A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 02:00 UTC
[Model][QwenVL] Optimize `Qwen2_5_VisionAttention` q,k preparation (#28769)
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com> Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Lukas Geiger committed
Nov 16, 2025 at 17:37 UTC
5a87076d6ee60a2cf681dada9e971b4ee3e6063e
2 files changed
+25
-27
vllm/model_executor/models/dots_ocr.py
+2
-2
@@ -39,8 +39,8 @@ from vllm.model_executor.models.interfaces import (
39
)
40
from vllm.model_executor.models.module_mapping import MultiModelKeys
41
from vllm.model_executor.models.qwen2 import Qwen2ForCausalLM
42
-from vllm.model_executor.models.qwen2_5_vl import Qwen2_5_VisionAttention
42
from vllm.model_executor.models.qwen2_vl import (
43
+ Qwen2VisionAttention,
44
Qwen2VLDummyInputsBuilder,
45
Qwen2VLMultiModalProcessor,
46
Qwen2VLProcessingInfo,
@@ -328,7 +328,7 @@ class DotsVisionAttention(nn.Module):
328
# [S, C] -> [S, B=1, C]
329
x = hidden_states.unsqueeze(1)
330
x, _ = self.qkv(x)
331
- q, k, v = Qwen2_5_VisionAttention.split_qkv(self, x)
331
+ q, k, v = Qwen2VisionAttention.split_qkv(self, x)
332
bs = q.shape[1]
333
# [S,B,H,D] -> [B,S,H,D]
334
q = q.permute(1, 0, 2, 3).contiguous()
vllm/model_executor/models/qwen2_5_vl.py
+23
-25
@@ -359,23 +359,6 @@ class Qwen2_5_VisionAttention(nn.Module):
359
AttentionBackendEnum.ROCM_AITER_FA,
360
}
361
362
- def split_qkv(self, qkv: torch.Tensor) -> tuple[torch.Tensor, ...]:
363
- # [s, b, 3 * head * head_dim]
364
- seq_len, bs, _ = qkv.shape
365
-
366
- # [s, b, 3 * head * head_dim] -> 3 * [s, b, head * head_dim]
367
- q, k, v = qkv.chunk(3, dim=2)
368
-
369
- # 3 * [s, b, head * head_dim] -> 3 * [s, b, head, head_dim]
370
- new_shape = (
371
- seq_len,
372
- bs,
373
- self.num_attention_heads_per_partition,
374
- self.hidden_size_per_attention_head,
375
- )
376
- q, k, v = (x.view(*new_shape) for x in (q, k, v))
377
- return q, k, v
378
-
362
def forward(
363
self,
364
x: torch.Tensor,
@@ -386,17 +369,32 @@ class Qwen2_5_VisionAttention(nn.Module):
369
) -> torch.Tensor:
370
# [s, b, c] --> [s, b, head * 3 * head_dim]
371
x, _ = self.qkv(x)
372
+ seq_len, batch_size, _ = x.shape
373
390
- # [s, b, 3 * head * head_dim] -> 3 * [s, b, head, head_dim]
391
- q, k, v = self.split_qkv(x)
392
- batch_size = q.shape[1]
374
+ qkv = einops.rearrange(
375
+ x,
376
+ "s b (three head head_dim) -> b s three head head_dim",
377
+ three=3,
378
+ head=self.num_attention_heads_per_partition,
379
+ )
380
394
- q, k, v = (einops.rearrange(x, "s b ... -> b s ...") for x in (q, k, v))
381
if rotary_pos_emb is not None:
396
- # [2 * b, s, heads, head_dim]
397
- qk_concat = torch.cat([q, k], dim=0)
398
- qk_rotated = apply_rotary_pos_emb_vision(qk_concat, rotary_pos_emb)
399
- q, k = torch.chunk(qk_rotated, 2, dim=0)
382
+ qk, v = qkv[:, :, :2], qkv[:, :, 2]
383
+
384
+ qk_reshaped = einops.rearrange(
385
+ qk, "b s two head head_dim -> (two b) s head head_dim", two=2
386
+ )
387
+ qk_rotated = apply_rotary_pos_emb_vision(qk_reshaped, rotary_pos_emb)
388
+ qk_rotated = qk_rotated.view(
389
+ 2,
390
+ batch_size,
391
+ seq_len,
392
+ self.num_attention_heads_per_partition,
393
+ self.hidden_size_per_attention_head,
394
+ )
395
+ q, k = qk_rotated.unbind(dim=0)
396
+ else:
397
+ q, k, v = qkv.unbind(dim=2)
398
399
if self.is_flash_attn_backend:
400
context_layer = vit_flash_attn_wrapper(