[8/n] Migrate merge_attn_states, mamba, sampler to torch stable ABI (continued) (#43361)

Signed-off-by: Mikayla Gawarecki <mikaylagawarecki@gmail.com>
Signed-off-by: Chris Leonard <chleonar@redhat.com>
Co-authored-by: Mikayla Gawarecki <mikaylagawarecki@gmail.com>
Co-authored-by: Shengqi Chen <harry-chen@outlook.com>
This commit is contained in:
Chris Leonard
2026-05-27 09:35:24 -07:00
committed by GitHub
co-authored by Mikayla Gawarecki Shengqi Chen
parent 05c50c721e
commit 284e6f543d
13 changed files with 432 additions and 403 deletions
+9 -7
View File
@@ -126,10 +126,10 @@ struct RadixRowState {
// ============================================================================
struct PersistentTopKParams {
const float* __restrict__ input; // [num_rows, stride]
int32_t* __restrict__ output; // [num_rows, top_k]
int32_t* __restrict__ lengths; // [num_rows]
RadixRowState* row_states; // large path: per-group state
const float* __restrict__ input; // [num_rows, stride]
int32_t* __restrict__ output; // [num_rows, top_k]
const int32_t* __restrict__ lengths; // [num_rows]
RadixRowState* row_states; // large path: per-group state
uint32_t num_rows;
uint32_t stride;
uint32_t top_k; // actual k value for output stride
@@ -1269,9 +1269,11 @@ constexpr int ComputeFilteredTopKVecSize(uint32_t max_len) {
}
template <typename DType, typename IdType, uint32_t MAX_K = 2048>
cudaError_t FilteredTopKRaggedTransform(DType* input, IdType* output_indices,
IdType* lengths, uint32_t num_rows,
uint32_t top_k_val, uint32_t max_len,
cudaError_t FilteredTopKRaggedTransform(const DType* input,
IdType* output_indices,
const IdType* lengths,
uint32_t num_rows, uint32_t top_k_val,
uint32_t max_len,
cudaStream_t stream = 0) {
constexpr size_t smem_size = FILTERED_TOPK_SMEM_DYNAMIC;
constexpr int MAX_VEC = 16 / sizeof(DType);