update

* [Bug Fix][Qwen-Image-Edit] Fix Qwen-Image-Edit series on NPU

* Enhance NPU attention handling by converting attention mask to boolean and refining mask checks.

* Refine attention mask handling in NPU attention function to improve validation and conversion logic.

* Clean Code

* Refine attention mask processing in NPU attention functions to enhance performance and validation.

* Remove item() ops on npu fa backend.

* Reuse NPU attention mask by `_maybe_modify_attn_mask_npu`

* Apply style fixes

* Update src/diffusers/models/attention_dispatch.py

---------

Co-authored-by: zhangtao <zhangtao529@huawei.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Dhruv Nair <dhruv.nair@gmail.com>

.github/workflows/pr_tests.yml

@@ -114,7 +114,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality]"
+        uv pip install -e ".[quality,test]"
         #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
         uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
@@ -191,7 +191,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality]"
+        uv pip install -e ".[quality,test]"
 
     - name: Environment
       run: |
@@ -242,7 +242,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality]"
+        uv pip install -e ".[quality,test]"
         # TODO (sayakpaul, DN6): revisit `--no-deps`
         uv pip install -U peft@git+https://github.com/huggingface/peft.git --no-deps
         uv pip install -U tokenizers

.github/workflows/push_tests.yml

@@ -77,8 +77,7 @@ jobs:
           uv pip install -e ".[quality]"
           uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
           #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
-          uv pip uninstall transformers huggingface_hub && uv pip install transformers
-
+          uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
       - name: Environment
         run: |
           python utils/print_env.py
@@ -136,7 +135,7 @@ jobs:
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
         #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
-        uv pip uninstall transformers huggingface_hub && uv pip install transformers
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
 
     - name: Environment
       run: |
@@ -189,7 +188,7 @@ jobs:
       run: |
         uv pip install -e ".[quality,training]"
         #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
-        uv pip uninstall transformers huggingface_hub && uv pip install transformers
+        uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
     - name: Environment
       run: |
         python utils/print_env.py

setup.py

@@ -101,6 +101,7 @@ _deps = [
     "datasets",
     "filelock",
     "flax>=0.4.1",
+    "ftfy",
     "hf-doc-builder>=0.3.0",
     "httpx<1.0.0",
     "huggingface-hub>=0.34.0,<2.0",
@@ -221,12 +222,14 @@ extras["docs"] = deps_list("hf-doc-builder")
 extras["training"] = deps_list("accelerate", "datasets", "protobuf", "tensorboard", "Jinja2", "peft", "timm")
 extras["test"] = deps_list(
     "compel",
+    "ftfy",
     "GitPython",
     "datasets",
     "Jinja2",
     "invisible-watermark",
     "librosa",
     "parameterized",
+    "protobuf",
     "pytest",
     "pytest-timeout",
     "pytest-xdist",
@@ -235,6 +238,7 @@ extras["test"] = deps_list(
     "sentencepiece",
     "scipy",
     "tiktoken",
+    "torchsde",
     "torchvision",
     "transformers",
     "phonemizer",

src/diffusers/dependency_versions_table.py

@@ -8,6 +8,7 @@ deps = {
     "datasets": "datasets",
     "filelock": "filelock",
     "flax": "flax>=0.4.1",
+    "ftfy": "ftfy",
     "hf-doc-builder": "hf-doc-builder>=0.3.0",
     "httpx": "httpx<1.0.0",
     "huggingface-hub": "huggingface-hub>=0.34.0,<2.0",

src/diffusers/models/attention_dispatch.py

@@ -1117,6 +1117,26 @@ def _sage_attention_backward_op(
     raise NotImplementedError("Backward pass is not implemented for Sage attention.")
 
 
+def _maybe_modify_attn_mask_npu(query: torch.Tensor, key: torch.Tensor, attn_mask: torch.Tensor | None = None):
+    # Skip Attention Mask if all values are 1, `None` mask can speedup the computation
+    if attn_mask is not None and torch.all(attn_mask != 0):
+        attn_mask = None
+
+    # Reshape Attention Mask: [batch_size, seq_len_k] -> [batch_size, 1, sqe_len_q, seq_len_k]
+    # https://www.hiascend.com/document/detail/zh/Pytorch/730/apiref/torchnpuCustomsapi/docs/context/torch_npu-npu_fusion_attention.md
+    if (
+        attn_mask is not None
+        and attn_mask.ndim == 2
+        and attn_mask.shape[0] == query.shape[0]
+        and attn_mask.shape[1] == key.shape[1]
+    ):
+        B, Sq, Skv = attn_mask.shape[0], query.shape[1], key.shape[1]
+        attn_mask = ~attn_mask.to(torch.bool)
+        attn_mask = attn_mask.unsqueeze(1).expand(B, Sq, Skv).unsqueeze(1).contiguous()
+
+    return attn_mask
+
+
 def _npu_attention_forward_op(
     ctx: torch.autograd.function.FunctionCtx,
     query: torch.Tensor,
@@ -1134,11 +1154,14 @@ def _npu_attention_forward_op(
     if return_lse:
         raise ValueError("NPU attention backend does not support setting `return_lse=True`.")
 
+    attn_mask = _maybe_modify_attn_mask_npu(query, key, attn_mask)
+
     out = npu_fusion_attention(
         query,
         key,
         value,
         query.size(2),  # num_heads
+        atten_mask=attn_mask,
         input_layout="BSND",
         pse=None,
         scale=1.0 / math.sqrt(query.shape[-1]) if scale is None else scale,
@@ -2668,16 +2691,17 @@ def _native_npu_attention(
     return_lse: bool = False,
     _parallel_config: "ParallelConfig" | None = None,
 ) -> torch.Tensor:
-    if attn_mask is not None:
-        raise ValueError("`attn_mask` is not supported for NPU attention")
     if return_lse:
         raise ValueError("NPU attention backend does not support setting `return_lse=True`.")
     if _parallel_config is None:
+        attn_mask = _maybe_modify_attn_mask_npu(query, key, attn_mask)
+
         out = npu_fusion_attention(
             query,
             key,
             value,
             query.size(2),  # num_heads
+            atten_mask=attn_mask,
             input_layout="BSND",
             pse=None,
             scale=1.0 / math.sqrt(query.shape[-1]) if scale is None else scale,
@@ -2692,7 +2716,7 @@ def _native_npu_attention(
             query,
             key,
             value,
-            None,
+            attn_mask,
             dropout_p,
             None,
             scale,

src/diffusers/models/transformers/transformer_qwenimage.py

@@ -164,7 +164,11 @@ def compute_text_seq_len_from_mask(
     position_ids = torch.arange(text_seq_len, device=encoder_hidden_states.device, dtype=torch.long)
     active_positions = torch.where(encoder_hidden_states_mask, position_ids, position_ids.new_zeros(()))
     has_active = encoder_hidden_states_mask.any(dim=1)
-    per_sample_len = torch.where(has_active, active_positions.max(dim=1).values + 1, torch.as_tensor(text_seq_len))
+    per_sample_len = torch.where(
+        has_active,
+        active_positions.max(dim=1).values + 1,
+        torch.as_tensor(text_seq_len, device=encoder_hidden_states.device),
+    )
     return text_seq_len, per_sample_len, encoder_hidden_states_mask