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 5, 18:30 UTC
Restore non-hf processor path for Nano-Nemotron-VL (bypass `call_hf_processor_mm_only`) - fixes #38018 (#38567)
Signed-off-by: Netanel Haber <58652339+netanel-haber@users.noreply.github.com> Co-authored-by: tomeras91 <57313761+tomeras91@users.noreply.github.com>
Netanel Haber committed
Mar 31, 2026 at 00:56 UTC
e812bf70bd668b4d28e7135ae1577d252c08ee5c
1 file changed
+14
vllm/model_executor/models/nano_nemotron_vl.py
+14
@@ -307,6 +307,20 @@ class NanoNemotronVLProcessingInfo(BaseProcessingInfo):
307
class NanoNemotronVLMultiModalProcessor(
308
BaseMultiModalProcessor[NanoNemotronVLProcessingInfo]
309
):
310
+ def _call_hf_processor(
311
+ self,
312
+ prompt: str,
313
+ mm_data: Mapping[str, object],
314
+ mm_kwargs: Mapping[str, object],
315
+ tok_kwargs: Mapping[str, object],
316
+ ) -> BatchFeature:
317
+ """
318
+ Bypass `call_hf_processor_mm_only` by no-op overriding`_call_hf_processor`,
319
+ so it chooses this path:
320
+ `type(self)._call_hf_processor != BaseMultiModalProcessor._call_hf_processor`
321
+ """
322
+ return super()._call_hf_processor(prompt, mm_data, mm_kwargs, tok_kwargs)
323
+
324
def _get_image_fields_config(self, hf_inputs: BatchFeature):
325
if self.info.is_dynamic_tiler:
326
pixel_values_flat = MultiModalFieldConfig.batched("image")