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 9, 16:00 UTC
[Bugfix]: resolve torch.compile cache conflict between mm_encoder_tp_modes (#32842)
Signed-off-by: Hongjian Zhang <zhanghongjian@xiaohongshu.com> Signed-off-by: Xingran Wang <wangxingran123456@outlook.com> Co-authored-by: Xingran Wang <wangxingran123456@outlook.com>
Hiroken. committed
Jan 24, 2026 at 22:45 UTC
1209b784f2ba976eff2ea24bc33c61f35c6eb213
2 files changed
+8
-1
vllm/config/multimodal.py
+2
-1
@@ -213,7 +213,8 @@ class MultiModalConfig:
213
factors: list[Any] = [
214
self.mm_encoder_attn_backend.name
215
if self.mm_encoder_attn_backend is not None
216
- else None
216
+ else None,
217
+ self.mm_encoder_tp_mode,
218
]
219
hash_str = safe_hash(str(factors).encode(), usedforsecurity=False).hexdigest()
220
return hash_str
vllm/config/vllm.py
+6
@@ -263,6 +263,12 @@ class VllmConfig:
263
vllm_factors.append(__version__)
264
if self.model_config:
265
vllm_factors.append(self.model_config.compute_hash())
266
+ if (
267
+ self.compilation_config
268
+ and getattr(self.compilation_config, "compile_mm_encoder", False)
269
+ and self.model_config.multimodal_config
270
+ ):
271
+ vllm_factors.append(self.model_config.multimodal_config.compute_hash())
272
else:
273
vllm_factors.append("None")
274
if self.cache_config: