Skip to content

[CI/Build] Avoid CUDA IMA on Ada (SM 8.9) by fallback to apply_repetition_penalties_torch #28180

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

Closed
wants to merge 1 commit into from
Closed

[CI/Build] Avoid CUDA IMA on Ada (SM 8.9) by fallback to apply_repetition_penalties_torch #28180

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
15 changes: 14 additions & 1 deletion vllm/_custom_ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

from functools import cache
from typing import TYPE_CHECKING, Literal

import torch
Expand Down Expand Up @@ -366,6 +367,14 @@ def apply_repetition_penalties_cuda(
)


@cache
def _should_use_cuda_repetition_penalties() -> bool:
"""Whether to run the CUDA repetition-penalty kernel on this platform.
Disable on Ada (SM 8.9); fall back to PyTorch there.
"""
return not current_platform.is_device_capability(89)


def apply_repetition_penalties(
logits: torch.Tensor,
prompt_mask: torch.Tensor,
Expand All @@ -380,7 +389,11 @@ def apply_repetition_penalties(
output_mask: A boolean tensor indicating which tokens appear in the output.
repetition_penalties: The repetition penalties of shape (num_seqs, ).
"""
if logits.is_cuda and logits.is_contiguous():
if (
logits.is_cuda
and logits.is_contiguous()
and _should_use_cuda_repetition_penalties()
):
apply_repetition_penalties_cuda(
logits, prompt_mask, output_mask, repetition_penalties
)
Expand Down