Skip to content

[Misc] Enable V1 FP16 inference on pre-Ampere GPUs #24022

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
DarkLight1337 merged 1 commit into from
Sep 1, 2025
Merged

[Misc] Enable V1 FP16 inference on pre-Ampere GPUs #24022

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 0 additions & 11 deletions vllm/engine/arg_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1432,17 +1432,6 @@ def _is_v1_supported_oracle(self, model_config: ModelConfig) -> bool:
recommend_to_remove=True)
return False

# 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 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":
supported = current_platform.is_kv_cache_dtype_supported(
self.kv_cache_dtype, model_config)
Expand Down