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 14, 05:30 UTC
[Misc] Enable V1 FP16 inference on pre-Ampere GPUs (#24022)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Sep 1, 2025 at 16:12 UTC
d7fbc6ddaccbcbd514cc6e5a48a04666d9930329
1 file changed
-11
vllm/engine/arg_utils.py
-11
@@ -1436,17 +1436,6 @@ class EngineArgs:
1436
recommend_to_remove=True)
1437
return False
1438
1439
- # Triton v3.3 has f16 conversion regression issue on Turing and Volta,
1440
- # which broke fp16 inference
1441
- # see: https://github.com/triton-lang/triton/issues/6698
1442
- if (current_platform.is_cuda()
1443
- and not current_platform.has_device_capability(80)
1444
- and model_config.dtype == torch.float16):
1445
- _raise_or_fallback(
1446
- feature_name="Compute Capability < 8.0 with FP16",
1447
- recommend_to_remove=False)
1448
- return False
1449
-
1439
if self.kv_cache_dtype != "auto":
1440
supported = current_platform.is_kv_cache_dtype_supported(
1441
self.kv_cache_dtype, model_config)