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][Perf] Use cos and sin cache in QwenVL (#28798)
Signed-off-by: gcanlin <canlinguosdu@gmail.com>
Canlin Guo committed
Nov 18, 2025 at 19:51 UTC
b9489f51e1c61c96378e12c9523f9de7043ca294
6 files changed
+218
-217
vllm/model_executor/layers/rotary_embedding/base.py
+5
@@ -83,6 +83,11 @@ class RotaryEmbeddingBase(CustomOp):
83
):
84
self.cos_sin_cache = self.cos_sin_cache.to(query.device, dtype=query.dtype)
85
86
+ def get_cos_sin(self, seqlen: int) -> tuple[torch.Tensor, torch.Tensor]:
87
+ cos_sin = self.cos_sin_cache[:seqlen]
88
+ cos, sin = cos_sin.chunk(2, dim=-1)
89
+ return cos, sin
90
+
91
92
class RotaryEmbedding(RotaryEmbeddingBase):
93
def __init__(
vllm/model_executor/models/glm4_1v.py
+38
-50
@@ -65,6 +65,7 @@ from vllm.model_executor.layers.linear import (
65
RowParallelLinear,
66
)
67
from vllm.model_executor.layers.quantization import QuantizationConfig
68
+from vllm.model_executor.layers.rotary_embedding import get_rope
69
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
70
from vllm.model_executor.models.module_mapping import MultiModelKeys
71
from vllm.multimodal import MULTIMODAL_REGISTRY
@@ -341,7 +342,8 @@ class Glm4vVisionAttention(nn.Module):
342
self,
343
x: torch.Tensor,
344
cu_seqlens: torch.Tensor,
344
- rotary_pos_emb: torch.Tensor,
345
+ rotary_pos_emb_cos: torch.Tensor,
346
+ rotary_pos_emb_sin: torch.Tensor,
347
max_seqlen: int | None = None, # Only used for Flash Attention
348
seqlens: list[int] | None = None, # Only used for xFormers
349
) -> torch.Tensor:
@@ -353,10 +355,12 @@ class Glm4vVisionAttention(nn.Module):
355
batch_size = q.shape[1]
356
357
q, k, v = (rearrange(x, "s b ... -> b s ...").contiguous() for x in (q, k, v))
356
- if rotary_pos_emb is not None:
358
+ if rotary_pos_emb_cos is not None and rotary_pos_emb_sin is not None:
359
# [2 * b, s, heads, head_dim]
360
qk_concat = torch.cat([q, k], dim=0)
359
- qk_rotated = apply_rotary_pos_emb_vision(qk_concat, rotary_pos_emb)
361
+ qk_rotated = apply_rotary_pos_emb_vision(
362
+ qk_concat, rotary_pos_emb_cos, rotary_pos_emb_sin
363
+ )
364
q, k = torch.chunk(qk_rotated, 2, dim=0)
365
366
if self.is_flash_attn_backend:
@@ -454,14 +458,16 @@ class Glm4vVisionBlock(nn.Module):
458
self,
459
x: torch.Tensor,
460
cu_seqlens: torch.Tensor,
457
- rotary_pos_emb: torch.Tensor,
461
+ rotary_pos_emb_cos: torch.Tensor,
462
+ rotary_pos_emb_sin: torch.Tensor,
463
max_seqlen: int | None = None, # Only used for Flash Attention
464
seqlens: list[int] | None = None, # Only used for xFormers
465
) -> torch.Tensor:
466
x_attn = self.attn(
467
self.norm1(x),
468
cu_seqlens=cu_seqlens,
464
- rotary_pos_emb=rotary_pos_emb,
469
+ rotary_pos_emb_cos=rotary_pos_emb_cos,
470
+ rotary_pos_emb_sin=rotary_pos_emb_sin,
471
max_seqlen=max_seqlen,
472
seqlens=seqlens,
473
)
@@ -660,44 +666,6 @@ class Glm4vVisionEmbeddings(nn.Module):
666
return embeddings
667
668
663
-class Glm4vVisionRotaryEmbedding(nn.Module):
664
- def __init__(self, dim: int, theta: float = 10000.0) -> None:
665
- super().__init__()
666
- self.dim = dim
667
- self.theta = theta
668
- inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
669
- self.register_buffer("inv_freq", inv_freq, persistent=False)
670
- self._seq_len_cached = 0
671
- self._freqs_cached = None
672
-
673
- def update_freqs_cache(self, seqlen: int) -> None:
674
- if seqlen > self._seq_len_cached:
675
- seqlen *= 2
676
- self._seq_len_cached = seqlen
677
- self.inv_freq = 1.0 / (
678
- self.theta
679
- ** (
680
- torch.arange(
681
- 0,
682
- self.dim,
683
- 2,
684
- dtype=torch.float,
685
- device=self.inv_freq.device,
686
- )
687
- / self.dim
688
- )
689
- )
690
- seq = torch.arange(
691
- seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype
692
- )
693
- freqs = torch.outer(seq, self.inv_freq)
694
- self._freqs_cached = freqs
695
-
696
- def forward(self, seqlen: int) -> torch.Tensor:
697
- self.update_freqs_cache(seqlen)
698
- return self._freqs_cached[:seqlen]
699
-
700
-
669
class Glm4vVisionTransformer(nn.Module):
670
def __init__(
671
self,
@@ -731,7 +699,13 @@ class Glm4vVisionTransformer(nn.Module):
699
700
norm_layer = partial(RMSNorm, eps=norm_eps)
701
head_dim = self.hidden_size // self.num_heads
734
- self.rotary_pos_emb = Glm4vVisionRotaryEmbedding(head_dim // 2)
702
+ self.rotary_pos_emb = get_rope(
703
+ head_size=head_dim,
704
+ rotary_dim=head_dim // 2,
705
+ max_position=8192,
706
+ base=10000.0,
707
+ is_neox_style=True,
708
+ )
709
self.blocks = nn.ModuleList(
710
[
711
Glm4vVisionBlock(
@@ -789,7 +763,9 @@ class Glm4vVisionTransformer(nn.Module):
763
def device(self) -> torch.device:
764
return self.patch_embed.proj.weight.device
765
792
- def rot_pos_emb(self, grid_thw: torch.Tensor) -> torch.Tensor:
766
+ def rot_pos_emb(
767
+ self, grid_thw: torch.Tensor
768
+ ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
769
pos_ids = []
770
for t, h, w in grid_thw:
771
hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
@@ -817,9 +793,18 @@ class Glm4vVisionTransformer(nn.Module):
793
pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
794
pos_ids = torch.cat(pos_ids, dim=0)
795
max_grid_size = grid_thw[:, 1:].max()
820
- rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
821
- rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
822
- return rotary_pos_emb, pos_ids
796
+
797
+ # Use pre-computed cos_sin_cache from RotaryEmbedding
798
+ cos, sin = self.rotary_pos_emb.get_cos_sin(max_grid_size)
799
+
800
+ cos_h = cos[pos_ids[:, 0]] # (num_tokens, rotary_dim // 2)
801
+ cos_w = cos[pos_ids[:, 1]]
802
+ sin_h = sin[pos_ids[:, 0]]
803
+ sin_w = sin[pos_ids[:, 1]]
804
+
805
+ cos_combined = torch.cat([cos_h, cos_w], dim=-1)
806
+ sin_combined = torch.cat([sin_h, sin_w], dim=-1)
807
+ return cos_combined, sin_combined, pos_ids
808
809
def compute_attn_mask_seqlen(
810
self,
@@ -848,7 +833,9 @@ class Glm4vVisionTransformer(nn.Module):
833
x = self.post_conv_layernorm(x)
834
835
# compute position embedding
851
- rotary_pos_emb, image_type_ids = self.rot_pos_emb(grid_thw)
836
+ rotary_pos_emb_cos, rotary_pos_emb_sin, image_type_ids = self.rot_pos_emb(
837
+ grid_thw
838
+ )
839
# compute cu_seqlens
840
cu_seqlens = torch.repeat_interleave(
841
grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
@@ -867,7 +854,8 @@ class Glm4vVisionTransformer(nn.Module):
854
x = blk(
855
x,
856
cu_seqlens=cu_seqlens,
870
- rotary_pos_emb=rotary_pos_emb,
857
+ rotary_pos_emb_cos=rotary_pos_emb_cos,
858
+ rotary_pos_emb_sin=rotary_pos_emb_sin,
859
max_seqlen=max_seqlen,
860
seqlens=seqlens,
861
)
vllm/model_executor/models/qwen2_5_vl.py
+65
-58
@@ -64,6 +64,7 @@ from vllm.model_executor.layers.linear import (
64
RowParallelLinear,
65
)
66
from vllm.model_executor.layers.quantization import QuantizationConfig
67
+from vllm.model_executor.layers.rotary_embedding import get_rope
68
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
69
from vllm.model_executor.models.module_mapping import MultiModelKeys
70
from vllm.model_executor.models.vision import should_torch_compile_mm_vit
@@ -363,7 +364,8 @@ class Qwen2_5_VisionAttention(nn.Module):
364
self,
365
x: torch.Tensor,
366
cu_seqlens: torch.Tensor,
366
- rotary_pos_emb: torch.Tensor,
367
+ rotary_pos_emb_cos: torch.Tensor,
368
+ rotary_pos_emb_sin: torch.Tensor,
369
max_seqlen: torch.Tensor, # Only used for Flash Attention
370
seqlens: torch.Tensor, # Only used for xFormers
371
) -> torch.Tensor:
@@ -378,13 +380,15 @@ class Qwen2_5_VisionAttention(nn.Module):
380
head=self.num_attention_heads_per_partition,
381
)
382
381
- if rotary_pos_emb is not None:
383
+ if rotary_pos_emb_cos is not None and rotary_pos_emb_sin is not None:
384
qk, v = qkv[:, :, :2], qkv[:, :, 2]
385
386
qk_reshaped = einops.rearrange(
387
qk, "b s two head head_dim -> (two b) s head head_dim", two=2
388
)
387
- qk_rotated = apply_rotary_pos_emb_vision(qk_reshaped, rotary_pos_emb)
389
+ qk_rotated = apply_rotary_pos_emb_vision(
390
+ qk_reshaped, cos=rotary_pos_emb_cos, sin=rotary_pos_emb_sin
391
+ )
392
qk_rotated = qk_rotated.view(
393
2,
394
batch_size,
@@ -434,7 +438,8 @@ class Qwen2_5_VisionAttention(nn.Module):
438
dynamic_arg_dims={
439
"x": 0,
440
"cu_seqlens": 0,
437
- "rotary_pos_emb": 0,
441
+ "rotary_pos_emb_cos": 0,
442
+ "rotary_pos_emb_sin": 0,
443
"seqlens": 0,
444
},
445
mark_unbacked_dims={"seqlens": 0},
@@ -485,14 +490,16 @@ class Qwen2_5_VisionBlock(nn.Module):
490
self,
491
x: torch.Tensor,
492
cu_seqlens: torch.Tensor,
488
- rotary_pos_emb: torch.Tensor,
493
+ rotary_pos_emb_cos: torch.Tensor,
494
+ rotary_pos_emb_sin: torch.Tensor,
495
max_seqlen: torch.Tensor, # Only used for Flash Attention
496
seqlens: torch.Tensor, # Only used for xFormers
497
) -> torch.Tensor:
498
x_attn = self.attn(
499
self.norm1(x),
500
cu_seqlens=cu_seqlens,
495
- rotary_pos_emb=rotary_pos_emb,
501
+ rotary_pos_emb_cos=rotary_pos_emb_cos,
502
+ rotary_pos_emb_sin=rotary_pos_emb_sin,
503
max_seqlen=max_seqlen,
504
seqlens=seqlens,
505
)
@@ -588,42 +595,6 @@ class Qwen2_5_VisionPatchMerger(nn.Module):
595
return out
596
597
591
-class Qwen2_5_VisionRotaryEmbedding(nn.Module):
592
- def __init__(self, dim: int, theta: float = 10000.0) -> None:
593
- super().__init__()
594
- self.dim = dim
595
- self.theta = theta
596
- inv_freq = 1.0 / (
597
- theta ** (torch.arange(0, dim, 2, dtype=torch.float, device="cpu") / dim)
598
- )
599
- self.register_buffer("inv_freq", inv_freq, persistent=False)
600
- self._seq_len_cached = 0
601
- self._freqs_cached = None
602
-
603
- def update_freqs_cache(self, seqlen: int) -> None:
604
- if seqlen > self._seq_len_cached:
605
- seqlen *= 2
606
- self._seq_len_cached = seqlen
607
- self.inv_freq = 1.0 / (
608
- self.theta
609
- ** (
610
- torch.arange(
611
- 0, self.dim, 2, dtype=torch.float, device=self.inv_freq.device
612
- )
613
- / self.dim
614
- )
615
- )
616
- seq = torch.arange(
617
- seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype
618
- )
619
- freqs = torch.outer(seq, self.inv_freq)
620
- self._freqs_cached = freqs
621
-
622
- def forward(self, seqlen: int) -> torch.Tensor:
623
- self.update_freqs_cache(seqlen)
624
- return self._freqs_cached[:seqlen]
625
-
626
-
598
class Qwen2_5_VisionTransformer(nn.Module):
599
def __init__(
600
self,
@@ -666,7 +637,13 @@ class Qwen2_5_VisionTransformer(nn.Module):
637
638
norm_layer = partial(RMSNorm, eps=norm_eps)
639
head_dim = self.hidden_size // self.num_heads
669
- self.rotary_pos_emb = Qwen2_5_VisionRotaryEmbedding(head_dim // 2)
640
+ self.rotary_pos_emb = get_rope(
641
+ head_size=head_dim,
642
+ rotary_dim=head_dim // 2,
643
+ max_position=8192,
644
+ base=10000.0,
645
+ is_neox_style=True,
646
+ )
647
648
use_upstream_fa = False
649
self.attn_backend = get_vit_attn_backend(
@@ -757,15 +734,30 @@ class Qwen2_5_VisionTransformer(nn.Module):
734
)
735
pos_ids = torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1)
736
max_size = max(h, w)
760
- rotary_pos_emb_full = self.rotary_pos_emb(max_size)
761
- rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
762
- rotary_pos_emb = rotary_pos_emb.reshape(
763
- rotary_pos_emb.shape[0] // self.spatial_merge_unit,
737
+
738
+ # Use pre-computed cos_sin_cache from RotaryEmbedding
739
+ cos, sin = self.rotary_pos_emb.get_cos_sin(max_size)
740
+
741
+ cos_h = cos[pos_ids[:, 0]] # (num_tokens, rotary_dim // 2)
742
+ cos_w = cos[pos_ids[:, 1]]
743
+ sin_h = sin[pos_ids[:, 0]]
744
+ sin_w = sin[pos_ids[:, 1]]
745
+
746
+ cos_combined = torch.cat([cos_h, cos_w], dim=-1)
747
+ sin_combined = torch.cat([sin_h, sin_w], dim=-1)
748
+
749
+ cos_combined = cos_combined.reshape(
750
+ cos_combined.shape[0] // self.spatial_merge_unit,
751
+ self.spatial_merge_unit,
752
+ -1,
753
+ )
754
+ sin_combined = sin_combined.reshape(
755
+ sin_combined.shape[0] // self.spatial_merge_unit,
756
self.spatial_merge_unit,
757
-1,
758
)
759
768
- return rotary_pos_emb
760
+ return cos_combined, sin_combined
761
762
def get_window_index_thw(self, grid_t, grid_h, grid_w):
763
vit_merger_window_size = (
@@ -807,14 +799,19 @@ class Qwen2_5_VisionTransformer(nn.Module):
799
@lru_cache(maxsize=1024) # noqa: B019
800
def get_rope_by_thw(self, t, h, w):
801
window_index_thw, cu_seqlens_window_thw = self.get_window_index_thw(t, h, w)
810
- rotary_pos_emb_thw = self.rotary_pos_emb_thw(t, h, w)
811
- rotary_pos_emb_thw = rotary_pos_emb_thw[window_index_thw, :, :]
812
- rotary_pos_emb_thw = rotary_pos_emb_thw.flatten(start_dim=0, end_dim=1)
802
+ cos_thw, sin_thw = self.rotary_pos_emb_thw(t, h, w)
803
+
804
+ cos_thw = cos_thw[window_index_thw, :, :]
805
+ cos_thw = cos_thw.flatten(start_dim=0, end_dim=1)
806
+ sin_thw = sin_thw[window_index_thw, :, :]
807
+ sin_thw = sin_thw.flatten(start_dim=0, end_dim=1)
808
+
809
cu_seqlens_thw = torch.repeat_interleave(
810
torch.tensor([h * w], dtype=torch.int32), t
811
)
812
return (
817
- rotary_pos_emb_thw,
813
+ cos_thw,
814
+ sin_thw,
815
window_index_thw,
816
cu_seqlens_window_thw,
817
cu_seqlens_thw,
@@ -849,7 +846,8 @@ class Qwen2_5_VisionTransformer(nn.Module):
846
) -> torch.Tensor:
847
# patchify
848
seq_len, _ = x.size()
852
- rotary_pos_emb = []
849
+ rotary_pos_emb_cos = []
850
+ rotary_pos_emb_sin = []
851
window_index: list = []
852
cu_window_seqlens: list = [torch.tensor([0], dtype=torch.int32)]
853
cu_seqlens: list = []
@@ -865,7 +863,8 @@ class Qwen2_5_VisionTransformer(nn.Module):
863
llm_w = w // self.spatial_merge_size
864
865
(
868
- rotary_pos_emb_thw,
866
+ cos_thw,
867
+ sin_thw,
868
window_index_thw,
869
cu_seqlens_window_thw,
870
cu_seqlens_thw,
@@ -878,11 +877,13 @@ class Qwen2_5_VisionTransformer(nn.Module):
877
cu_window_seqlens_last = cu_seqlens_window_thw[-1]
878
cu_window_seqlens.append(cu_seqlens_window_thw)
879
881
- rotary_pos_emb.append(rotary_pos_emb_thw)
880
+ rotary_pos_emb_cos.append(cos_thw)
881
+ rotary_pos_emb_sin.append(sin_thw)
882
883
cu_seqlens.append(cu_seqlens_thw)
884
885
- rotary_pos_emb = torch.cat(rotary_pos_emb)
885
+ rotary_pos_emb_cos = torch.cat(rotary_pos_emb_cos)
886
+ rotary_pos_emb_sin = torch.cat(rotary_pos_emb_sin)
887
window_index = torch.cat(window_index)
888
# compute reverse indices
889
reverse_indices = self.invert_permutation(window_index)
@@ -901,7 +902,12 @@ class Qwen2_5_VisionTransformer(nn.Module):
902
903
cu_seqlens = cu_seqlens.to(device=self.device, non_blocking=True)
904
cu_window_seqlens = cu_window_seqlens.to(device=self.device, non_blocking=True)
904
- rotary_pos_emb = rotary_pos_emb.to(device=self.device, non_blocking=True)
905
+ rotary_pos_emb_cos = rotary_pos_emb_cos.to(
906
+ device=self.device, non_blocking=True
907
+ )
908
+ rotary_pos_emb_sin = rotary_pos_emb_sin.to(
909
+ device=self.device, non_blocking=True
910
+ )
911
window_index = window_index.to(device=hidden_states.device, non_blocking=True)
912
reverse_indices = reverse_indices.to(
913
device=hidden_states.device, non_blocking=True
@@ -928,7 +934,8 @@ class Qwen2_5_VisionTransformer(nn.Module):
934
hidden_states = blk(
935
hidden_states,
936
cu_seqlens=cu_seqlens_now,
931
- rotary_pos_emb=rotary_pos_emb,
937
+ rotary_pos_emb_cos=rotary_pos_emb_cos,
938
+ rotary_pos_emb_sin=rotary_pos_emb_sin,
939
max_seqlen=max_seqlen_now,
940
seqlens=seqlens_now,
941
)
vllm/model_executor/models/qwen2_vl.py
+46
-89
@@ -32,7 +32,7 @@ from typing import Annotated, Any, Literal, TypeAlias
32
import torch
33
import torch.nn as nn
34
import torch.nn.functional as F
35
-from einops import rearrange, repeat
35
+from einops import rearrange
36
from transformers import BatchFeature
37
from transformers.models.qwen2_vl import Qwen2VLImageProcessor, Qwen2VLProcessor
38
from transformers.models.qwen2_vl.configuration_qwen2_vl import (
@@ -59,7 +59,9 @@ from vllm.model_executor.layers.linear import (
59
RowParallelLinear,
60
)
61
from vllm.model_executor.layers.quantization import QuantizationConfig
62
+from vllm.model_executor.layers.rotary_embedding import get_rope
63
from vllm.model_executor.layers.rotary_embedding.common import (
64
+ apply_rotary_emb_torch,
65
dispatch_rotary_emb_function,
66
)
67
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
@@ -275,47 +277,13 @@ class Qwen2VisionMLP(nn.Module):
277
return x
278
279
278
-def rotate_half(x: torch.Tensor, interleaved: bool = False) -> torch.Tensor:
279
- if not interleaved:
280
- x1, x2 = x.chunk(2, dim=-1)
281
- return torch.cat((-x2, x1), dim=-1)
282
- else:
283
- x1, x2 = x[..., ::2], x[..., 1::2]
284
- return rearrange(
285
- torch.stack((-x2, x1), dim=-1), "... d two -> ... (d two)", two=2
286
- )
287
-
288
-
289
-def apply_rotary_emb_torch(
290
- x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor, interleaved: bool = False
280
+def apply_rotary_pos_emb_vision(
281
+ t: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
282
) -> torch.Tensor:
292
- """
293
- x: (batch_size, seqlen, nheads, headdim)
294
- cos, sin: (seqlen, rotary_dim / 2) or (batch_size, seqlen, rotary_dim / 2)
295
- """
296
- ro_dim = cos.shape[-1] * 2
297
- assert ro_dim <= x.shape[-1]
298
- cos = repeat(
299
- cos, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)"
300
- )
301
- sin = repeat(
302
- sin, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)"
283
+ rotary_emb_function = dispatch_rotary_emb_function(
284
+ default=partial(apply_rotary_emb_torch, is_neox_style=True)
285
)
304
- return torch.cat(
305
- [
306
- x[..., :ro_dim] * cos + rotate_half(x[..., :ro_dim], interleaved) * sin,
307
- x[..., ro_dim:],
308
- ],
309
- dim=-1,
310
- )
311
-
312
-
313
-def apply_rotary_pos_emb_vision(t: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor:
314
- rotary_emb_function = dispatch_rotary_emb_function(default=apply_rotary_emb_torch)
315
- t_ = t.float()
316
- cos = freqs.cos()
317
- sin = freqs.sin()
318
- output = rotary_emb_function(t_, cos, sin).type_as(t)
286
+ output = rotary_emb_function(t, cos, sin).type_as(t)
287
return output
288
289
@@ -412,7 +380,8 @@ class Qwen2VisionAttention(nn.Module):
380
self,
381
x: torch.Tensor,
382
cu_seqlens: torch.Tensor,
415
- rotary_pos_emb: torch.Tensor,
383
+ rotary_pos_emb_cos: torch.Tensor,
384
+ rotary_pos_emb_sin: torch.Tensor,
385
max_seqlen: int | None = None, # Only used for Flash Attention
386
seqlens: list[int] | None = None, # Only used for xFormers
387
) -> torch.Tensor:
@@ -424,11 +393,13 @@ class Qwen2VisionAttention(nn.Module):
393
batch_size = q.shape[1]
394
395
q, k, v = (rearrange(x, "s b ... -> b s ...") for x in (q, k, v))
427
- if rotary_pos_emb is not None:
428
- # [2 * b, s, heads, head_dim]
429
- qk_concat = torch.cat([q, k], dim=0)
430
- qk_rotated = apply_rotary_pos_emb_vision(qk_concat, rotary_pos_emb)
431
- q, k = torch.chunk(qk_rotated, 2, dim=0)
396
+
397
+ # [2 * b, s, heads, head_dim]
398
+ qk_concat = torch.cat([q, k], dim=0)
399
+ qk_rotated = apply_rotary_pos_emb_vision(
400
+ qk_concat, rotary_pos_emb_cos, rotary_pos_emb_sin
401
+ )
402
+ q, k = torch.chunk(qk_rotated, 2, dim=0)
403
404
if self.is_flash_attn_backend:
405
q, k, v = (rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v])
@@ -534,14 +505,16 @@ class Qwen2VisionBlock(nn.Module):
505
self,
506
x: torch.Tensor,
507
cu_seqlens: torch.Tensor,
537
- rotary_pos_emb: torch.Tensor,
508
+ rotary_pos_emb_cos: torch.Tensor,
509
+ rotary_pos_emb_sin: torch.Tensor,
510
max_seqlen: int | None = None, # Only used for Flash Attention
511
seqlens: list[int] | None = None, # Only used for xFormers
512
) -> torch.Tensor:
513
x = x + self.attn(
514
self.norm1(x),
515
cu_seqlens=cu_seqlens,
544
- rotary_pos_emb=rotary_pos_emb,
516
+ rotary_pos_emb_cos=rotary_pos_emb_cos,
517
+ rotary_pos_emb_sin=rotary_pos_emb_sin,
518
max_seqlen=max_seqlen,
519
seqlens=seqlens,
520
)
@@ -628,40 +601,6 @@ class Qwen2VisionPatchMerger(nn.Module):
601
return out
602
603
631
-class Qwen2VisionRotaryEmbedding(nn.Module):
632
- def __init__(self, dim: int, theta: float = 10000.0) -> None:
633
- super().__init__()
634
- self.dim = dim
635
- self.theta = theta
636
- inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
637
- self.register_buffer("inv_freq", inv_freq, persistent=False)
638
- self._seq_len_cached = 0
639
- self._freqs_cached = None
640
-
641
- def update_freqs_cache(self, seqlen: int) -> None:
642
- if seqlen > self._seq_len_cached:
643
- seqlen *= 2
644
- self._seq_len_cached = seqlen
645
- self.inv_freq = 1.0 / (
646
- self.theta
647
- ** (
648
- torch.arange(
649
- 0, self.dim, 2, dtype=torch.float, device=self.inv_freq.device
650
- )
651
- / self.dim
652
- )
653
- )
654
- seq = torch.arange(
655
- seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype
656
- )
657
- freqs = torch.outer(seq, self.inv_freq)
658
- self._freqs_cached = freqs
659
-
660
- def forward(self, seqlen: int) -> torch.Tensor:
661
- self.update_freqs_cache(seqlen)
662
- return self._freqs_cached[:seqlen]
663
-
664
-
604
class Qwen2VisionTransformer(nn.Module):
605
def __init__(
606
self,
@@ -700,7 +639,13 @@ class Qwen2VisionTransformer(nn.Module):
639
640
norm_layer = partial(nn.LayerNorm, eps=norm_eps)
641
head_dim = embed_dim // num_heads
703
- self.rotary_pos_emb = Qwen2VisionRotaryEmbedding(head_dim // 2)
642
+ self.rotary_pos_emb = get_rope(
643
+ head_size=head_dim,
644
+ rotary_dim=head_dim // 2,
645
+ max_position=8192,
646
+ base=10000.0,
647
+ is_neox_style=True,
648
+ )
649
650
self.blocks = nn.ModuleList(
651
[
@@ -744,7 +689,9 @@ class Qwen2VisionTransformer(nn.Module):
689
def device(self) -> torch.device:
690
return self.patch_embed.proj.weight.device
691
747
- def rot_pos_emb(self, grid_thw: list[list[int]]) -> torch.Tensor:
692
+ def rot_pos_emb(
693
+ self, grid_thw: list[list[int]]
694
+ ) -> tuple[torch.Tensor, torch.Tensor]:
695
pos_ids = []
696
max_grid_size = 0
697
for t, h, w in grid_thw:
@@ -773,9 +720,18 @@ class Qwen2VisionTransformer(nn.Module):
720
pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
721
max_grid_size = max(max_grid_size, h, w)
722
pos_ids = torch.cat(pos_ids, dim=0)
776
- rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
777
- rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
778
- return rotary_pos_emb
723
+
724
+ # Use pre-computed cos_sin_cache from RotaryEmbedding
725
+ cos, sin = self.rotary_pos_emb.get_cos_sin(max_grid_size)
726
+
727
+ cos_h = cos[pos_ids[:, 0]] # (num_tokens, rotary_dim // 2)
728
+ cos_w = cos[pos_ids[:, 1]]
729
+ sin_h = sin[pos_ids[:, 0]]
730
+ sin_w = sin[pos_ids[:, 1]]
731
+
732
+ cos_combined = torch.cat([cos_h, cos_w], dim=-1)
733
+ sin_combined = torch.cat([sin_h, sin_w], dim=-1)
734
+ return cos_combined, sin_combined
735
736
def compute_attn_mask_seqlen(
737
self, cu_seqlens: torch.Tensor
@@ -806,7 +762,7 @@ class Qwen2VisionTransformer(nn.Module):
762
grid_thw_list = grid_thw.tolist()
763
764
# compute position embedding
809
- rotary_pos_emb = self.rot_pos_emb(grid_thw_list)
765
+ rotary_pos_emb_cos, rotary_pos_emb_sin = self.rot_pos_emb(grid_thw_list)
766
767
# compute cu_seqlens
768
cu_seqlens = torch.repeat_interleave(
@@ -824,7 +780,8 @@ class Qwen2VisionTransformer(nn.Module):
780
x = blk(
781
x,
782
cu_seqlens=cu_seqlens,
827
- rotary_pos_emb=rotary_pos_emb,
783
+ rotary_pos_emb_cos=rotary_pos_emb_cos,
784
+ rotary_pos_emb_sin=rotary_pos_emb_sin,
785
max_seqlen=max_seqlen,
786
seqlens=seqlens,
787
)
vllm/model_executor/models/qwen3_omni_moe_thinker.py
+30
-10
@@ -60,6 +60,7 @@ from vllm.model_executor.layers.linear import (
60
)
61
from vllm.model_executor.layers.logits_processor import LogitsProcessor
62
from vllm.model_executor.layers.quantization import QuantizationConfig
63
+from vllm.model_executor.layers.rotary_embedding import get_rope
64
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
65
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
66
from vllm.model_executor.models.qwen2_audio import Qwen2AudioProcessingInfo
@@ -90,7 +91,6 @@ from .qwen2_5_omni_thinker import (
91
)
92
from .qwen2_5_vl import (
93
Qwen2_5_VisionAttention,
93
- Qwen2_5_VisionRotaryEmbedding,
94
Qwen2_5_VLProcessingInfo,
95
)
96
from .qwen3_moe import Qwen3MoeForCausalLM, Qwen3MoeModel
@@ -221,14 +221,16 @@ class Qwen3_VisionBlock(nn.Module):
221
self,
222
x: torch.Tensor,
223
cu_seqlens: torch.Tensor,
224
- rotary_pos_emb: torch.Tensor,
224
+ rotary_pos_emb_cos: torch.Tensor,
225
+ rotary_pos_emb_sin: torch.Tensor,
226
max_seqlen: torch.Tensor, # Only used for Flash Attention
227
seqlens: torch.Tensor, # Only used for xFormers
228
) -> torch.Tensor:
229
x = x + self.attn(
230
self.norm1(x),
231
cu_seqlens=cu_seqlens,
231
- rotary_pos_emb=rotary_pos_emb,
232
+ rotary_pos_emb_cos=rotary_pos_emb_cos,
233
+ rotary_pos_emb_sin=rotary_pos_emb_sin,
234
max_seqlen=max_seqlen,
235
seqlens=seqlens,
236
)
@@ -332,7 +334,13 @@ class Qwen3Omni_VisionTransformer(nn.Module):
334
335
norm_layer = partial(nn.LayerNorm, eps=norm_eps)
336
head_dim = self.hidden_size // self.num_heads
335
- self.rotary_pos_emb = Qwen2_5_VisionRotaryEmbedding(head_dim // 2)
337
+ self.rotary_pos_emb = get_rope(
338
+ head_size=head_dim,
339
+ rotary_dim=head_dim // 2,
340
+ max_position=8192,
341
+ base=10000.0,
342
+ is_neox_style=True,
343
+ )
344
345
self.blocks = nn.ModuleList(
346
[
@@ -416,9 +424,19 @@ class Qwen3Omni_VisionTransformer(nn.Module):
424
pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
425
pos_ids = torch.cat(pos_ids, dim=0)
426
max_grid_size = grid_thw[:, 1:].max()
419
- rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
420
- rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
421
- return rotary_pos_emb
427
+
428
+ # Use pre-computed cos_sin_cache from RotaryEmbedding
429
+ cos, sin = self.rotary_pos_emb.get_cos_sin(max_grid_size)
430
+
431
+ cos_h = cos[pos_ids[:, 0]] # (num_tokens, rotary_dim // 2)
432
+ cos_w = cos[pos_ids[:, 1]]
433
+ sin_h = sin[pos_ids[:, 0]]
434
+ sin_w = sin[pos_ids[:, 1]]
435
+
436
+ cos_combined = torch.cat([cos_h, cos_w], dim=-1)
437
+ sin_combined = torch.cat([sin_h, sin_w], dim=-1)
438
+
439
+ return cos_combined, sin_combined
440
441
def fast_pos_embed_interpolate(self, grid_thw: list[list[int]]) -> torch.Tensor:
442
num_grid_per_side = self.num_grid_per_side
@@ -508,7 +526,7 @@ class Qwen3Omni_VisionTransformer(nn.Module):
526
if self.apply_vit_abs_pos_embed:
527
pos_embeds = self.fast_pos_embed_interpolate(grid_thw)
528
hidden_states = hidden_states + pos_embeds
511
- rotary_pos_emb = self.rot_pos_emb(grid_thw)
529
+ rotary_pos_emb_cos, rotary_pos_emb_sin = self.rot_pos_emb(grid_thw)
530
531
cu_seqlens = torch.repeat_interleave(
532
grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
@@ -519,7 +537,8 @@ class Qwen3Omni_VisionTransformer(nn.Module):
537
cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
538
539
hidden_states = hidden_states.unsqueeze(1)
522
- rotary_pos_emb = rotary_pos_emb.to(hidden_states.device)
540
+ rotary_pos_emb_cos = rotary_pos_emb_cos.to(hidden_states.device)
541
+ rotary_pos_emb_sin = rotary_pos_emb_sin.to(hidden_states.device)
542
max_seqlen, seqlens = self.compute_attn_mask_seqlen(cu_seqlens)
543
544
hidden_states_list = []
@@ -529,7 +548,8 @@ class Qwen3Omni_VisionTransformer(nn.Module):
548
hidden_states = blk(
549
hidden_states,
550
cu_seqlens=cu_seqlens,
532
- rotary_pos_emb=rotary_pos_emb,
551
+ rotary_pos_emb_cos=rotary_pos_emb_cos,
552
+ rotary_pos_emb_sin=rotary_pos_emb_sin,
553
max_seqlen=max_seqlen,
554
seqlens=seqlens,
555
)
vllm/model_executor/models/qwen3_vl.py
+34
-10
@@ -63,6 +63,7 @@ from vllm.model_executor.layers.linear import (
63
)
64
from vllm.model_executor.layers.logits_processor import LogitsProcessor
65
from vllm.model_executor.layers.quantization import QuantizationConfig
66
+from vllm.model_executor.layers.rotary_embedding import get_rope
67
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
68
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
69
from vllm.model_executor.models.module_mapping import MultiModelKeys
@@ -95,7 +96,6 @@ from .interfaces import (
96
)
97
from .qwen2_5_vl import (
98
Qwen2_5_VisionAttention,
98
- Qwen2_5_VisionRotaryEmbedding,
99
Qwen2_5_VLImageEmbeddingInputs,
100
Qwen2_5_VLImageInputs,
101
Qwen2_5_VLImagePixelInputs,
@@ -232,14 +232,16 @@ class Qwen3_VisionBlock(nn.Module):
232
self,
233
x: torch.Tensor,
234
cu_seqlens: torch.Tensor,
235
- rotary_pos_emb: torch.Tensor,
235
+ rotary_pos_emb_cos: torch.Tensor,
236
+ rotary_pos_emb_sin: torch.Tensor,
237
max_seqlen: torch.Tensor, # Only used for Flash Attention
238
seqlens: torch.Tensor, # Only used for xFormers
239
) -> torch.Tensor:
240
x = x + self.attn(
241
self.norm1(x),
242
cu_seqlens=cu_seqlens,
242
- rotary_pos_emb=rotary_pos_emb,
243
+ rotary_pos_emb_cos=rotary_pos_emb_cos,
244
+ rotary_pos_emb_sin=rotary_pos_emb_sin,
245
max_seqlen=max_seqlen,
246
seqlens=seqlens,
247
)
@@ -339,7 +341,13 @@ class Qwen3_VisionTransformer(nn.Module):
341
342
norm_layer = partial(nn.LayerNorm, eps=norm_eps)
343
head_dim = self.hidden_size // self.num_heads
342
- self.rotary_pos_emb = Qwen2_5_VisionRotaryEmbedding(head_dim // 2)
344
+ self.rotary_pos_emb = get_rope(
345
+ head_size=head_dim,
346
+ rotary_dim=head_dim // 2,
347
+ max_position=8192,
348
+ base=10000.0,
349
+ is_neox_style=True,
350
+ )
351
352
self.merger = Qwen3_VisionPatchMerger(
353
d_model=vision_config.out_hidden_size,
@@ -452,9 +460,19 @@ class Qwen3_VisionTransformer(nn.Module):
460
for t, h, w in grid_thw
461
]
462
pos_ids = torch.cat(pos_ids, dim=0)
455
- rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
456
- rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
457
- return rotary_pos_emb
463
+
464
+ # Use pre-computed cos_sin_cache from RotaryEmbedding
465
+ cos, sin = self.rotary_pos_emb.get_cos_sin(max_grid_size)
466
+
467
+ cos_h = cos[pos_ids[:, 0]] # (num_tokens, rotary_dim // 2)
468
+ cos_w = cos[pos_ids[:, 1]]
469
+ sin_h = sin[pos_ids[:, 0]]
470
+ sin_w = sin[pos_ids[:, 1]]
471
+
472
+ cos_combined = torch.cat([cos_h, cos_w], dim=-1)
473
+ sin_combined = torch.cat([sin_h, sin_w], dim=-1)
474
+
475
+ return cos_combined, sin_combined
476
477
def fast_pos_embed_interpolate(self, grid_thw: list[list[int]]) -> torch.Tensor:
478
num_grid_per_side = self.num_grid_per_side
@@ -547,8 +565,13 @@ class Qwen3_VisionTransformer(nn.Module):
565
566
pos_embeds = self.fast_pos_embed_interpolate(grid_thw_list)
567
hidden_states = hidden_states + pos_embeds
550
- rotary_pos_emb = self.rot_pos_emb(grid_thw_list)
551
- rotary_pos_emb = rotary_pos_emb.to(hidden_states.device, non_blocking=True)
568
+ rotary_pos_emb_cos, rotary_pos_emb_sin = self.rot_pos_emb(grid_thw_list)
569
+ rotary_pos_emb_cos = rotary_pos_emb_cos.to(
570
+ hidden_states.device, non_blocking=True
571
+ )
572
+ rotary_pos_emb_sin = rotary_pos_emb_sin.to(
573
+ hidden_states.device, non_blocking=True
574
+ )
575
576
cu_seqlens = torch.repeat_interleave(
577
grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
@@ -564,7 +587,8 @@ class Qwen3_VisionTransformer(nn.Module):
587
hidden_states = blk(
588
hidden_states,
589
cu_seqlens=cu_seqlens,
567
- rotary_pos_emb=rotary_pos_emb,
590
+ rotary_pos_emb_cos=rotary_pos_emb_cos,
591
+ rotary_pos_emb_sin=rotary_pos_emb_sin,
592
max_seqlen=max_seqlen,
593
seqlens=seqlens,
594
)