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 4, 06:15 UTC
[Spec Decode] Allow multimodal models with a warning (#41752)
Signed-off-by: Li Zhang <lzhanga@amazon.com> Co-authored-by: Li Zhang <lzhanga@amazon.com>
Li committed
May 5, 2026 at 22:16 UTC
2d7d6cf765d0c796070feaa34fcfcc26494d5c31
2 files changed
+6
-6
vllm/v1/spec_decode/dflash.py
+1
-2
@@ -80,9 +80,8 @@ class DFlashProposer(SpecDecodeBaseProposer):
80
)
81
82
@override
83
- def _raise_if_multimodal(self):
83
+ def _warn_if_multimodal(self):
84
# Override to allow multimodal inputs since DFlash supports Qwen3.5 models
85
- # Support for multimodal inputs has not been tested.
85
pass
86
87
@override
vllm/v1/spec_decode/llm_base_proposer.py
+5
-4
@@ -193,7 +193,7 @@ class SpecDecodeBaseProposer:
193
194
if self.needs_extra_input_slots:
195
self._raise_if_padded_drafter_batch_disabled()
196
- self._raise_if_multimodal()
196
+ self._warn_if_multimodal()
197
self._raise_if_mrope()
198
199
self.is_rejected_token_mask: torch.Tensor | None = None
@@ -309,11 +309,12 @@ class SpecDecodeBaseProposer:
309
"disable_padded_drafter_batch in the speculative_config."
310
)
311
312
- def _raise_if_multimodal(self):
312
+ def _warn_if_multimodal(self):
313
if self.supports_mm_inputs:
314
- raise NotImplementedError(
314
+ logger.warning(
315
"Speculative Decoding with draft models or parallel drafting "
316
- "does not support multimodal models yet"
316
+ "does not fully support multimodal models yet. "
317
+ "Proceeding with text-only speculative decoding."
318
)
319
320
def _raise_if_mrope(self):