Skip to content

[sgl-kernel] avoid per_token_quant_fp8.cu hardcode sm_count #8738

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
ispobock merged 1 commit into from
Aug 4, 2025
Merged

[sgl-kernel] avoid per_token_quant_fp8.cu hardcode sm_count #8738

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
5 changes: 2 additions & 3 deletions sgl-kernel/csrc/gemm/per_token_quant_fp8.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -173,9 +173,8 @@ void sgl_per_token_quant_fp8(torch::Tensor input, torch::Tensor output_q, torch:
TORCH_CHECK(hidden_dim % 8 == 0, "Hidden dimension must be divisible by 8, but got ", hidden_dim);

cudaStream_t stream = at::cuda::getCurrentCUDAStream();
// Hard-code sm_count
int sm_count = 132;
constexpr int TOKENS_PER_CTA = 8;
const int sm_count = at::cuda::getCurrentDeviceProperties()->multiProcessorCount;
const int TOKENS_PER_CTA = 8;
const bool use_warp_kernel = (num_tokens >= sm_count * 2 * TOKENS_PER_CTA);
const bool use_vec16 = (hidden_dim % 16 == 0);

Expand Down
Loading