xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-11 00:44:57 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Support compile for Transformers multimodal (#23095)

Browse Source 
Signed-off-by: raushan <raushan@huggingface.co>
This commit is contained in:
Raushan Turganbay 2025-08-18 15:35:48 +02:00 committed by GitHub
parent 569aefd134
commit 0e3bb543f0
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 7 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/model_executor/models/transformers.py
View File

@ -709,6 +709,13 @@ def flatten_and_concat(x: list[torch.Tensor]) -> torch.Tensor:
MultiModalProcessor, MultiModalProcessor,
info=MultiModalProcessingInfo, info=MultiModalProcessingInfo,
dummy_inputs=MultiModalDummyInputsBuilder) dummy_inputs=MultiModalDummyInputsBuilder)
@support_torch_compile(
dynamic_arg_dims={
"input_ids": 0,
"positions": -1,
"intermediate_tensors": 0,
"inputs_embeds": 0,
}) # set `positions` to last dim to support Qwen-mrope
class TransformersForMultimodalLM(TransformersForCausalLM, SupportsMultiModal): class TransformersForMultimodalLM(TransformersForCausalLM, SupportsMultiModal):
# Backwards compatibility for prev released models. State dicts back then # Backwards compatibility for prev released models. State dicts back then
# had different formats and cannot be loaded with `AutoModel` mapping as is # had different formats and cannot be loaded with `AutoModel` mapping as is