xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-17 01:35:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix Mistral3 support on SM100/SM120 (#20998)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin 2025-07-15 23:08:41 -04:00 committed by GitHub
parent e9534c7202
commit f46098335b
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 7 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/model_executor/models/pixtral.py
View File

@ -43,6 +43,7 @@ from vllm.multimodal.processing import (BaseMultiModalProcessor,
PromptReplacement, PromptUpdate, PromptReplacement, PromptUpdate,
PromptUpdateDetails) PromptUpdateDetails)
from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.tokenizer import (MistralTokenizer, from vllm.transformers_utils.tokenizer import (MistralTokenizer,
cached_tokenizer_from_config) cached_tokenizer_from_config)
@ -54,7 +55,12 @@ from .vision import VisionEncoderInfo, resolve_visual_encoder_outputs
try: try:
from xformers import ops as xops from xformers import ops as xops
USE_XFORMERS_OPS = True if (current_platform.is_cuda()
and current_platform.has_device_capability(100)):
# Xformers FA is not compatible with B200
USE_XFORMERS_OPS = False
else:
USE_XFORMERS_OPS = True
except ImportError: except ImportError:
USE_XFORMERS_OPS = False USE_XFORMERS_OPS = False
@ -1082,7 +1088,6 @@ class PixtralHFAttention(nn.Module):
# Transpose q and k back for attention # Transpose q and k back for attention
q = q.transpose(1, 2).contiguous() q = q.transpose(1, 2).contiguous()
k = k.transpose(1, 2).contiguous() k = k.transpose(1, 2).contiguous()
out = xops.memory_efficient_attention(q, out = xops.memory_efficient_attention(q,
k, k,
v, v,