remove max_seq_len_to_capture (#1200)

Signed-off-by: Ao Tang <[email protected]>

Author: suiyoubi

nemo_curator/models/qwen_vl.py

@@ -91,7 +91,6 @@ def setup(self) -> None:
             model=self.weight_file,
             limit_mm_per_prompt={"image": 0, "video": 1},
             quantization="fp8" if self.fp8 else None,
-            max_seq_len_to_capture=32768,
             max_model_len=32768,
             gpu_memory_utilization=0.85,
             mm_processor_kwargs=mm_processor_kwargs,

pyproject.toml

@@ -144,7 +144,7 @@ video_cuda12 = [
     "PyNvVideoCodec==2.0.2; (platform_machine == 'x86_64' and platform_system != 'Darwin')",
     "torch<=2.8.0",
     "torchaudio",
-    "vllm==0.10.2; (platform_machine == 'x86_64' and platform_system != 'Darwin')",
+    "vllm==0.11.0; (platform_machine == 'x86_64' and platform_system != 'Darwin')",
 ]
 
 # All dependencies

tests/models/test_qwen_vl.py

@@ -128,7 +128,6 @@ def test_setup_with_fp8(self, mock_logger: Mock, mock_sampling_params: Mock, moc
             model=self.qwen_vl.weight_file,
             limit_mm_per_prompt={"image": 0, "video": 1},
             quantization="fp8",
-            max_seq_len_to_capture=32768,
             max_model_len=32768,
             gpu_memory_utilization=0.85,
             mm_processor_kwargs=expected_mm_processor_kwargs,