[feature] enable pre compile jit deep_gemm (#5580)

python/sglang/srt/model_executor/model_runner.py

@@ -42,6 +42,10 @@ from sglang.srt.layers.dp_attention import (
 )
 from sglang.srt.layers.logits_processor import LogitsProcessorOutput
 from sglang.srt.layers.quantization import monkey_patch_isinstance_for_vllm_base_layer
+from sglang.srt.layers.quantization.deep_gemm import (
+    _ENABLE_JIT_DEEPGEMM,
+    update_deep_gemm_config,
+)
 from sglang.srt.layers.sampler import Sampler
 from sglang.srt.layers.torchao_utils import apply_torchao_config_to_model
 from sglang.srt.lora.lora_manager import LoRAManager
@@ -169,6 +173,10 @@ class ModelRunner:
         # Get memory before model loading
         min_per_gpu_memory = self.init_torch_distributed()
 
+        # Update deep gemm configure
+        if _ENABLE_JIT_DEEPGEMM:
+            update_deep_gemm_config(gpu_id, server_args)
+
         # If it is a draft model tp_group can be different.
         self.initialize(min_per_gpu_memory)