obscura/vllm
2
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2026-08-25 07:00:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix DeepSeek v4 topk numerical issue for unaligned max-model-len (#42169)

Browse Source
This commit is contained in:
Wei Zhao
2026-05-09 20:30:08 -07:00
committed by GitHub
parent 27d3bac272
commit 986edc858a
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
csrc/topk.cu
+2 -2
View File
@@ -20,7 +20,7 @@ void launch_persistent_topk(const torch::Tensor& logits,
namespace P = vllm::persistent;
const int64_t num_rows = logits.size(0);
const int64_t stride = logits.size(1);
const int64_t stride = logits.stride(0);
cudaStream_t stream = at::cuda::getCurrentCUDAStream();
static int num_sms = 0;
@@ -243,7 +243,7 @@ void persistent_topk(const torch::Tensor& logits, const torch::Tensor& lengths,
TORCH_CHECK(output.dim() == 2, "output must be 2D");
const int64_t num_rows = logits.size(0);
const int64_t stride = logits.size(1);
const int64_t stride = logits.stride(0);
TORCH_CHECK(lengths.numel() == num_rows, "lengths size mismatch");
TORCH_CHECK(output.size(0) == num_rows && output.size(1) == k,