py

@@ -313,7 +313,7 @@ def from_pretrained(
         ]
         hub_kwargs = {name: kwargs.pop(name) for name in hub_kwargs_names if name in kwargs}
 
-        config = cls.load_config(pretrained_model_name_or_path)
+        config = cls.load_config(pretrained_model_name_or_path, **hub_kwargs)
         has_remote_code = "auto_map" in config and cls.__name__ in config["auto_map"]
         trust_remote_code = resolve_trust_remote_code(
             trust_remote_code, pretrained_model_name_or_path, has_remote_code