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 3, 01:45 UTC
[BugFix] wire make_empty_intermediate_tensors on AyaVision and Voxtral (#43118)
Signed-off-by: Keyi Li <likey6688@gmail.com> Co-authored-by: Keyi Li <likey6688@gmail.com>
Keyi Li committed
May 22, 2026 at 05:26 UTC
79ff0ffa98dc8dd14a8651bce36ce6265ff4d35d
2 files changed
+8
vllm/model_executor/models/aya_vision.py
+4
@@ -358,6 +358,10 @@ class AyaVisionForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsP
358
architectures=["Cohere2ForCausalLM"],
359
)
360
361
+ self.make_empty_intermediate_tensors = (
362
+ self.language_model.make_empty_intermediate_tensors
363
+ )
364
+
365
@property
366
def dtype(self):
367
return next(self.parameters()).dtype
vllm/model_executor/models/voxtral.py
+4
@@ -349,6 +349,10 @@ class VoxtralForConditionalGeneration(
349
dim=config.text_config.hidden_size,
350
)
351
352
+ self.make_empty_intermediate_tensors = (
353
+ self.language_model.make_empty_intermediate_tensors
354
+ )
355
+
356
def get_mm_mapping(self) -> MultiModelKeys:
357
"""Get module prefix for multimodal models to filter LoRA modules."""
358
return MultiModelKeys.from_string_field(