Skip to content

support all logit values #1323

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
valarLip merged 3 commits into from
Nov 10, 2025
Merged

support all logit values #1323

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
fbd213b
support all logit values
ukannika Nov 3, 2025
9dc5f66
Merge branch 'main' into support_top_k_per_row_all_logit_values
gyohuangxin Nov 4, 2025
74c21ed
fix tests
ukannika Nov 9, 2025
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
2 changes: 1 addition & 1 deletion aiter/jit/optCompilerConfig.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -935,7 +935,7 @@
"verbose": "False",
"blob_gen_cmd": "''"
},
"module_topk_per_row": {
"module_top_k_per_row": {
"srcs": [
"f'{AITER_CSRC_DIR}/kernels/topk_per_row_kernels.cu'",
"f'{AITER_CSRC_DIR}/pybind/topk_per_row_pybind.cu'"
Expand Down
8 changes: 4 additions & 4 deletions aiter/ops/topk.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -196,8 +196,8 @@ def grouped_topk_torch(
return topk_weights.to(dtypes.fp32), topk_ids.to(dtypes.i32)


@compile_ops("module_topk_per_row")
def topk_per_row(
@compile_ops("module_top_k_per_row")
def top_k_per_row_prefill(
logits: torch.Tensor,
rowStarts: torch.Tensor,
rowEnds: torch.Tensor,
Expand All @@ -208,8 +208,8 @@ def topk_per_row(
) -> None: ...


@compile_ops("module_topk_per_row")
def topk_per_row_decode(
@compile_ops("module_top_k_per_row")
def top_k_per_row_decode(
logits: torch.Tensor,
next_n: int,
seqLens: torch.Tensor,
Expand Down
98 changes: 49 additions & 49 deletions csrc/include/rocm_ops.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1289,56 +1289,56 @@ namespace py = pybind11;
#define GEMM_COMMON_PYBIND \
m.def("get_padded_m", &getPaddedM, py::arg("M"), py::arg("N"), py::arg("K"), py::arg("gl"));

#define TOPK_PER_ROW_PYBIND \
m.def("topk_per_row", \
&topk_per_row, \
py::arg("logits"), \
py::arg("rowStarts"), \
py::arg("rowEnds"), \
py::arg("indices"), \
py::arg("numRows"), \
py::arg("stride0"), \
py::arg("stride1")); \
m.def("topk_per_row_decode", \
&topk_per_row_decode, \
py::arg("logits"), \
py::arg("next_n"), \
py::arg("seqLens"), \
py::arg("indices"), \
py::arg("numRows"), \
py::arg("stride0"), \
#define TOP_K_PER_ROW_PYBIND \
m.def("top_k_per_row_prefill", \
&top_k_per_row_prefill, \
py::arg("logits"), \
py::arg("rowStarts"), \
py::arg("rowEnds"), \
py::arg("indices"), \
py::arg("numRows"), \
py::arg("stride0"), \
py::arg("stride1")); \
m.def("top_k_per_row_decode", \
&top_k_per_row_decode, \
py::arg("logits"), \
py::arg("next_n"), \
py::arg("seqLens"), \
py::arg("indices"), \
py::arg("numRows"), \
py::arg("stride0"), \
py::arg("stride1"));

#define MLA_METADATA_PYBIND \
m.def("get_mla_metadata_v1", \
&get_mla_metadata_v1, \
"get_mla_metadata_v1", \
py::arg("seqlens_qo_indptr"), \
py::arg("seqlens_kv_indptr"), \
py::arg("num_heads_per_head_k"), \
py::arg("num_heads_k"), \
py::arg("is_causal"), \
py::arg("work_metadata_ptrs"), \
py::arg("work_info_set"), \
py::arg("work_indptr"), \
py::arg("reduce_indptr"), \
py::arg("reduce_final_map"), \
py::arg("reduce_partial_map"), \
py::arg("kv_granularity") = 16, \
py::arg("max_seqlen_qo") = -1, \
py::arg("uni_seqlen_qo") = -1, \
py::arg("fast_mode") = true, \
py::arg("topk") = -1); \
#define MLA_METADATA_PYBIND \
m.def("get_mla_metadata_v1", \
&get_mla_metadata_v1, \
"get_mla_metadata_v1", \
py::arg("seqlens_qo_indptr"), \
py::arg("seqlens_kv_indptr"), \
py::arg("num_heads_per_head_k"), \
py::arg("num_heads_k"), \
py::arg("is_causal"), \
py::arg("work_metadata_ptrs"), \
py::arg("work_info_set"), \
py::arg("work_indptr"), \
py::arg("reduce_indptr"), \
py::arg("reduce_final_map"), \
py::arg("reduce_partial_map"), \
py::arg("kv_granularity") = 16, \
py::arg("max_seqlen_qo") = -1, \
py::arg("uni_seqlen_qo") = -1, \
py::arg("fast_mode") = true, \
py::arg("topk") = -1); \
m.def("get_mla_metadata_v1_no_redundant", &get_mla_metadata_v1_no_redundant);

#define MLA_REDUCE_PYBIND \
m.def("mla_reduce_v1", \
&mla_reduce_v1, \
"mla_reduce_v1", \
py::arg("partial_output"), \
py::arg("partial_lse"), \
py::arg("reduce_indptr"), \
py::arg("reduce_final_map"), \
py::arg("reduce_partial_map"), \
py::arg("final_output"), \
py::arg("final_lse") = std::nullopt);
#define MLA_REDUCE_PYBIND \
m.def("mla_reduce_v1", \
&mla_reduce_v1, \
"mla_reduce_v1", \
py::arg("partial_output"), \
py::arg("partial_lse"), \
py::arg("reduce_indptr"), \
py::arg("reduce_final_map"), \
py::arg("reduce_partial_map"), \
py::arg("final_output"), \
py::arg("final_lse") = std::nullopt);
28 changes: 14 additions & 14 deletions csrc/include/topk_per_row.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,18 +2,18 @@
// Copyright (C) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <torch/extension.h>

void topk_per_row(const torch::Tensor& logits,
const torch::Tensor& rowStarts,
const torch::Tensor& rowEnds,
torch::Tensor& indices,
int64_t numRows,
int64_t stride0,
int64_t stride1);
void top_k_per_row_prefill(const torch::Tensor& logits,
const torch::Tensor& rowStarts,
const torch::Tensor& rowEnds,
torch::Tensor& indices,
int64_t numRows,
int64_t stride0,
int64_t stride1);

void topk_per_row_decode(const torch::Tensor& logits,
int64_t next_n,
const torch::Tensor& seqLens,
torch::Tensor& indices,
int64_t numRows,
int64_t stride0,
int64_t stride1);
void top_k_per_row_decode(const torch::Tensor& logits,
int64_t next_n,
const torch::Tensor& seqLens,
torch::Tensor& indices,
int64_t numRows,
int64_t stride0,
int64_t stride1);
Loading