[misc] deep_gemm fallback to NVRTC when NVCC not found (sgl-project#6252)

Author: Alcanderian

python/sglang/srt/layers/quantization/deep_gemm.py

@@ -15,6 +15,7 @@
 if is_cuda():
     import deep_gemm
     from deep_gemm import get_num_sms
+    from deep_gemm.jit.compiler import get_nvcc_compiler
     from deep_gemm.jit_kernels.gemm import get_best_configs
     from deep_gemm.jit_kernels.runtime import FP8GemmRuntime, GemmType
     from deep_gemm.jit_kernels.tuner import jit_tuner
@@ -48,7 +49,17 @@ def get_enable_jit_deepgemm():
 # Refer to https://github.com/deepseek-ai/DeepGEMM/commit/d75b218b7b8f4a5dd5406ac87905039ead3ae42f
 # NVRTC may have performance loss with some cases.
 # And NVCC JIT speed is also 9x faster in the ref commit
-os.environ["DG_JIT_USE_NVRTC"] = os.getenv("SGL_DG_USE_NVRTC", "0")
+_USE_NVRTC_DEFAULT = "0"
+if _ENABLE_JIT_DEEPGEMM:
+    try:
+        get_nvcc_compiler()
+    except:
+        logger.warning(
+            "NVCC Compiler not found, use NVRTC for DeepGEMM JIT "
+            "and may have performance loss with some cases."
+        )
+        _USE_NVRTC_DEFAULT = "1"
+os.environ["DG_JIT_USE_NVRTC"] = os.getenv("SGL_DG_USE_NVRTC", _USE_NVRTC_DEFAULT)
 
 
 def update_deep_gemm_config(gpu_id: int, server_args: ServerArgs):