frozenleaves/vllm-dev
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

[V1] Enable V1 for compute capability < 8.0 + FP32 (#23614)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-08-26 18:00:18 +08:00
committed by GitHub
parent b5d34af328
commit 50fede6634
1 changed files with 8 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
vllm/engine/arg_utils.py
View File

@ -1433,15 +1433,15 @@ class EngineArgs:
recommend_to_remove=True)
return False
# Need at least Ampere for now (FA support required).
# Skip this check if we are running on a non-GPU platform,
# or if the device capability is not available
# (e.g. in a Ray actor without GPUs).
# Triton v3.3 has f16 conversion regression issue on Turing and Volta,
# which broke fp16 inference
# see: https://github.com/triton-lang/triton/issues/6698
if (current_platform.is_cuda()
and current_platform.get_device_capability()
and current_platform.get_device_capability().major < 8):
_raise_or_fallback(feature_name="Compute Capability < 8.0",
recommend_to_remove=False)
and not current_platform.has_device_capability(80)
and model_config.dtype == torch.float16):
_raise_or_fallback(
feature_name="Compute Capability < 8.0 with FP16",
recommend_to_remove=False)
return False
if self.kv_cache_dtype != "auto":