A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 17, 05:30 UTC
ROCm: Allow setting compilation target (#2581)
Rasmus Larsen committed
Jan 29, 2024 at 19:52 UTC
ea8489fce266d69f2fbe314c1385956b1a342e12
1 file changed
+9
-5
setup.py
+9
-5
@@ -287,11 +287,15 @@ if _is_cuda():
287
},
288
))
289
elif _is_hip():
290
- amd_arch = get_amdgpu_offload_arch()
291
- if amd_arch not in ROCM_SUPPORTED_ARCHS:
292
- raise RuntimeError(
293
- f"Only the following arch is supported: {ROCM_SUPPORTED_ARCHS}"
294
- f"amdgpu_arch_found: {amd_arch}")
290
+ amd_archs = os.getenv("GPU_ARCHS")
291
+ if amd_archs is None:
292
+ amd_archs = get_amdgpu_offload_arch()
293
+ for arch in amd_archs.split(";"):
294
+ if arch not in ROCM_SUPPORTED_ARCHS:
295
+ raise RuntimeError(
296
+ f"Only the following arch is supported: {ROCM_SUPPORTED_ARCHS}"
297
+ f"amdgpu_arch_found: {arch}")
298
+ NVCC_FLAGS += [f"--offload-arch={arch}"]
299
300
elif _is_neuron():
301
neuronxcc_version = get_neuronxcc_version()