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 14, 05:30 UTC
[Bugfix] Fix dummy video number of frames calculation (#25553)
Signed-off-by: Roger Wang <hey@rogerw.io>
Roger Wang committed
Sep 24, 2025 at 02:47 UTC
42488dae690b69cbf5ad022ac22070608e65dc0e
2 files changed
+2
-2
vllm/v1/worker/gpu_model_runner.py
+1
-1
@@ -2807,7 +2807,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
2807
2808
dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
2809
model_config=self.model_config,
2810
- seq_len=self.max_num_tokens,
2810
+ seq_len=self.max_model_len,
2811
mm_counts={modality: 1},
2812
cache=self.mm_budget.cache,
2813
)
vllm/v1/worker/tpu_model_runner.py
+1
-1
@@ -1795,7 +1795,7 @@ class TPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
1795
1796
dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
1797
model_config=self.model_config,
1798
- seq_len=self.max_num_tokens,
1798
+ seq_len=self.max_model_len,
1799
mm_counts={modality: 1},
1800
cache=self.mm_budget.cache,
1801
)