onnx-diagnostic 0.8.2__py3-none-any.whl → 0.8.4__py3-none-any.whl

onnx_diagnostic/torch_export_patches/patches/_patch_transformers_qwen2_5.py

@@ -0,0 +1,735 @@
+import os
+from typing import Callable, Optional, Tuple
+import onnx
+import onnx.helper as oh
+import torch
+import torch.nn.functional as F
+from ...export.onnx_plug import EagerDirectReplacementWithOnnx
+from ...helpers.torch_helper import torch_dtype_to_onnx_dtype
+from .patch_helper import _is_torchdynamo_exporting
+from ._patch_transformers_attention import patched_sdpa_attention_forward
+
+try:
+    import transformers.models.qwen2_5_vl
+    import transformers.models.qwen2_5_vl.modeling_qwen2_5_vl
+
+    patch_qwen2_5 = True
+except ImportError:
+    patch_qwen2_5 = False
+
+PLUGS = []
+
+if patch_qwen2_5:
+    import onnxscript
+
+    onnx_plugs_op = onnxscript.values.Opset("onnx_plug", 1)
+    op = onnxscript.opset22
+    op24 = onnxscript.onnx_opset.opset24
+    msft_op = onnxscript.values.Opset("com.microsoft", 1)
+    STOPAT = (
+        int(os.environ.get("STOPAT", None))
+        if os.environ.get("STOPAT", None) is not None
+        else None
+    )
+
+    def _add_com_microsoft_opset(function_proto: onnx.FunctionProto) -> onnx.FunctionProto:
+        opsets = {d.domain: d.version for d in function_proto.opset_import}
+        if "com.microsoft" not in opsets:
+            d = function_proto.opset_import.add()
+            d.domain = "com.microsoft"
+            d.version = 1
+        return function_proto
+
+    def _update_sequence_type(
+        itype: int, function_proto: onnx.FunctionProto
+    ) -> onnx.FunctionProto:
+        proto = oh.make_function(
+            function_proto.domain,
+            function_proto.name,
+            function_proto.input,
+            function_proto.output,
+            [
+                (
+                    oh.make_node("SequenceEmpty", node.input, node.output, dtype=itype)
+                    if node.op_type == "SequenceEmpty"
+                    else node
+                )
+                for node in function_proto.node
+            ],
+            attributes=function_proto.attribute,
+            attribute_protos=function_proto.attribute_proto,
+            opset_imports=function_proto.opset_import,
+        )
+        return proto
+
+    @onnxscript.script(opset=onnx_plugs_op)
+    def LoopMHAAttention(
+        query_states,
+        key_states,
+        value_states,
+        cu_seqlens,
+        scaling: float = 0.11180339887498948,
+        num_heads: int = 16,
+    ):
+        to_3d_shape = op.Constant(value_ints=[0, 0, -1])
+        query_transposed = op.Transpose(query_states, perm=[0, 2, 1, 3])
+        output_shape = op.Shape(query_transposed)
+        query_3d = op.Reshape(query_transposed, to_3d_shape)
+        value_3d = op.Reshape(op.Transpose(value_states, perm=[0, 2, 1, 3]), to_3d_shape)
+        key_3d = op.Reshape(op.Transpose(key_states, perm=[0, 2, 1, 3]), to_3d_shape)
+        cu_seqlens = op.Cast(cu_seqlens, to=onnx.TensorProto.INT32)
+        num_patches = op.Size(cu_seqlens) - 1
+        seq_axis = op.Constant(value_ints=[1])
+        seq_axis_int32 = op.Cast(seq_axis, to=onnx.TensorProto.INT32)
+        # attn_output = op.Slice(value_3d, [0], [0], seq_axis)
+        seq_attn = op.SequenceEmpty(dtype=onnx.TensorProto.FLOAT)
+        for i_patch in range(num_patches):
+            i_1d = op.Reshape(i_patch, [1])
+            i_plus_1_1d = i_1d + 1
+            start = op.Gather(cu_seqlens, i_1d, axis=0)
+            end = op.Gather(cu_seqlens, i_plus_1_1d, axis=0)
+            query_i = op.Slice(query_3d, start, end, seq_axis_int32)
+            key_i = op.Slice(key_3d, start, end, seq_axis_int32)
+            value_i = op.Slice(value_3d, start, end, seq_axis_int32)
+            mha_output = msft_op.MultiHeadAttention(
+                query_i, key_i, value_i, num_heads=num_heads, scale=scaling
+            )
+            # attn_output = op.Concat(attn_output, mha_output, axis=1)
+            seq_attn = op.SequenceInsert(seq_attn, mha_output)
+        attn_output = op.ConcatFromSequence(seq_attn, axis=1)
+        attn_output_4d = op.Reshape(attn_output, output_shape)
+        return attn_output_4d
+
+    @onnxscript.script(opset=onnx_plugs_op)
+    def LoopAttention24(
+        query_states,
+        key_states,
+        value_states,
+        cu_seqlens,
+        scaling: float = 0.11180339887498948,
+        num_heads: int = 16,
+    ):
+        to_3d_shape = op24.Constant(value_ints=[0, 0, -1])
+        query_transposed = op24.Transpose(query_states, perm=[0, 2, 1, 3])
+        output_shape = op24.Shape(query_transposed)
+        query_3d = op24.Reshape(query_transposed, to_3d_shape)
+        value_3d = op24.Reshape(op24.Transpose(value_states, perm=[0, 2, 1, 3]), to_3d_shape)
+        key_3d = op24.Reshape(op24.Transpose(key_states, perm=[0, 2, 1, 3]), to_3d_shape)
+        cu_seqlens = op24.Cast(cu_seqlens, to=onnx.TensorProto.INT32)
+        num_patches = op24.Size(cu_seqlens) - 1
+        seq_axis = op24.Constant(value_ints=[1])
+        seq_axis_int32 = op24.Cast(seq_axis, to=onnx.TensorProto.INT32)
+        seq_attn = op24.SequenceEmpty(dtype=onnx.TensorProto.FLOAT)
+        for i_patch in range(num_patches):
+            i_1d = op24.Reshape(i_patch, [1])
+            i_plus_1_1d = i_1d + 1
+            start = op24.Gather(cu_seqlens, i_1d, axis=0)
+            end = op24.Gather(cu_seqlens, i_plus_1_1d, axis=0)
+            query_i = op24.Slice(query_3d, start, end, seq_axis_int32)
+            key_i = op24.Slice(key_3d, start, end, seq_axis_int32)
+            value_i = op24.Slice(value_3d, start, end, seq_axis_int32)
+            mha_output = op24.Attention(
+                query_i,
+                key_i,
+                value_i,
+                scale=scaling,
+                q_num_heads=num_heads,
+                kv_num_heads=num_heads,
+                softmax_precision=onnx.TensorProto.FLOAT,
+            )
+            seq_attn = op24.SequenceInsert(seq_attn, mha_output)
+        attn_output = op24.ConcatFromSequence(seq_attn, axis=1)
+        attn_output_4d = op24.Reshape(attn_output, output_shape)
+        return attn_output_4d
+
+    @onnxscript.script(opset=onnx_plugs_op)
+    def PackedAttention(
+        query,
+        key,
+        value,
+        cu_seqlens,
+        scaling: float = 0.11180339887498948,
+        num_heads: int = 16,
+    ):
+        num_patches = op.Cast(op.Size(cu_seqlens), to=onnx.TensorProto.INT32) - 1
+        starts = op.Slice(cu_seqlens, [0], [-1], [0])
+        ends = op.Slice(cu_seqlens, [1], [9223372036854775807], [0])
+        lengths = ends - starts
+        max_length = op.ReduceMax(lengths, [0], keepdims=0)  # max_seqlen
+        rows = op.Range(0, num_patches, 1)
+        rows_2d = op.Unsqueeze(rows, [1])
+        cols = op.Range(0, max_length, 1)
+        cols_2d = op.Unsqueeze(cols, [0])
+
+        position_matrix = op.Cast(rows_2d, to=onnx.TensorProto.INT32) * op.Cast(
+            max_length, to=onnx.TensorProto.INT32
+        ) + op.Cast(cols_2d, to=onnx.TensorProto.INT32)
+        position_matrix_shape = op.Shape(position_matrix)
+        token_mask = cols_2d < op.Unsqueeze(lengths, [1])
+        token_mask_1d = op.Reshape(token_mask, [-1])
+        padded_mask_1d = op.Not(token_mask_1d)
+        valid_token_positions = op.Compress(position_matrix, token_mask)
+        padded_token_positions = op.Compress(position_matrix, padded_mask_1d)
+        token_offset_1d = op.Concat(valid_token_positions, padded_token_positions, axis=0)
+        token_offset = op.Reshape(token_offset_1d, position_matrix_shape)
+
+        query_3d = op.Transpose(op.Squeeze(query, [0]), perm=[1, 0, 2])
+        shape_3d = op.Shape(query_3d)
+        query_2d = op.Reshape(query_3d, [0, -1])
+        key_2d = op.Reshape(op.Transpose(op.Squeeze(key, [0]), perm=[1, 0, 2]), [0, -1])
+        value_2d = op.Reshape(op.Transpose(op.Squeeze(value, [0]), perm=[1, 0, 2]), [0, -1])
+
+        packed_attn_output_2d = msft_op.PackedMultiHeadAttention(
+            query_2d,
+            key_2d,
+            value_2d,
+            None,
+            op.Cast(token_offset, to=onnx.TensorProto.INT32),
+            op.Cast(cu_seqlens, to=onnx.TensorProto.INT32),
+            scale=scaling,
+            num_heads=num_heads,
+        )
+        packed_attn_output_3d = op.Reshape(packed_attn_output_2d, shape_3d)
+        return op.Unsqueeze(packed_attn_output_3d, [0])
+
+    def qwen_sdpa_attention(
+        query_states: torch.Tensor,  # F10s1x16xs47x80
+        key_states: torch.Tensor,  # F10s1x16xs47x80
+        value_states: torch.Tensor,  # F10s1x16xs47x80
+        cu_seqlens: torch.Tensor,  # F7su19
+        scaling: float = 0,
+        num_heads: int = 16,
+    ) -> torch.Tensor:
+        """
+        The loop can be removed with the following code
+        but it hits memory overflow for big inputs.
+
+        .. code-block:: python
+
+            # make square mask
+            indices = torch.arange(
+                cu_seqlens.max(), dtype=cu_seqlens.dtype, device=cu_seqlens.device
+            )
+            dot = (cu_seqlens.unsqueeze(1) <= indices.unsqueeze(0)).to(
+                cu_seqlens.dtype
+            )
+            dot = dot.sum(dim=0)
+            mask = dot.unsqueeze(1) - dot.unsqueeze(0)
+            bool_mask = mask == 0
+            bool_mask = bool_mask.unsqueeze(0).unsqueeze(0)
+
+            torch._check(bool_mask.shape[2] == key_states.shape[2])
+            torch._check(bool_mask.shape[3] == key_states.shape[2])
+
+            attn_output, _ = attention_interface(
+                self,
+                query_states,
+                key_states,
+                value_states,
+                attention_mask=bool_mask,
+                scaling=self.scaling,
+                dropout=0.0 if not self.training else self.attention_dropout,
+                is_causal=False,
+                **kwargs,
+            )
+        """
+        lengths = cu_seqlens[1:] - cu_seqlens[:-1]
+        splits = [
+            torch.split(tensor, lengths.tolist(), dim=2)
+            for tensor in (query_states, key_states, value_states)
+        ]
+
+        attn_outputs = [
+            patched_sdpa_attention_forward(
+                None,
+                q,
+                k,
+                v,
+                attention_mask=None,
+                scaling=scaling,
+                dropout=0.0,
+                is_causal=False,
+            )[0]
+            for q, k, v in zip(*splits)
+        ]
+        attn_output = torch.cat(attn_outputs, dim=1)
+        return attn_output
+
+    def qwen_version_selector(opset: int, *args: torch.Tensor) -> Tuple[str, torch.dtype]:
+        first_tensor = next(a for a in args if a is not None)
+        dtype = first_tensor.dtype
+        strategy = patched_Qwen2_5_VLVisionAttention.STRATEGY_FOR_ATTENTION()
+        itype = torch_dtype_to_onnx_dtype(dtype)
+        if strategy is not None:
+            return strategy, itype
+        if dtype == torch.float32 or itype == onnx.TensorProto.FLOAT:
+            if opset >= 24:
+                return "LOOPA24", itype
+            return "LOOPMHA", itype
+        if dtype == torch.float16 or itype == onnx.TensorProto.FLOAT16:
+            # first_tensor may be a SymbolicTensor (onnx).
+            # is_cuda is not available.
+            if hasattr(first_tensor, "is_cuda") and first_tensor.is_cuda:
+                return "PACKED", itype
+            return "LOOPMHA", itype
+        raise AssertionError(
+            f"Unable to handle type {torch.dtype} (itype={itype}) "
+            f"on device {torch.device} with opset={opset}"
+        )
+
+    qwen_sdpa_attention_versatile = EagerDirectReplacementWithOnnx(
+        qwen_sdpa_attention,
+        lambda qs, *args, **kwargs: torch.empty(
+            (qs.shape[0], qs.shape[2], qs.shape[1], qs.shape[3]),
+            dtype=qs.dtype,
+            device=qs.device,
+        ),
+        {
+            ("PACKED", onnx.TensorProto.FLOAT16): _add_com_microsoft_opset(
+                PackedAttention.to_function_proto()
+            ),
+            ("LOOPA24", onnx.TensorProto.FLOAT): LoopAttention24.to_function_proto(),
+            ("LOOPA24", onnx.TensorProto.FLOAT16): _update_sequence_type(
+                onnx.TensorProto.FLOAT16, LoopAttention24.to_function_proto()
+            ),
+            ("LOOPMHA", onnx.TensorProto.FLOAT): _add_com_microsoft_opset(
+                LoopMHAAttention.to_function_proto()
+            ),
+            ("LOOPMHA", onnx.TensorProto.FLOAT16): _update_sequence_type(
+                onnx.TensorProto.FLOAT16,
+                _add_com_microsoft_opset(LoopMHAAttention.to_function_proto()),
+            ),
+        },
+        n_inputs=4,
+        n_outputs=1,
+        kwargs=dict(scaling=0.11180339887498948, num_heads=16),
+        name="qwen_sdpa_attention_versatile",
+        version_selector=qwen_version_selector,
+    )
+    PLUGS.append(qwen_sdpa_attention_versatile)
+
+    class patched_Qwen2_5_VLForConditionalGeneration:
+        _PATCHES_ = ["prepare_inputs_for_generation"]
+        _PATCHED_CLASS_ = (
+            transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLForConditionalGeneration
+        )
+
+        def prepare_inputs_for_generation(
+            self,
+            input_ids,
+            past_key_values=None,
+            attention_mask=None,
+            inputs_embeds=None,
+            cache_position=None,
+            position_ids=None,
+            use_cache=True,
+            pixel_values=None,
+            pixel_values_videos=None,
+            image_grid_thw=None,
+            video_grid_thw=None,
+            second_per_grid_ts=None,
+            **kwargs,
+        ):
+            # Overwritten -- in specific circumstances we don't want to
+            # forward image inputs to the model
+            from transformers.generation import GenerationMixin
+
+            model_inputs = GenerationMixin.prepare_inputs_for_generation(
+                self,
+                input_ids,
+                past_key_values=past_key_values,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                cache_position=cache_position,
+                position_ids=position_ids,
+                pixel_values=pixel_values,
+                pixel_values_videos=pixel_values_videos,
+                image_grid_thw=image_grid_thw,
+                video_grid_thw=video_grid_thw,
+                second_per_grid_ts=second_per_grid_ts,
+                use_cache=use_cache,
+                **kwargs,
+            )
+
+            # Qwen2-5-VL position_ids are prepared with rope_deltas
+            if position_ids is None:
+                # Calculate RoPE index once per generation in the pre-fill stage only.
+                # When compiling, we can't check tensor values thus we check only input length
+                # It is safe to assume that `length!=1` means we're in pre-fill
+                # because compiled models currently cannot do assisted decoding
+                if cache_position[0] == 0 or self.model.rope_deltas is None:
+                    vision_positions, rope_deltas = self.model.get_rope_index(
+                        model_inputs.get("input_ids", None),
+                        image_grid_thw=image_grid_thw,
+                        video_grid_thw=video_grid_thw,
+                        second_per_grid_ts=second_per_grid_ts,
+                        attention_mask=attention_mask,
+                    )
+                    self.model.rope_deltas = rope_deltas
+                # then use the prev pre-calculated rope-deltas to get the correct position ids
+                elif (
+                    "position_ids" in model_inputs and model_inputs["position_ids"] is not None
+                ):
+                    batch_size, seq_length = model_inputs["position_ids"].shape
+                    device = model_inputs["position_ids"].device
+                    position_ids = torch.arange(seq_length, device=device)
+                    position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1)
+                    delta = cache_position[0] + self.model.rope_deltas
+                    delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
+                    vision_positions = position_ids + delta.expand_as(position_ids)
+
+                # Concatenate "text + vision" positions into [4, bs, seq-len]
+                if "position_ids" not in model_inputs or model_inputs["position_ids"] is None:
+                    text_positions = torch.arange(input_ids.shape[1], device=input_ids.device)[
+                        None, None, :
+                    ]
+                else:
+                    text_positions = model_inputs["position_ids"][None, ...]
+                # text_positions = model_inputs["position_ids"][None, ...]
+                assert vision_positions is not None, "vision_positions are missing"
+                model_inputs["position_ids"] = torch.cat(
+                    [text_positions, vision_positions], dim=0
+                )
+
+            if cache_position[0] != 0:
+                model_inputs["pixel_values"] = None
+                model_inputs["pixel_values_videos"] = None
+
+            return model_inputs
+
+    class patched_Qwen2_5_VisionTransformerPretrainedModel:
+        _PATCHES_ = ["get_window_index", "forward", "rot_pos_emb"]
+        _PATCHED_CLASS_ = (
+            transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VisionTransformerPretrainedModel
+        )
+
+        def rot_pos_emb(self, grid_thw):
+            pos_ids = []
+            for thw_ in grid_thw:
+                # PATCHED: avoid unbind
+                t = thw_[0]
+                h = thw_[1]
+                w = thw_[2]
+                hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
+                hpos_ids = hpos_ids.reshape(
+                    h // self.spatial_merge_size,
+                    self.spatial_merge_size,
+                    w // self.spatial_merge_size,
+                    self.spatial_merge_size,
+                )
+                hpos_ids = hpos_ids.permute(0, 2, 1, 3)
+                hpos_ids = hpos_ids.flatten()
+
+                wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
+                wpos_ids = wpos_ids.reshape(
+                    h // self.spatial_merge_size,
+                    self.spatial_merge_size,
+                    w // self.spatial_merge_size,
+                    self.spatial_merge_size,
+                )
+                wpos_ids = wpos_ids.permute(0, 2, 1, 3)
+                wpos_ids = wpos_ids.flatten()
+                pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
+            pos_ids = torch.cat(pos_ids, dim=0)
+            max_grid_size = grid_thw[:, 1:].max()
+            rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+            rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
+            return rotary_pos_emb
+
+        def get_window_index(self, grid_thw):
+            window_index: list = []  # type: ignore[annotation-unchecked]
+            # PATCHED
+            cu_window_seqlens: list = [torch.tensor([0], dtype=torch.int64)]  # type: ignore[annotation-unchecked]
+            window_index_id = 0
+            vit_merger_window_size = (
+                self.window_size // self.spatial_merge_size // self.patch_size
+            )
+
+            for _thw in grid_thw:
+                # PATCHED: avoid unbind
+                grid_t = _thw[0]
+                grid_h = _thw[1]
+                grid_w = _thw[2]
+                llm_grid_h, llm_grid_w = (
+                    grid_h // self.spatial_merge_size,
+                    grid_w // self.spatial_merge_size,
+                )
+                index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(
+                    grid_t, llm_grid_h, llm_grid_w
+                )
+                pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size
+                pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size
+                num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
+                num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
+                index_padded = F.pad(index, (0, pad_w, 0, pad_h), "constant", -100)
+                index_padded = index_padded.reshape(
+                    grid_t,
+                    num_windows_h,
+                    vit_merger_window_size,
+                    num_windows_w,
+                    vit_merger_window_size,
+                )
+                index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape(
+                    grid_t,
+                    num_windows_h * num_windows_w,
+                    vit_merger_window_size,
+                    vit_merger_window_size,
+                )
+                seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
+                index_padded = index_padded.reshape(-1)
+                index_new = index_padded[index_padded != -100]
+                window_index.append(index_new + window_index_id)
+                cu_seqlens_tmp = (
+                    seqlens.cumsum(0) * self.spatial_merge_unit + cu_window_seqlens[-1][-1:]
+                )
+                # PATCHED
+                # cu_window_seqlens.extend(cu_seqlens_tmp.tolist())
+                cu_window_seqlens.append(cu_seqlens_tmp)
+                window_index_id += (grid_t * llm_grid_h * llm_grid_w).item()
+            window_index = torch.cat(window_index, dim=0)
+
+            return window_index, torch.cat(cu_window_seqlens, dim=0)
+
+        def forward(
+            self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs
+        ) -> torch.Tensor:
+            """
+            Args:
+                hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
+                    The final hidden states of the model.
+                grid_thw (`torch.Tensor` of shape `(num_images_or_videos, 3)`):
+                    The temporal, height and width of feature shape of each image in LLM.
+
+            Returns:
+                `torch.Tensor`: hidden_states.
+            """
+            hidden_states = self.patch_embed(hidden_states)
+            rotary_pos_emb = self.rot_pos_emb(grid_thw)
+            window_index, cu_window_seqlens = self.get_window_index(grid_thw)
+            # PATCHED
+            # cu_window_seqlens = torch.tensor(
+            #    cu_window_seqlens,
+            #    device=hidden_states.device,
+            #    dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
+            # )
+            cu_window_seqlens = cu_window_seqlens.to(hidden_states.device).to(grid_thw.dtype)
+            cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)
+
+            seq_len, _ = hidden_states.size()
+            hidden_states = hidden_states.reshape(
+                seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1
+            )
+            hidden_states = hidden_states[window_index, :, :]
+            hidden_states = hidden_states.reshape(seq_len, -1)
+            rotary_pos_emb = rotary_pos_emb.reshape(
+                seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1
+            )
+            rotary_pos_emb = rotary_pos_emb[window_index, :, :]
+            rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)
+            emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
+            position_embeddings = (emb.cos(), emb.sin())
+
+            cu_seqlens = torch.repeat_interleave(
+                grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
+            ).cumsum(
+                dim=0,
+                # Select dtype based on the following factors:
+                #  - FA2 requires that cu_seqlens_q must have dtype int32
+                #  - torch.onnx.export requires that cu_seqlens_q must have same dtype
+                # as grid_thw
+                # See https://github.com/huggingface/transformers/pull/34852
+                # for more information
+                dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
+            )
+            cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+
+            for layer_num, blk in enumerate(self.blocks):
+                if layer_num in self.fullatt_block_indexes:
+                    cu_seqlens_now = cu_seqlens
+                else:
+                    cu_seqlens_now = cu_window_seqlens
+
+                hidden_states = blk(
+                    hidden_states,
+                    cu_seqlens=cu_seqlens_now,
+                    position_embeddings=position_embeddings,
+                    **kwargs,
+                )
+                if STOPAT is not None and layer_num > STOPAT:
+                    break
+
+            hidden_states = self.merger(hidden_states)
+            reverse_indices = torch.argsort(window_index)
+            hidden_states = hidden_states[reverse_indices, :]
+            return hidden_states
+
+    class patched_Qwen2_5_VLVisionAttentionOneIteration(torch.nn.Module):
+        def forward(
+            self,
+            start_end,
+            query_states,
+            key_states,
+            value_states,
+            scaling: float = 1.0,
+            dropout: float = 0.0,
+            **kwargs,
+        ):
+            a = start_end[0].item()
+            b = start_end[1].item()
+            q = query_states[:, :, a:b, :]
+            k = key_states[:, :, a:b, :]
+            v = value_states[:, :, a:b, :]
+            return patched_sdpa_attention_forward(
+                self,
+                q,
+                k,
+                v,
+                attention_mask=None,
+                scaling=scaling,
+                dropout=dropout,
+                is_causal=False,
+                **kwargs,
+            )[0]
+
+    class patched_Qwen2_5_VLVisionAttention:
+        _PATCHES_ = ["forward"]
+        _PATCHED_CLASS_ = (
+            transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLVisionAttention
+        )
+        STRATEGY_FOR_ATTENTION = lambda: os.environ.get("QWEN25ATTENTION", None)  # noqa: E731
+
+        def forward(
+            self,
+            hidden_states: torch.Tensor,
+            cu_seqlens: torch.Tensor,
+            rotary_pos_emb: Optional[torch.Tensor] = None,
+            position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+            **kwargs,
+        ) -> torch.Tensor:
+            seq_length = hidden_states.shape[0]
+            # PATCHED: avoid the use of unbind
+            qkv = (
+                self.qkv(hidden_states)
+                .reshape(seq_length, 3, self.num_heads, -1)
+                .permute(1, 0, 2, 3)
+            )
+
+            query_states, key_states, value_states = qkv[0], qkv[1], qkv[2]
+            cos, sin = position_embeddings
+
+            # This part should be moved into the loop
+            # iteration to enable fusion inside the loop.
+            query_states, key_states = (
+                transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.apply_rotary_pos_emb_vision(
+                    query_states, key_states, cos, sin
+                )
+            )
+
+            query_states = query_states.transpose(0, 1).unsqueeze(0)
+            key_states = key_states.transpose(0, 1).unsqueeze(0)
+            value_states = value_states.transpose(0, 1).unsqueeze(0)
+
+            attention_interface: Callable = (
+                transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.eager_attention_forward
+            )
+            if self.config._attn_implementation != "eager":
+                # PATCHED
+                # attention_interface = ALL_ATTENTION_FUNCTIONS[
+                #       self.config._attn_implementation]
+                attention_interface = transformers.modeling_utils.ALL_ATTENTION_FUNCTIONS[
+                    self.config._attn_implementation
+                ]
+
+            is_sdpa_or_eager = (
+                attention_interface
+                is transformers.integrations.sdpa_attention.sdpa_attention_forward
+                or attention_interface is patched_sdpa_attention_forward
+                or attention_interface
+                is transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.eager_attention_forward
+            )
+            if is_sdpa_or_eager:
+                attn_output = qwen_sdpa_attention_versatile(
+                    query_states,
+                    key_states,
+                    value_states,
+                    cu_seqlens,
+                    self.scaling,
+                    self.num_heads,
+                )
+            elif _is_torchdynamo_exporting():
+                if self.config._attn_implementation == "flash_attention_2":
+                    max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
+                    attn_output = torch.onnx.ops.symbolic(
+                        "custom::qwen25_flash_attention",
+                        (
+                            query_states,
+                            key_states,
+                            value_states,
+                            cu_seqlens,
+                            cu_seqlens,
+                            max_seqlen,
+                            max_seqlen,
+                            torch.tensor(self.scaling, dtype=torch.float32),
+                        ),
+                        dtype=query_states.dtype,
+                        shape=(
+                            query_states.shape[0],  # batch_size
+                            query_states.shape[2],  # sequence_length (total patches)
+                            query_states.shape[1],  # num_heads
+                            query_states.shape[3],  # head_size
+                        ),
+                        version=1,
+                    )
+                else:
+                    raise NotImplementedError(
+                        f"No corresponding export strategy for implementation "
+                        f"{self.config._attn_implementation!r}, "
+                        f"(use QWEN25ATTENTION to change it), and attention_interface="
+                        f"{attention_interface!r} (use sdpa)"
+                    )
+            elif self.config._attn_implementation == "flash_attention_2":
+                # Flash Attention 2: Use cu_seqlens for variable length attention
+                max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
+                attn_output, _ = attention_interface(
+                    self,
+                    query_states,
+                    key_states,
+                    value_states,
+                    attention_mask=None,
+                    scaling=self.scaling,
+                    dropout=0.0 if not self.training else self.attention_dropout,
+                    cu_seq_lens_q=cu_seqlens,
+                    cu_seq_lens_k=cu_seqlens,
+                    max_length_q=max_seqlen,
+                    max_length_k=max_seqlen,
+                    is_causal=False,
+                    **kwargs,
+                )
+            else:
+                # Other implementations: Process each chunk separately
+                # = qwen_sdpa_attention
+                lengths = cu_seqlens[1:] - cu_seqlens[:-1]
+                splits = [
+                    torch.split(tensor, lengths.tolist(), dim=2)
+                    for tensor in (query_states, key_states, value_states)
+                ]
+
+                attn_outputs = [
+                    attention_interface(
+                        self,
+                        q,
+                        k,
+                        v,
+                        attention_mask=None,
+                        scaling=self.scaling,
+                        dropout=0.0 if not self.training else self.attention_dropout,
+                        is_causal=False,
+                        **kwargs,
+                    )[0]
+                    for q, k, v in zip(*splits)
+                ]
+                attn_output = torch.cat(attn_outputs, dim=1)
+
+            attn_output = attn_output.reshape(seq_length, -1).contiguous()
+            attn_output = self.proj(attn_output)
+            return attn_output