Add PDL support for quant kernel and rope kernel (#9106)

python/sglang/srt/entrypoints/engine.py

@@ -635,6 +635,8 @@ def _set_envs_and_config(server_args: ServerArgs):
         os.environ["NCCL_NVLS_ENABLE"] = str(int(server_args.enable_nccl_nvls))
     os.environ["CUDA_DEVICE_MAX_CONNECTIONS"] = "4"
     os.environ["CUDA_MODULE_LOADING"] = "AUTO"
+    # flashinfer uses this environment variable for various kernels from MoE to quant kernels
+    os.environ["TRTLLM_ENABLE_PDL"] = "1"
 
     # Set prometheus env vars
     if server_args.enable_metrics:

python/sglang/srt/server_args.py

@@ -550,7 +550,6 @@ class ServerArgs:
             assert (
                 self.quantization == "modelopt_fp4"
             ), "modelopt_fp4 quantization is required for Flashinfer MOE"
-            os.environ["TRTLLM_ENABLE_PDL"] = "1"
             assert self.ep_size in [
                 1,
                 self.tp_size,