fix: hyperparameter settings

Author: Silviase

examples/sample_vllm.py

@@ -49,7 +49,7 @@ def parse_args():
     parser.add_argument(
         "--gpu_memory_utilization",
         type=float,
-        default=0.85,
+        default=0.95,
         help="GPU memory utilization for vLLM (default: 0.85)",
     )
     parser.add_argument(

examples/vllm_registry.py

@@ -154,7 +154,7 @@ def _load_qwen3_vl(
     def _engine_args_kimi_vl(self) -> EngineArgs:
         return EngineArgs(
             model=self.model_id,
-            max_model_len=32768,
+            max_model_len=8192,
             trust_remote_code=True,
             limit_mm_per_prompt={self.modality: 5},
         )
@@ -272,11 +272,11 @@ def _load_deepseek_vl2(
     def _engine_args_glm4_5v(self) -> EngineArgs:
         return EngineArgs(
             model=self.model_id,
-            max_model_len=32768,
+            max_model_len=8192,
             max_num_seqs=2,
             trust_remote_code=True,
             enforce_eager=True,
-            limit_mm_per_prompt={self.modality: 5},
+            limit_mm_per_prompt={"image": 5, "video": 0},
         )
 
     def _load_glm4_5v(
@@ -359,7 +359,7 @@ def _load_ovis2(
     def _engine_args_ovis2_5(self) -> EngineArgs:
         return EngineArgs(
             model=self.model_id,
-            max_model_len=4096,
+            max_model_len=8192,
             max_num_seqs=2,
             trust_remote_code=True,
             dtype="half",