fb15fec662
### What this PR does / why we need it? Remove patch for https://github.com/vllm-project/vllm/pull/28798. - vLLM version: v0.12.0 Signed-off-by: shen-shanshan <467638484@qq.com>
86 lines
3.5 KiB
Python
86 lines
3.5 KiB
Python
#
|
|
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
|
|
# This file is a part of the vllm-ascend project.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
#
|
|
|
|
import numpy as np
|
|
import torch
|
|
import torch.nn as nn
|
|
from vllm.model_executor.models.qwen3_vl import Qwen3_VisionTransformer
|
|
|
|
|
|
class AscendQwen3_VisionTransformer(nn.Module):
|
|
|
|
def forward(
|
|
self,
|
|
x: torch.Tensor,
|
|
grid_thw: torch.Tensor | list[list[int]],
|
|
) -> torch.Tensor:
|
|
hidden_states = x.to(device=self.device,
|
|
dtype=self.dtype,
|
|
non_blocking=True)
|
|
hidden_states = self.patch_embed(hidden_states)
|
|
|
|
if isinstance(grid_thw, list):
|
|
grid_thw_list = grid_thw
|
|
grid_thw = np.array(grid_thw, dtype=np.int32)
|
|
else:
|
|
grid_thw = grid_thw.to("cpu")
|
|
grid_thw_list = grid_thw.tolist()
|
|
grid_thw = grid_thw.numpy()
|
|
|
|
pos_embeds = self.fast_pos_embed_interpolate(grid_thw_list)
|
|
hidden_states = hidden_states + pos_embeds
|
|
rotary_pos_emb_cos, rotary_pos_emb_sin = self.rot_pos_emb(
|
|
grid_thw_list)
|
|
rotary_pos_emb_cos = rotary_pos_emb_cos.to(hidden_states.device,
|
|
non_blocking=True)
|
|
rotary_pos_emb_sin = rotary_pos_emb_sin.to(hidden_states.device,
|
|
non_blocking=True)
|
|
|
|
cu_seqlens = np.repeat(grid_thw[:, 1] * grid_thw[:, 2],
|
|
grid_thw[:, 0]).cumsum(axis=0, dtype=np.int32)
|
|
cu_seqlens = np.concatenate([np.zeros(1, dtype=np.int32), cu_seqlens])
|
|
cu_seqlens = torch.from_numpy(cu_seqlens)
|
|
|
|
hidden_states = hidden_states.unsqueeze(1)
|
|
max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens)
|
|
cu_seqlens = cu_seqlens.to(self.device, non_blocking=True)
|
|
|
|
deepstack_feature_lists = []
|
|
for layer_num, blk in enumerate(self.blocks):
|
|
hidden_states = blk(
|
|
hidden_states,
|
|
cu_seqlens=cu_seqlens,
|
|
rotary_pos_emb_cos=rotary_pos_emb_cos,
|
|
rotary_pos_emb_sin=rotary_pos_emb_sin,
|
|
max_seqlen=max_seqlen,
|
|
)
|
|
if layer_num in self.deepstack_visual_indexes:
|
|
deepstack_merger_idx = self.deepstack_visual_indexes.index(
|
|
layer_num)
|
|
deepstack_feature = self.deepstack_merger_list[
|
|
deepstack_merger_idx](hidden_states)
|
|
deepstack_feature_lists.append(deepstack_feature)
|
|
hidden_states = self.merger(hidden_states)
|
|
hidden_states = torch.cat(
|
|
[hidden_states] + deepstack_feature_lists,
|
|
dim=1) # [seq_len, hidden_size * (1 + depth_of_deepstack)]
|
|
return hidden_states
|
|
|
|
|
|
# NOTE: This will be removed after implementing multimodal_cpu_fields in vllm-ascend model_runner.
|
|
Qwen3_VisionTransformer.forward = AscendQwen3_VisionTransformer.forward
|