[MM Encoder] Add Triton ViT attention backend (#32183)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py
2026-02-15 06:32:47 -08:00
committed by GitHub
parent 19fab44152
commit 71cd89264f
14 changed files with 178 additions and 51 deletions
@@ -12,6 +12,7 @@ from vllm.v1.attention.backends.registry import AttentionBackendEnum
from vllm.v1.attention.ops.vit_attn_wrappers import (
vit_flash_attn_wrapper,
vit_torch_sdpa_wrapper,
vit_triton_attn_wrapper,
)
logger = init_logger(__name__)
@@ -165,6 +166,41 @@ class MMEncoderAttention(CustomOp):
output = output.reshape(bsz, q_len, -1)
return output
def _forward_triton(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
cu_seqlens: torch.Tensor | None = None,
max_seqlen: torch.Tensor | None = None, # Only used for Flash Attention
) -> torch.Tensor:
"""Input shape:
(batch_size x seq_len x hidden_size) or
(batch_size x seq_len x num_heads x head_size)
"""
assert (cu_seqlens is not None and max_seqlen is not None) or (
cu_seqlens is None and max_seqlen is None
), "cu_seqlens and max_seqlen should be both set or both None."
bsz, q_len = query.size()[:2]
kv_len = key.size(1)
is_reshaped = query.dim() != 4
query, key, value = self.view_qkv_to_4d(query, key, value, bsz, q_len, kv_len)
output = vit_triton_attn_wrapper(
q=query,
k=key,
v=value,
batch_size=bsz,
scale=self.scale,
cu_seqlens=cu_seqlens,
max_seqlen=max_seqlen,
)
if is_reshaped:
output = output.reshape(bsz, q_len, -1)
return output
def forward_native(
self,
query: torch.Tensor,
@@ -185,6 +221,8 @@ class MMEncoderAttention(CustomOp):
) -> torch.Tensor:
if self.is_flash_attn_backend:
return self._forward_fa(query, key, value, cu_seqlens, max_seqlen)
elif self.attn_backend == AttentionBackendEnum.TRITON_ATTN:
return self._forward_triton(query, key, value, cu_seqlens, max_seqlen)
elif self.attn_backend == AttentionBackendEnum.TORCH_SDPA:
return self._forward_sdpa(query, key, value, cu_seqlens)
else:
+5 -4
View File
@@ -573,10 +573,11 @@ class DotsVisionTransformer(nn.Module):
def compute_attn_mask_seqlen(self, cu_seqlens: torch.Tensor) -> int | None:
max_seqlen = None
if (
self.attn_backend == AttentionBackendEnum.FLASH_ATTN
or self.attn_backend == AttentionBackendEnum.ROCM_AITER_FA
):
if self.attn_backend in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.ROCM_AITER_FA,
AttentionBackendEnum.TRITON_ATTN,
}:
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
return max_seqlen
+5 -4
View File
@@ -446,10 +446,11 @@ class Ernie4_5_VisionTransformer(nn.Module):
def compute_attn_mask_seqlen(self, cu_seqlens: torch.Tensor) -> torch.Tensor | None:
max_seqlen = None
if (
self.attn_backend == AttentionBackendEnum.FLASH_ATTN
or self.attn_backend == AttentionBackendEnum.ROCM_AITER_FA
):
if self.attn_backend in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.ROCM_AITER_FA,
AttentionBackendEnum.TRITON_ATTN,
}:
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
return max_seqlen
+5 -4
View File
@@ -723,10 +723,11 @@ class Glm4vVisionTransformer(nn.Module):
cu_seqlens: torch.Tensor,
) -> torch.Tensor | None:
max_seqlen = None
if (
self.attn_backend == AttentionBackendEnum.FLASH_ATTN
or self.attn_backend == AttentionBackendEnum.ROCM_AITER_FA
):
if self.attn_backend in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.ROCM_AITER_FA,
AttentionBackendEnum.TRITON_ATTN,
}:
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
return max_seqlen
+2 -8
View File
@@ -730,14 +730,7 @@ class SiglipEncoder(nn.Module):
head_size=head_dim,
dtype=torch.get_default_dtype(),
)
if self.attn_backend not in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.TORCH_SDPA,
AttentionBackendEnum.ROCM_AITER_FA,
}:
raise RuntimeError(
f"PaddleOCR-VL does not support {self.attn_backend} backend now."
)
self.layers = nn.ModuleList(
[
SiglipEncoderLayer(
@@ -805,6 +798,7 @@ class SiglipEncoder(nn.Module):
if self.attn_backend in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.ROCM_AITER_FA,
AttentionBackendEnum.TRITON_ATTN,
}:
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
+1 -9
View File
@@ -607,15 +607,6 @@ class Qwen2_5_VisionTransformer(nn.Module):
dtype=torch.get_default_dtype(),
)
if self.attn_backend not in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.TORCH_SDPA,
AttentionBackendEnum.ROCM_AITER_FA,
}:
raise RuntimeError(
f"Qwen2.5-VL does not support {self.attn_backend} backend now."
)
with set_model_tag("Qwen2_5_VisionBlock", is_encoder=True):
self.blocks = nn.ModuleList(
[
@@ -761,6 +752,7 @@ class Qwen2_5_VisionTransformer(nn.Module):
if self.attn_backend in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.ROCM_AITER_FA,
AttentionBackendEnum.TRITON_ATTN,
}:
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
return max_seqlen
+1
View File
@@ -642,6 +642,7 @@ class Qwen2VisionTransformer(nn.Module):
if self.attn_backend in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.ROCM_AITER_FA,
AttentionBackendEnum.TRITON_ATTN,
}:
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
return max_seqlen
@@ -391,6 +391,7 @@ class Qwen3OmniMoeAudioEncoder(nn.Module):
if self.attn_backend in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.ROCM_AITER_FA,
AttentionBackendEnum.TRITON_ATTN,
}:
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
return max_seqlen
@@ -919,6 +920,7 @@ class Qwen3Omni_VisionTransformer(nn.Module):
if self.attn_backend in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.ROCM_AITER_FA,
AttentionBackendEnum.TRITON_ATTN,
}:
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
return max_seqlen
+4 -11
View File
@@ -385,14 +385,6 @@ class Qwen3_VisionTransformer(nn.Module):
dtype=torch.get_default_dtype(),
)
if self.attn_backend not in {
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.TORCH_SDPA,
AttentionBackendEnum.ROCM_AITER_FA,
}:
raise RuntimeError(
f"Qwen3-VL does not support {self.attn_backend} backend now."
)
self.blocks = nn.ModuleList(
[
Qwen3_VisionBlock(
@@ -526,9 +518,10 @@ class Qwen3_VisionTransformer(nn.Module):
cu_seqlens: torch.Tensor,
) -> torch.Tensor:
max_seqlen = torch.zeros([], device=cu_seqlens.device)
if (
self.attn_backend == AttentionBackendEnum.FLASH_ATTN
or self.attn_backend == AttentionBackendEnum.ROCM_AITER_FA
if self.attn_backend in (
AttentionBackendEnum.FLASH_ATTN,
AttentionBackendEnum.ROCM_AITER_FA,
AttentionBackendEnum.TRITON_ATTN,
):
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
return max_seqlen
+2 -2
View File
@@ -108,7 +108,7 @@ def get_vit_attn_backend(
multimodal_config: MultiModalConfig | None = (
model_config.multimodal_config if model_config is not None else None
)
except AssertionError:
except (AssertionError, AttributeError):
multimodal_config = None
attn_backend_override = (
@@ -134,7 +134,7 @@ def is_vit_use_data_parallel():
multimodal_config: MultiModalConfig | None = (
model_config.multimodal_config if model_config is not None else None
)
except AssertionError:
except (AssertionError, AttributeError):
multimodal_config = None
mm_encoder_tp_mode = (