repo stringclasses 3
values | commit stringlengths 40 40 | date stringdate 2023-04-02 02:30:17-0500 2026-10-18 07:42:16-0500 | pr_number int64 20 61.9k ⌀ | pr_url stringlengths 44 53 ⌀ | title stringlengths 8 263 | message stringlengths 8 8k | path stringlengths 10 155 | language stringclasses 4
values | license stringclasses 4
values | before_code stringlengths 160 118k | after_code stringlengths 340 120k | diff stringlengths 149 51k | lines_added int64 0 729 | lines_removed int64 0 488 | files_in_commit int64 1 10 | models_mentioned listlengths 0 4 | hardware_mentioned listlengths 0 16 | speedup_claims listlengths 0 10 | prompt stringlengths 569 119k | pr_body stringlengths 0 24k | gains stringlengths 2 197k | kernel_speedup float64 0.78 150 ⌀ | e2e_speedup float64 0.76 136 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
sglang | 70d25873246bb02335b0a107575e289a35662f96 | 2025-12-04T16:38:47+08:00 | 12,446 | https://github.com/sgl-project/sglang/pull/12446 | [CPU] Optimize small oc GEMM for Qwen3-next on CPU (#12446) | [CPU] Optimize small oc GEMM for Qwen3-next on CPU (#12446)
Co-authored-by: Zheng, Beilei <beilei.zheng@intel.com> | sgl-kernel/csrc/cpu/torch_extension_cpu.cpp | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/cpu/torch_extension_cpu.cpp
+++ sgl-kernel/csrc/cpu/torch_extension_cpu.cpp
@@ -123,6 +123,14 @@
// gemm
at::Tensor
weight_packed_linear(at::Tensor& mat1, at::Tensor& mat2, const std::optional<at::Tensor>& bias, bool is_vnni);
+
+// gemm fusion
+at::Tensor fused_linear_sigmoid_mul(
+ at::Tenso... | 13 | 0 | 3 | [
"qwen"
] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/torch_extension_cpu.cpp).
Goal: [CPU] Optimize small oc GEMM for Qwen3-next on CPU
Target hardware: CPU
Target models: qwen
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "Lic... | This PR adds optimization for BF16 GEMMs with small oc shapes ( oc < 16), including:
1. `weight_packed_linear`， which is common for all GEMMs with oc < 16
2. `fused_linear_sigmoid_mul`, which is specific for cases found in Qwen3-next (shared_expert_gate, where N = 1 and having post op sigmoid+mul)
| [] | null | null |
sglang | 46f21a5956f85462208ff005e6cbc86f142f8046 | 2025-12-04T12:34:05+08:00 | 12,316 | https://github.com/sgl-project/sglang/pull/12316 | use faster covnersion from float8_e4m3fn to bfloat16 (#12316) | use faster covnersion from float8_e4m3fn to bfloat16 (#12316) | sgl-kernel/csrc/cpu/gemm_fp8.cpp | cpp | Apache-2.0 | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
template <typename scalar_t>
inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ input, int64_t size) {
using bVec = at::vec::Vectorized<scalar_t>;
using fVec = at::vec::Vectorized<float>;
constexpr int kVecSize = bVec::s... | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
template <typename scalar_t>
inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ input, int64_t size) {
using bVec = at::vec::Vectorized<scalar_t>;
using fVec = at::vec::Vectorized<float>;
constexpr int kVecSize = bVec::s... | --- sgl-kernel/csrc/cpu/gemm_fp8.cpp
+++ sgl-kernel/csrc/cpu/gemm_fp8.cpp
@@ -56,7 +56,8 @@
const int K2 = K >> 1;
const int ldb2 = ldb; // ldb * 2 >> 1;
const uint16_t* b_ptr = reinterpret_cast<const uint16_t*>(packed_B);
- const __m512 vd = _mm512_set1_ps(scale);
+ const __m512 vexp = _mm512_castsi512_ps(... | 9 | 5 | 2 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/gemm_fp8.cpp).
Goal: use faster covnersion from float8_e4m3fn to bfloat16
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
template <typename scalar_t>
inline void co... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
improve float8 gemm performance. the old impl... | [{"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": null, "metric": "performance", "unit": null, "higher_is_better": true, "baseline": null, "optimized": null, "stated_speedup": 1.11, "speedup": 1.11}] | 1.11 | null |
sglang | 46f21a5956f85462208ff005e6cbc86f142f8046 | 2025-12-04T12:34:05+08:00 | 12,316 | https://github.com/sgl-project/sglang/pull/12316 | use faster covnersion from float8_e4m3fn to bfloat16 (#12316) | use faster covnersion from float8_e4m3fn to bfloat16 (#12316) | sgl-kernel/csrc/cpu/vec.h | cpp | Apache-2.0 | #pragma once
#if defined(__AVX512F__) && defined(__AVX512BF16__) && defined(__AMX_BF16__)
#define CPU_CAPABILITY_AVX512
#endif
#include <ATen/cpu/vec/functional.h>
#include <ATen/cpu/vec/vec.h>
namespace {
using namespace at::vec;
template <typename scalar_t, typename std::enable_if_t<is_reduced_floating_point_v<s... | #pragma once
#if defined(__AVX512F__) && defined(__AVX512BF16__) && defined(__AMX_BF16__)
#define CPU_CAPABILITY_AVX512
#endif
#include <ATen/cpu/vec/functional.h>
#include <ATen/cpu/vec/vec.h>
namespace {
using namespace at::vec;
template <typename scalar_t, typename std::enable_if_t<is_reduced_floating_point_v<s... | --- sgl-kernel/csrc/cpu/vec.h
+++ sgl-kernel/csrc/cpu/vec.h
@@ -121,6 +121,30 @@
#endif
}
+// faster version of float8_e4m3fn conversion to bfloat16
+//
+// we mapped cuda implementation from below link and vectorized with avx512:
+// https://github.com/thu-pacman/chitu/blob/1ed2078ec26581ebdca05b7306d4385f86edaa7c... | 24 | 0 | 2 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/vec.h).
Goal: use faster covnersion from float8_e4m3fn to bfloat16
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#pragma once
#if defined(__AVX512F__) && defined(__AVX512BF16__) && defined(__AMX_BF16__)
#define CPU_CAPABILITY_AVX51... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
improve float8 gemm performance. the old impl... | [{"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": null, "metric": "performance", "unit": null, "higher_is_better": true, "baseline": null, "optimized": null, "stated_speedup": 1.11, "speedup": 1.11}] | 1.11 | null |
sglang | 043f13171fb9688b21fc4fa076c57e80cf83c89f | 2025-12-02T18:53:06-08:00 | 13,812 | https://github.com/sgl-project/sglang/pull/13812 | [Performance] Optimize NSA Indexer K/S Buffer Access with Fused Triton Kernels (#13812) | [Performance] Optimize NSA Indexer K/S Buffer Access with Fused Triton Kernels (#13812)
Co-authored-by: Johnsonms <johnson@together.ai> | python/sglang/srt/layers/attention/nsa/index_buf_accessor.py | triton | Apache-2.0 | from typing import TYPE_CHECKING
import torch
import triton
import triton.language as tl
if TYPE_CHECKING:
from sglang.srt.mem_cache.memory_pool import NSATokenToKVPool
"""
k: data, 128 item per token, fp8
s: scale, 1 item per token, fp32
"""
class GetK:
@classmethod
def execute(cls, *args, **kwargs):
... | from typing import TYPE_CHECKING
import torch
import triton
import triton.language as tl
if TYPE_CHECKING:
from sglang.srt.mem_cache.memory_pool import NSATokenToKVPool
"""
k: data, 128 item per token, fp8
s: scale, 1 item per token, fp32
"""
class GetK:
@classmethod
def execute(cls, *args, **kwargs):
... | --- python/sglang/srt/layers/attention/nsa/index_buf_accessor.py
+++ python/sglang/srt/layers/attention/nsa/index_buf_accessor.py
@@ -16,7 +16,7 @@
class GetK:
@classmethod
def execute(cls, *args, **kwargs):
- return cls.torch_fast(*args, **kwargs)
+ return cls.triton(*args, **kwargs)
@c... | 318 | 2 | 2 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/nsa/index_buf_accessor.py).
Goal: [Performance] Optimize NSA Indexer K/S Buffer Access with Fused Triton Kernels
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from typing import TYPE_CHECKING
import tor... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
https://github.com/sgl-project/sglang/issues/1... | [{"scope": "e2e", "hardware": null, "model": null, "shape": null, "batch_size": null, "metric": "CPU time", "unit": "us", "higher_is_better": false, "baseline": 166, "optimized": 63, "stated_speedup": null, "speedup": 2.6349}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": null, "metr... | 7 | 2.6349 |
sglang | 432ecf841e261b6dcaacdf250a8691390b86558e | 2025-11-25T19:44:24+08:00 | 12,078 | https://github.com/sgl-project/sglang/pull/12078 | [Ascend] qwen optimization (#12078) | [Ascend] qwen optimization (#12078)
Co-authored-by: Even Zhou <even.y.zhou@outlook.com> | python/sglang/srt/layers/rotary_embedding.py | triton | Apache-2.0 | # Adapted from https://raw.githubusercontent.com/vllm-project/vllm/refs/tags/v0.6.6.post1/vllm/model_executor/layers/rotary_embedding.py
"""Rotary Positional Embeddings."""
import itertools
import math
from typing import Any, Dict, List, Optional, Tuple, Union
import torch
import torch.nn as nn
import triton
import t... | # Adapted from https://raw.githubusercontent.com/vllm-project/vllm/refs/tags/v0.6.6.post1/vllm/model_executor/layers/rotary_embedding.py
"""Rotary Positional Embeddings."""
import itertools
import math
from typing import Any, Dict, List, Optional, Tuple, Union
import torch
import torch.nn as nn
import triton
import t... | --- python/sglang/srt/layers/rotary_embedding.py
+++ python/sglang/srt/layers/rotary_embedding.py
@@ -135,6 +135,7 @@
self._apply_rotary_emb_wrapped = torch.compile(dynamic=True)(
self._apply_rotary_emb_wrapped
)
+ self.position_cos, self.position_sin = None, None
... | 13 | 0 | 8 | [
"qwen"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/rotary_embedding.py).
Goal: [Ascend] qwen optimization
Target hardware: unspecified
Target models: qwen
Current file:
```python
# Adapted from https://raw.githubusercontent.com/vllm-project/vllm/refs/tags/v0.6.6.post1/vllm/model_executor/layers/ro... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
related to #10337
## Modifications
-bugfix:
... | [] | null | null |
sglang | 9535015d055a1d0a3b1f5fc336d9b5da44565944 | 2025-11-25T03:05:38+08:00 | 10,027 | https://github.com/sgl-project/sglang/pull/10027 | [Perf] Optimize DeepSeek-R1 w4afp8 glue kernels (#10027) | [Perf] Optimize DeepSeek-R1 w4afp8 glue kernels (#10027)
Co-authored-by: Fan Yin <1106310035@qq.com> | python/sglang/srt/layers/moe/ep_moe/kernels.py | triton | Apache-2.0 | import logging
import torch
import triton
from sglang.srt.utils import ceil_div, is_cuda
logger = logging.getLogger(__name__)
_is_cuda = is_cuda()
if _is_cuda:
from sglang.srt.layers.quantization.fp8_kernel import (
sglang_per_token_group_quant_fp8 as per_token_group_quant_fp8,
)
import triton.lang... | import logging
import torch
import triton
from sglang.srt.utils import ceil_div, is_cuda
logger = logging.getLogger(__name__)
_is_cuda = is_cuda()
if _is_cuda:
from sglang.srt.layers.quantization.fp8_kernel import (
sglang_per_token_group_quant_fp8 as per_token_group_quant_fp8,
)
import triton.lang... | --- python/sglang/srt/layers/moe/ep_moe/kernels.py
+++ python/sglang/srt/layers/moe/ep_moe/kernels.py
@@ -14,6 +14,60 @@
)
import triton.language as tl
+
+
+def _get_launch_config_1d(device, numel):
+ MAX_THREADS_PER_BLOCK = 1024
+ MIN_THREADS_PER_BLOCK = 512
+ MAX_WAVES = 8 # empirical numbers
+
+ ... | 239 | 66 | 3 | [
"deepseek"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/ep_moe/kernels.py).
Goal: [Perf] Optimize DeepSeek-R1 w4afp8 glue kernels
Target hardware: unspecified
Target models: deepseek
Current file:
```python
import logging
import torch
import triton
from sglang.srt.utils import ceil_div, is_cuda
... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [{"scope": "e2e", "hardware": "H800", "model": "DeepSeek-R1-W4AFP8", "shape": "ISL=10240,OSL=512", "batch_size": 4, "metric": "E2E Latency", "unit": "ms", "higher_is_better": false, "baseline": 8321.31, "optimized": 6815.18, "stated_speedup": null, "speedup": 1.221}, {"scope": "e2e", "hardware": "H800", "model": "DeepS... | null | 1.1161 |
sglang | 3451fc3280892c6555705e24abea8cb047825990 | 2025-11-01T19:47:56-07:00 | 11,133 | https://github.com/sgl-project/sglang/pull/11133 | [Feature] Qwen3-Next & FLA: Support MTP topk>1; Up to 6% faster (#11133) | [Feature] Qwen3-Next & FLA: Support MTP topk>1; Up to 6% faster (#11133)
Co-authored-by: Stefan He <hebiaobuaa@gmail.com> | python/sglang/srt/layers/attention/fla/fused_recurrent.py | triton | Apache-2.0 | # Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/gated_delta_rule/fused_recurrent.py
# -*- coding: utf-8 -*-
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
from typing import Optional, Tuple
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.fl... | # Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/gated_delta_rule/fused_recurrent.py
# -*- coding: utf-8 -*-
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
from typing import Optional, Tuple
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.fl... | --- python/sglang/srt/layers/attention/fla/fused_recurrent.py
+++ python/sglang/srt/layers/attention/fla/fused_recurrent.py
@@ -330,11 +330,29 @@
return o, final_state
+# HAS_EAGLE_TREE_CUSTOM_ATTN_MASK is added to support eagle tree attention mask
+# retrieve_parent_token_ptr: [N, NP2_T], retrieve_next_siblin... | 70 | 0 | 4 | [
"qwen"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/fla/fused_recurrent.py).
Goal: [Feature] Qwen3-Next & FLA: Support MTP topk>1; Up to 6% faster
Target hardware: unspecified
Target models: qwen
Current file:
```python
# Adapt from https://github.com/fla-org/flash-linear-attention/blob/m... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
Co-authored-by: @hebiao064
## Motivation
MTP spec decodin... | [{"scope": "e2e", "hardware": "H100", "model": "Qwen3-Next-80B-A3B-Instruct", "shape": null, "batch_size": null, "metric": "throughput", "unit": "tok/s", "higher_is_better": true, "baseline": 343.33, "optimized": 349.29, "stated_speedup": null, "speedup": 1.0174}] | null | 1.0174 |
sglang | 3451fc3280892c6555705e24abea8cb047825990 | 2025-11-01T19:47:56-07:00 | 11,133 | https://github.com/sgl-project/sglang/pull/11133 | [Feature] Qwen3-Next & FLA: Support MTP topk>1; Up to 6% faster (#11133) | [Feature] Qwen3-Next & FLA: Support MTP topk>1; Up to 6% faster (#11133)
Co-authored-by: Stefan He <hebiaobuaa@gmail.com> | python/sglang/srt/layers/attention/mamba/causal_conv1d_triton.py | triton | Apache-2.0 | # Copyright (c) 2024, Tri Dao.
# Adapted from https://github.com/Dao-AILab/causal-conv1d/blob/main/causal_conv1d/causal_conv1d_interface.py
# and https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/mamba/ops/causal_conv1d.py
from typing import List, Optional, Union
import torch
import triton
imp... | # Copyright (c) 2024, Tri Dao.
# Adapted from https://github.com/Dao-AILab/causal-conv1d/blob/main/causal_conv1d/causal_conv1d_interface.py
# and https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/mamba/ops/causal_conv1d.py
from typing import List, Optional, Union
import torch
import triton
imp... | --- python/sglang/srt/layers/attention/mamba/causal_conv1d_triton.py
+++ python/sglang/srt/layers/attention/mamba/causal_conv1d_triton.py
@@ -186,7 +186,7 @@
)
mask = (idx_tokens_conv < state_len)[:, None] & (idx_feats < dim)[None, :]
- tl.debug_barrier() # NOTE: use this due to... | 251 | 54 | 4 | [
"qwen",
"mamba"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/mamba/causal_conv1d_triton.py).
Goal: [Feature] Qwen3-Next & FLA: Support MTP topk>1; Up to 6% faster
Target hardware: unspecified
Target models: qwen, mamba
Current file:
```python
# Copyright (c) 2024, Tri Dao.
# Adapted from https://g... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
Co-authored-by: @hebiao064
## Motivation
MTP spec decodin... | [{"scope": "e2e", "hardware": "H100", "model": "Qwen3-Next-80B-A3B-Instruct", "shape": null, "batch_size": null, "metric": "throughput", "unit": "tok/s", "higher_is_better": true, "baseline": 343.33, "optimized": 349.29, "stated_speedup": null, "speedup": 1.0174}] | null | 1.0174 |
sglang | d5b6e50fe8148a851eb9170ced0208132596c615 | 2025-10-31T22:48:02-07:00 | 12,435 | https://github.com/sgl-project/sglang/pull/12435 | perf: trtllm mla performance minor improvements (#12435) | perf: trtllm mla performance minor improvements (#12435) | python/sglang/srt/layers/attention/trtllm_mla_backend.py | triton | Apache-2.0 | from __future__ import annotations
"""
Support attention backend for TRTLLM MLA kernels from flashinfer.
"""
import math
from dataclasses import dataclass
from typing import TYPE_CHECKING, Optional, Union
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.flashinfer_mla_backend... | from __future__ import annotations
"""
Support attention backend for TRTLLM MLA kernels from flashinfer.
"""
import math
from dataclasses import dataclass
from typing import TYPE_CHECKING, Optional, Union
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.flashinfer_mla_backend... | --- python/sglang/srt/layers/attention/trtllm_mla_backend.py
+++ python/sglang/srt/layers/attention/trtllm_mla_backend.py
@@ -219,6 +219,7 @@
sum_seq_lens_q: Optional[int] = None
cu_seqlens_q: Optional[torch.Tensor] = None
seq_lens_q: Optional[torch.Tensor] = None
+ seq_lens_k: Optional[torch.Tensor] ... | 70 | 65 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/trtllm_mla_backend.py).
Goal: perf: trtllm mla performance minor improvements
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from __future__ import annotations
"""
Support attention backend for TRTLLM ML... | ## Motivation
The current `trtllm_mla` backend has several minor inefficiencies and confusing parts:
1. Element-wise operations on `seq_lens` are repeatedly executed inside CUDA Graph replay, instead of being done once during metadata initialization.
2. A redundant `torch.cat` operation (`k` + `k_rope`) exists in `... | [{"scope": "e2e", "hardware": null, "model": "nvidia/DeepSeek-V3-0324-FP4", "shape": null, "batch_size": null, "metric": "latency", "unit": "s", "higher_is_better": false, "baseline": 1.939, "optimized": 1.895, "stated_speedup": null, "speedup": 1.0232}, {"scope": "e2e", "hardware": null, "model": "nvidia/DeepSeek-V3-0... | null | 1.0232 |
sglang | cafebef1546e9c869672b9fe12a89f3a34a79f86 | 2025-10-30T21:52:16+08:00 | 11,969 | https://github.com/sgl-project/sglang/pull/11969 | [NPU] bugfix for Qwen3-Next and performance update (#11969) | [NPU] bugfix for Qwen3-Next and performance update (#11969) | python/sglang/srt/layers/attention/fla/layernorm_gated.py | triton | Apache-2.0 | # Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/modules/layernorm_gated.py
# Copyright (c) 2024, Tri Dao.
# Based on the Triton LayerNorm tutorial: https://triton-lang.org/main/getting-started/tutorials/05-layer-norm.html
# For the backward pass, we keep weight_grad and bias_grad in registe... | # Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/modules/layernorm_gated.py
# Copyright (c) 2024, Tri Dao.
# Based on the Triton LayerNorm tutorial: https://triton-lang.org/main/getting-started/tutorials/05-layer-norm.html
# For the backward pass, we keep weight_grad and bias_grad in registe... | --- python/sglang/srt/layers/attention/fla/layernorm_gated.py
+++ python/sglang/srt/layers/attention/fla/layernorm_gated.py
@@ -12,7 +12,9 @@
import triton.language as tl
from einops import rearrange
-from sglang.srt.utils import device_context
+from sglang.srt.utils import device_context, is_npu
+
+_is_npu = is_np... | 7 | 1 | 3 | [
"qwen"
] | [
"NPU"
] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/fla/layernorm_gated.py).
Goal: [NPU] bugfix for Qwen3-Next and performance update
Target hardware: NPU
Target models: qwen
Current file:
```python
# Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/modules/layer... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [] | null | null |
sglang | 13bf565d60f894acf6a496591c2f2fc0eecc4a98 | 2025-10-25T08:41:16+08:00 | 8,464 | https://github.com/sgl-project/sglang/pull/8464 | [2/N]Support DeepSeek-R1 w4a8 low latency deepep (#8464) | [2/N]Support DeepSeek-R1 w4a8 low latency deepep (#8464)
Co-authored-by: Hank Han <hanhan7630@outlook.com>
Co-authored-by: Shangchuan Huang <2510421000@qq.com> | python/sglang/srt/layers/moe/ep_moe/kernels.py | triton | Apache-2.0 | import logging
import torch
import triton
from sglang.srt.utils import ceil_div, is_cuda
logger = logging.getLogger(__name__)
_is_cuda = is_cuda()
if _is_cuda:
from sglang.srt.layers.quantization.fp8_kernel import (
sglang_per_token_group_quant_fp8 as per_token_group_quant_fp8,
)
import triton.lang... | import logging
import torch
import triton
from sglang.srt.utils import ceil_div, is_cuda
logger = logging.getLogger(__name__)
_is_cuda = is_cuda()
if _is_cuda:
from sglang.srt.layers.quantization.fp8_kernel import (
sglang_per_token_group_quant_fp8 as per_token_group_quant_fp8,
)
import triton.lang... | --- python/sglang/srt/layers/moe/ep_moe/kernels.py
+++ python/sglang/srt/layers/moe/ep_moe/kernels.py
@@ -1014,3 +1014,197 @@
)
return output
+
+
+@triton.jit
+def compute_problem_sizes_w4a8_kernel(
+ masked_m_ptr,
+ problem_sizes1_ptr,
+ problem_sizes2_ptr,
+ n,
+ k,
+ num_experts,
+ B... | 194 | 0 | 6 | [
"deepseek"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/ep_moe/kernels.py).
Goal: [2/N]Support DeepSeek-R1 w4a8 low latency deepep
Target hardware: unspecified
Target models: deepseek
Current file:
```python
import logging
import torch
import triton
from sglang.srt.utils import ceil_div, is_cuda
... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Follow https://github.com/sgl-project/sglang/pull/82... | [] | null | null |
sglang | 13bf565d60f894acf6a496591c2f2fc0eecc4a98 | 2025-10-25T08:41:16+08:00 | 8,464 | https://github.com/sgl-project/sglang/pull/8464 | [2/N]Support DeepSeek-R1 w4a8 low latency deepep (#8464) | [2/N]Support DeepSeek-R1 w4a8 low latency deepep (#8464)
Co-authored-by: Hank Han <hanhan7630@outlook.com>
Co-authored-by: Shangchuan Huang <2510421000@qq.com> | sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_get_group_starts.cuh | cuda | Apache-2.0 | #pragma once
#include <c10/cuda/CUDAStream.h>
#include <cuda.h>
#include <torch/all.h>
#include "cutlass/bfloat16.h"
#include "cutlass/float8.h"
template <typename ElementA, typename ElementB, typename ElementC, typename ElementAccumulator>
__global__ void int4_fp8_get_group_gemm_starts(
int32_t* expert_offsets,... | #pragma once
#include <c10/cuda/CUDAStream.h>
#include <cuda.h>
#include <torch/all.h>
#include "cutlass/bfloat16.h"
#include "cutlass/float8.h"
template <typename ElementA, typename ElementB, typename ElementC, typename ElementAccumulator>
__global__ void int4_fp8_get_group_gemm_starts(
int32_t* expert_offsets,... | --- sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_get_group_starts.cuh
+++ sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_get_group_starts.cuh
@@ -34,6 +34,40 @@
b_scales_offsets[expert_id] = b_scales_base_as_int + (per_out_ch ? expert_id * n * k / 128 : expert_id);
}
+template <typename ElementA, typename ElementB, typen... | 72 | 6 | 6 | [
"deepseek"
] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_get_group_starts.cuh).
Goal: [2/N]Support DeepSeek-R1 w4a8 low latency deepep
Target hardware: unspecified
Target models: deepseek
Current file:
```cuda
#pragma once
#include <c10/cuda/CUDAStream.h>
#include <cuda.h>
#include <torc... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Follow https://github.com/sgl-project/sglang/pull/82... | [] | null | null |
sglang | 13bf565d60f894acf6a496591c2f2fc0eecc4a98 | 2025-10-25T08:41:16+08:00 | 8,464 | https://github.com/sgl-project/sglang/pull/8464 | [2/N]Support DeepSeek-R1 w4a8 low latency deepep (#8464) | [2/N]Support DeepSeek-R1 w4a8 low latency deepep (#8464)
Co-authored-by: Hank Han <hanhan7630@outlook.com>
Co-authored-by: Shangchuan Huang <2510421000@qq.com> | sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_grouped_mm_c3x.cuh | cuda | Apache-2.0 | #pragma once
/**
* @file w4a8_grouped_mm_c3x.cuh
* @brief Implementation of grouped GEMM operation with int4 and fp8 mixed
* precision
*
* This file implements a grouped GEMM operation that multiplies FP8 matrices
* (A) with quantized INT4 matrices (B), applying per-block scaling factors.
* The implementation i... | #pragma once
/**
* @file w4a8_grouped_mm_c3x.cuh
* @brief Implementation of grouped GEMM operation with int4 and fp8 mixed
* precision
*
* This file implements a grouped GEMM operation that multiplies FP8 matrices
* (A) with quantized INT4 matrices (B), applying per-block scaling factors.
* The implementation i... | --- sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_grouped_mm_c3x.cuh
+++ sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_grouped_mm_c3x.cuh
@@ -174,7 +174,7 @@
bool per_out_ch = b_scales.numel() != num_experts;
// Check inputs
- TORCH_CHECK(a_tensors.dim() == 2, "A tensor must be 2D");
+ TORCH_CHECK(a_tensors.dim() == ... | 4 | 2 | 6 | [
"deepseek"
] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_grouped_mm_c3x.cuh).
Goal: [2/N]Support DeepSeek-R1 w4a8 low latency deepep
Target hardware: unspecified
Target models: deepseek
Current file:
```cuda
#pragma once
/**
* @file w4a8_grouped_mm_c3x.cuh
* @brief Implementation of gr... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Follow https://github.com/sgl-project/sglang/pull/82... | [] | null | null |
sglang | e51046beaa67c3dd39f1814488ffc147ce5e740d | 2025-10-24T16:05:44-07:00 | 12,093 | https://github.com/sgl-project/sglang/pull/12093 | perf: trtllm_mla attention backend spec decoding speedup w/ cuda graph (#12093) | perf: trtllm_mla attention backend spec decoding speedup w/ cuda graph (#12093) | python/sglang/srt/layers/attention/trtllm_mla_backend.py | triton | Apache-2.0 | from __future__ import annotations
"""
Support attention backend for TRTLLM MLA kernels from flashinfer.
"""
import math
from dataclasses import dataclass
from typing import TYPE_CHECKING, Optional, Union
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.flashinfer_mla_backend... | from __future__ import annotations
"""
Support attention backend for TRTLLM MLA kernels from flashinfer.
"""
import math
from dataclasses import dataclass
from typing import TYPE_CHECKING, Optional, Union
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.flashinfer_mla_backend... | --- python/sglang/srt/layers/attention/trtllm_mla_backend.py
+++ python/sglang/srt/layers/attention/trtllm_mla_backend.py
@@ -423,14 +423,9 @@
PAGED_SIZE=self.page_size,
)
- # Record the true maximum sequence length for this capture batch so that
- # the kernel launch path (which r... | 1 | 13 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/trtllm_mla_backend.py).
Goal: perf: trtllm_mla attention backend spec decoding speedup w/ cuda graph
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from __future__ import annotations
"""
Support attentio... | ## Motivation
The `trtllm_mla` backend was selecting kernels during CUDA graph capture using very small sequence lengths.
This often led to inefficient kernel choices in most runtime cases, degrading overall performance.
## Modifications
Now the kernel selection uses `max_context_len` instead of the real input seque... | [] | null | null |
sglang | 13fb8b5489447c7e70a3e9b13793bc5bceda9c49 | 2025-10-23T12:39:51+08:00 | 10,652 | https://github.com/sgl-project/sglang/pull/10652 | [CPU] Optimize FP16 decode_attention_cpu (#10652) | [CPU] Optimize FP16 decode_attention_cpu (#10652) | sgl-kernel/csrc/cpu/decode.cpp | cpp | Apache-2.0 | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
// [NOTE] TODO list for this kernel:
// 1. tune the value for BLOCK_N
// 2. planning for {batches, num_heads, num_kv_splits}
// and use actual num_kv_splits for small seq length
// 3. try fast impl of `.tanh()`
// 4. provide amx kernel fo... | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
// [NOTE] TODO list for this kernel:
// 1. tune the value for BLOCK_N
// 2. planning for {batches, num_heads, num_kv_splits}
// and use actual num_kv_splits for small seq length
// 3. try fast impl of `.tanh()`
// 4. provide amx kernel fo... | --- sgl-kernel/csrc/cpu/decode.cpp
+++ sgl-kernel/csrc/cpu/decode.cpp
@@ -308,6 +308,93 @@
};
#endif
+#if defined(CPU_CAPABILITY_AVX512)
+template <typename index_t, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nt<at::Half, index_t, BLOCK_M, BLOCK_N> {
+ static inline void apply(
+ const at::Half* __rest... | 171 | 2 | 3 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/decode.cpp).
Goal: [CPU] Optimize FP16 decode_attention_cpu
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
// [NOTE] TODO list for this kernel:
// 1. tune the val... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [] | null | null |
sglang | 13fb8b5489447c7e70a3e9b13793bc5bceda9c49 | 2025-10-23T12:39:51+08:00 | 10,652 | https://github.com/sgl-project/sglang/pull/10652 | [CPU] Optimize FP16 decode_attention_cpu (#10652) | [CPU] Optimize FP16 decode_attention_cpu (#10652) | sgl-kernel/csrc/cpu/vec.h | cpp | Apache-2.0 | #pragma once
#if defined(__AVX512F__) && defined(__AVX512BF16__) && defined(__AMX_BF16__)
#define CPU_CAPABILITY_AVX512
#endif
#include <ATen/cpu/vec/functional.h>
#include <ATen/cpu/vec/vec.h>
namespace {
using namespace at::vec;
template <typename scalar_t, typename std::enable_if_t<is_reduced_floating_point_v<s... | #pragma once
#if defined(__AVX512F__) && defined(__AVX512BF16__) && defined(__AMX_BF16__)
#define CPU_CAPABILITY_AVX512
#endif
#include <ATen/cpu/vec/functional.h>
#include <ATen/cpu/vec/vec.h>
namespace {
using namespace at::vec;
template <typename scalar_t, typename std::enable_if_t<is_reduced_floating_point_v<s... | --- sgl-kernel/csrc/cpu/vec.h
+++ sgl-kernel/csrc/cpu/vec.h
@@ -47,7 +47,7 @@
#define CVT_BF16_TO_FP32(a) _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(a), 16))
-#define CVT_FP16_TO_FP32(a) _mm512_cvtps_ph(a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC))
+#define CVT_FP16_TO_FP32(a) _mm512_cvtph_p... | 1 | 1 | 3 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/vec.h).
Goal: [CPU] Optimize FP16 decode_attention_cpu
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#pragma once
#if defined(__AVX512F__) && defined(__AVX512BF16__) && defined(__AMX_BF16__)
#define CPU_CAPABILITY_AVX512
#endif
#i... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [] | null | null |
sglang | 5a33c3aae72ae0c356bc113b328aa3a904a3365d | 2025-10-14T20:08:32+08:00 | 11,556 | https://github.com/sgl-project/sglang/pull/11556 | Optimize Triton Draft Backend (#11556) | Optimize Triton Draft Backend (#11556) | python/sglang/srt/layers/attention/triton_backend.py | triton | Apache-2.0 | from __future__ import annotations
from dataclasses import dataclass
from typing import TYPE_CHECKING, Optional, Union
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
from sglang.srt.layers.attention.utils import create_flashinfer_kv_... | from __future__ import annotations
from dataclasses import dataclass
from typing import TYPE_CHECKING, List, Optional
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
from sglang.srt.layers.attention.utils import create_flashinfer_kv_i... | --- python/sglang/srt/layers/attention/triton_backend.py
+++ python/sglang/srt/layers/attention/triton_backend.py
@@ -1,7 +1,7 @@
from __future__ import annotations
from dataclasses import dataclass
-from typing import TYPE_CHECKING, Optional, Union
+from typing import TYPE_CHECKING, List, Optional
import torch
... | 51 | 28 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/triton_backend.py).
Goal: Optimize Triton Draft Backend
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from __future__ import annotations
from dataclasses import dataclass
from typing import TYPE_CHECKIN... | This PR:
- Reduces the redundant copy when doing the multi-step decode (draft) forward.
- Introduce the `num_kv_splits_buf` for Triton backend, and only init once when replaying cuda graph.
| [] | null | null |
sglang | 21337b22b960c8bca5410c73ba4f36bfca72443c | 2025-10-11T15:59:03+08:00 | 10,312 | https://github.com/sgl-project/sglang/pull/10312 | Reland [1/2] Optimizations and refactors about quant kernel (#10312) | Reland [1/2] Optimizations and refactors about quant kernel (#10312)
Co-authored-by: Yineng Zhang <me@zhyncs.com> | python/sglang/srt/layers/quantization/fp8_kernel.py | triton | Apache-2.0 | # Copyright 2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, so... | # Copyright 2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, so... | --- python/sglang/srt/layers/quantization/fp8_kernel.py
+++ python/sglang/srt/layers/quantization/fp8_kernel.py
@@ -43,11 +43,17 @@
_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
if _is_cuda:
- from sgl_kernel import (
- sgl_per_tensor_quant_fp8,
- sgl_per_token_group_quant_fp8,
- ... | 35 | 8 | 7 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/quantization/fp8_kernel.py).
Goal: Reland [1/2] Optimizations and refactors about quant kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2024 SGLang Team
# Licensed under the Apache License, Versio... | ## Motivation
real change is tiny, most code are from the last pr which has been reviewed
before fix: either error or 0% acc
Blackwell:
```
(cd /data/numa0/tom/primary_synced/sglang && SGL_JIT_DEEPGEMM_PRECOMPILE=0 SGLANG_IS_IN_CI=1 HF_TOKEN=... HF_HUB_ENABLE_HF_TRANSFER=1 python3 test/srt/ep/test_deepep_small.py... | [] | null | null |
sglang | 21337b22b960c8bca5410c73ba4f36bfca72443c | 2025-10-11T15:59:03+08:00 | 10,312 | https://github.com/sgl-project/sglang/pull/10312 | Reland [1/2] Optimizations and refactors about quant kernel (#10312) | Reland [1/2] Optimizations and refactors about quant kernel (#10312)
Co-authored-by: Yineng Zhang <me@zhyncs.com> | python/sglang/srt/layers/quantization/int8_kernel.py | triton | Apache-2.0 | import functools
import json
import logging
import os
from typing import Any, Dict, List, Optional, Tuple
import torch
import triton
import triton.language as tl
from sglang.srt.utils import get_device_name, is_cuda
_is_cuda = is_cuda()
if _is_cuda:
from sgl_kernel import sgl_per_token_group_quant_int8
logger =... | import functools
import json
import logging
import os
from typing import Any, Dict, List, Optional, Tuple
import torch
import triton
import triton.language as tl
from sglang.srt.utils import get_bool_env_var, get_device_name, is_cuda
_is_cuda = is_cuda()
if _is_cuda:
# Temporary
try:
from sgl_kernel ... | --- python/sglang/srt/layers/quantization/int8_kernel.py
+++ python/sglang/srt/layers/quantization/int8_kernel.py
@@ -8,11 +8,17 @@
import triton
import triton.language as tl
-from sglang.srt.utils import get_device_name, is_cuda
+from sglang.srt.utils import get_bool_env_var, get_device_name, is_cuda
_is_cuda =... | 12 | 3 | 7 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/quantization/int8_kernel.py).
Goal: Reland [1/2] Optimizations and refactors about quant kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```python
import functools
import json
import logging
import os
from typing impor... | ## Motivation
real change is tiny, most code are from the last pr which has been reviewed
before fix: either error or 0% acc
Blackwell:
```
(cd /data/numa0/tom/primary_synced/sglang && SGL_JIT_DEEPGEMM_PRECOMPILE=0 SGLANG_IS_IN_CI=1 HF_TOKEN=... HF_HUB_ENABLE_HF_TRANSFER=1 python3 test/srt/ep/test_deepep_small.py... | [] | null | null |
sglang | 21337b22b960c8bca5410c73ba4f36bfca72443c | 2025-10-11T15:59:03+08:00 | 10,312 | https://github.com/sgl-project/sglang/pull/10312 | Reland [1/2] Optimizations and refactors about quant kernel (#10312) | Reland [1/2] Optimizations and refactors about quant kernel (#10312)
Co-authored-by: Yineng Zhang <me@zhyncs.com> | sgl-kernel/csrc/common_extension.cc | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/common_extension.cc
+++ sgl-kernel/csrc/common_extension.cc
@@ -136,14 +136,14 @@
m.impl("fp8_blockwise_scaled_mm", torch::kCUDA, &fp8_blockwise_scaled_mm);
m.def(
- "sgl_per_token_group_quant_fp8(Tensor input, Tensor output_q, Tensor output_s, int group_size,"
+ "sgl_per_token_gro... | 7 | 7 | 7 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/common_extension.cc).
Goal: Reland [1/2] Optimizations and refactors about quant kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Versio... | ## Motivation
real change is tiny, most code are from the last pr which has been reviewed
before fix: either error or 0% acc
Blackwell:
```
(cd /data/numa0/tom/primary_synced/sglang && SGL_JIT_DEEPGEMM_PRECOMPILE=0 SGLANG_IS_IN_CI=1 HF_TOKEN=... HF_HUB_ENABLE_HF_TRANSFER=1 python3 test/srt/ep/test_deepep_small.py... | [] | null | null |
sglang | 21337b22b960c8bca5410c73ba4f36bfca72443c | 2025-10-11T15:59:03+08:00 | 10,312 | https://github.com/sgl-project/sglang/pull/10312 | Reland [1/2] Optimizations and refactors about quant kernel (#10312) | Reland [1/2] Optimizations and refactors about quant kernel (#10312)
Co-authored-by: Yineng Zhang <me@zhyncs.com> | sgl-kernel/csrc/gemm/per_token_group_quant_8bit.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <cuda_fp8.h>
#include <cmath>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
__device__ __forceinline__ float GroupReduceMax(float val, const int tid) {
unsigned mask = threadIdx.x % 32 >= 16 ? 0xffff0000 : 0x0000ffff;
val = fmaxf(val, __shfl_xor_sync(mask, v... | #include <ATen/cuda/CUDAContext.h>
#include <cuda_fp8.h>
#include <cmath>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
__device__ __forceinline__ float GroupReduceMax(float val, const int tid) {
unsigned mask = threadIdx.x % 32 >= 16 ? 0xffff0000 : 0x0000ffff;
val = fmaxf(val, __shfl_xor_sync(mask, v... | --- sgl-kernel/csrc/gemm/per_token_group_quant_8bit.cu
+++ sgl-kernel/csrc/gemm/per_token_group_quant_8bit.cu
@@ -121,7 +121,7 @@
double eps,
double min_8bit,
double max_8bit,
- bool scale_ue8m0 = false) {
+ bool scale_ue8m0) {
CHECK_INPUT(input);
CHECK_INPUT(output_q);
@@ -215,26 +215,3 @@... | 1 | 24 | 7 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/gemm/per_token_group_quant_8bit.cu).
Goal: Reland [1/2] Optimizations and refactors about quant kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#include <cuda_fp8.h>
#include <cmath>
#i... | ## Motivation
real change is tiny, most code are from the last pr which has been reviewed
before fix: either error or 0% acc
Blackwell:
```
(cd /data/numa0/tom/primary_synced/sglang && SGL_JIT_DEEPGEMM_PRECOMPILE=0 SGLANG_IS_IN_CI=1 HF_TOKEN=... HF_HUB_ENABLE_HF_TRANSFER=1 python3 test/srt/ep/test_deepep_small.py... | [] | null | null |
sglang | 21337b22b960c8bca5410c73ba4f36bfca72443c | 2025-10-11T15:59:03+08:00 | 10,312 | https://github.com/sgl-project/sglang/pull/10312 | Reland [1/2] Optimizations and refactors about quant kernel (#10312) | Reland [1/2] Optimizations and refactors about quant kernel (#10312)
Co-authored-by: Yineng Zhang <me@zhyncs.com> | sgl-kernel/include/sgl_kernel_ops.h | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/include/sgl_kernel_ops.h
+++ sgl-kernel/include/sgl_kernel_ops.h
@@ -219,7 +219,7 @@
const torch::Dtype& out_dtype);
void scaled_fp4_quant(
torch::Tensor& output, torch::Tensor const& input, torch::Tensor& output_scale, torch::Tensor const& input_scale);
-void sgl_per_token_group_quant_fp8(
+v... | 7 | 4 | 7 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/include/sgl_kernel_ops.h).
Goal: Reland [1/2] Optimizations and refactors about quant kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Versio... | ## Motivation
real change is tiny, most code are from the last pr which has been reviewed
before fix: either error or 0% acc
Blackwell:
```
(cd /data/numa0/tom/primary_synced/sglang && SGL_JIT_DEEPGEMM_PRECOMPILE=0 SGLANG_IS_IN_CI=1 HF_TOKEN=... HF_HUB_ENABLE_HF_TRANSFER=1 python3 test/srt/ep/test_deepep_small.py... | [] | null | null |
sglang | 42245551efd8d556692aa518a49e8f15cfa136a5 | 2025-09-28T23:04:22+08:00 | 10,543 | https://github.com/sgl-project/sglang/pull/10543 | [sgl-kernel] Optimize concat_mla_k kernel (#10543) | [sgl-kernel] Optimize concat_mla_k kernel (#10543)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
Co-authored-by: PGFLMG <1106310035@qq.com> | sgl-kernel/csrc/elementwise/concat_mla.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/CUDADataType.h>
#include <cuda_runtime.h>
#include "pytorch_extension_utils.h"
constexpr int NUM_LOCAL_HEADS = 128;
constexpr int QK_NOPE_HEAD_DIM = 128;
constexpr int QK_ROPE_HEAD_DIM = 64;
constexpr int K_HEAD_DIM = QK_NOPE_HEAD_DIM + QK_ROPE_HEAD_DIM;
constex... | #include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/CUDADataType.h>
#include <cuda_runtime.h>
#include "pytorch_extension_utils.h"
#include "utils.cuh"
constexpr int NUM_LOCAL_HEADS = 128;
constexpr int QK_NOPE_HEAD_DIM = 128;
constexpr int QK_ROPE_HEAD_DIM = 64;
constexpr int K_HEAD_DIM = QK_NOPE_HEAD_DIM + QK_RO... | --- sgl-kernel/csrc/elementwise/concat_mla.cu
+++ sgl-kernel/csrc/elementwise/concat_mla.cu
@@ -3,6 +3,7 @@
#include <cuda_runtime.h>
#include "pytorch_extension_utils.h"
+#include "utils.cuh"
constexpr int NUM_LOCAL_HEADS = 128;
constexpr int QK_NOPE_HEAD_DIM = 128;
@@ -12,20 +13,10 @@
constexpr int HEAD_CHUN... | 44 | 46 | 1 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/elementwise/concat_mla.cu).
Goal: [sgl-kernel] Optimize concat_mla_k kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/CUDADataType.h>
#include <cuda_runtime.h>
#inclu... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [{"scope": "e2e", "hardware": null, "model": "DeepSeek-R1", "shape": "input=4k, output=1", "batch_size": 16, "metric": "Input token throughput", "unit": "tok/s", "higher_is_better": true, "baseline": 9165.76, "optimized": 9327.91, "stated_speedup": null, "speedup": 1.0177}, {"scope": "e2e", "hardware": null, "model": "... | null | 1.0187 |
sglang | 0f04a5f428b27b4f011473787de93f3115e832a6 | 2025-09-22T08:04:27+08:00 | 10,714 | https://github.com/sgl-project/sglang/pull/10714 | Optimize cutlass int8 gemm kernel for large M on SM89 Ada GPU (#10714) | Optimize cutlass int8 gemm kernel for large M on SM89 Ada GPU (#10714) | sgl-kernel/csrc/gemm/int8_gemm_kernel.cu | cuda | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/gemm/int8_gemm_kernel.cu
+++ sgl-kernel/csrc/gemm/int8_gemm_kernel.cu
@@ -409,8 +409,8 @@
cutlass_int8_scaled_mm<
ElementOutput,
ArchTag,
- cutlass::gemm::GemmShape<32, 64, 128>,
- cutlass::gemm::GemmShape<16, 64, 64>,
+ cutlass::gemm::GemmShape<128, 128, ... | 2 | 2 | 1 | [] | [
"ADA",
"SM89"
] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/gemm/int8_gemm_kernel.cu).
Goal: Optimize cutlass int8 gemm kernel for large M on SM89 Ada GPU
Target hardware: ADA, SM89
Target models: unspecified
Current file:
```cuda
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License,... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
Inspired by [10502](https://github.com/sgl-pro... | [] | null | null |
sglang | bd7eb0205add75a33047221e3322e376d314697c | 2025-09-20T00:12:43-07:00 | 10,680 | https://github.com/sgl-project/sglang/pull/10680 | [Performance] Qwen3-Next: optimize causal_conv1d_fn triton kernel - up to 9% faster (#10680) | [Performance] Qwen3-Next: optimize causal_conv1d_fn triton kernel - up to 9% faster (#10680) | python/sglang/srt/layers/attention/mamba/causal_conv1d_triton.py | triton | Apache-2.0 | # Copyright (c) 2024, Tri Dao.
# Adapted from https://github.com/Dao-AILab/causal-conv1d/blob/main/causal_conv1d/causal_conv1d_interface.py
# and https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/mamba/ops/causal_conv1d.py
from typing import Optional, Union
import numpy as np
import torch
PAD... | # Copyright (c) 2024, Tri Dao.
# Adapted from https://github.com/Dao-AILab/causal-conv1d/blob/main/causal_conv1d/causal_conv1d_interface.py
# and https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/mamba/ops/causal_conv1d.py
from typing import List, Optional, Union
import numpy as np
import torc... | --- python/sglang/srt/layers/attention/mamba/causal_conv1d_triton.py
+++ python/sglang/srt/layers/attention/mamba/causal_conv1d_triton.py
@@ -2,7 +2,7 @@
# Adapted from https://github.com/Dao-AILab/causal-conv1d/blob/main/causal_conv1d/causal_conv1d_interface.py
# and https://github.com/vllm-project/vllm/blob/main/vl... | 15 | 98 | 3 | [
"qwen",
"mamba"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/mamba/causal_conv1d_triton.py).
Goal: [Performance] Qwen3-Next: optimize causal_conv1d_fn triton kernel - up to 9% faster
Target hardware: unspecified
Target models: qwen, mamba
Current file:
```python
# Copyright (c) 2024, Tri Dao.
# Ad... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
(This PR fixed the merge conflict in this PR:... | [{"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": null, "metric": "runtime", "unit": "us", "higher_is_better": false, "baseline": 313, "optimized": 95, "stated_speedup": null, "speedup": 3.2947}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": null, "me... | 4.6844 | 1.0546 |
sglang | 3b25dc127ae7df0680cfe0fd7800ba5aa280c64c | 2025-09-16T02:53:21+08:00 | 10,473 | https://github.com/sgl-project/sglang/pull/10473 | [1/2] Speed up trtllm_mla attention backend (>10% e2e) (#10473) | [1/2] Speed up trtllm_mla attention backend (>10% e2e) (#10473) | sgl-kernel/csrc/common_extension.cc | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/common_extension.cc
+++ sgl-kernel/csrc/common_extension.cc
@@ -103,6 +103,9 @@
m.impl("copy_to_gpu_no_ce", torch::kCUDA, ©_to_gpu_no_ce);
m.def("concat_mla_k(Tensor! k, Tensor k_nope, Tensor k_rope) -> ()");
m.impl("concat_mla_k", torch::kCUDA, &concat_mla_k);
+
+ m.def("concat_mla_ab... | 3 | 0 | 5 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/common_extension.cc).
Goal: [1/2] Speed up trtllm_mla attention backend (>10% e2e)
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
EDIT: 17.5% e2e speedup for short seq len, th... | [{"scope": "e2e", "hardware": null, "model": null, "shape": null, "batch_size": 1, "metric": "latency", "unit": "s", "higher_is_better": false, "baseline": 0.010048, "optimized": 0.006176, "stated_speedup": null, "speedup": 1.6269}, {"scope": "e2e", "hardware": null, "model": null, "shape": null, "batch_size": 16, "met... | null | 5.4031 |
sglang | 3b25dc127ae7df0680cfe0fd7800ba5aa280c64c | 2025-09-16T02:53:21+08:00 | 10,473 | https://github.com/sgl-project/sglang/pull/10473 | [1/2] Speed up trtllm_mla attention backend (>10% e2e) (#10473) | [1/2] Speed up trtllm_mla attention backend (>10% e2e) (#10473) | sgl-kernel/csrc/elementwise/concat_mla.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/CUDADataType.h>
#include <cuda_runtime.h>
#include "pytorch_extension_utils.h"
constexpr int NUM_LOCAL_HEADS = 128;
constexpr int QK_NOPE_HEAD_DIM = 128;
constexpr int QK_ROPE_HEAD_DIM = 64;
constexpr int K_HEAD_DIM = QK_NOPE_HEAD_DIM + QK_ROPE_HEAD_DIM;
constex... | #include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/CUDADataType.h>
#include <cuda_runtime.h>
#include "pytorch_extension_utils.h"
constexpr int NUM_LOCAL_HEADS = 128;
constexpr int QK_NOPE_HEAD_DIM = 128;
constexpr int QK_ROPE_HEAD_DIM = 64;
constexpr int K_HEAD_DIM = QK_NOPE_HEAD_DIM + QK_ROPE_HEAD_DIM;
constex... | --- sgl-kernel/csrc/elementwise/concat_mla.cu
+++ sgl-kernel/csrc/elementwise/concat_mla.cu
@@ -115,3 +115,105 @@
cudaError_t err = cudaGetLastError();
TORCH_CHECK(err == cudaSuccess, "CUDA kernel launch failed: ", cudaGetErrorString(err));
}
+
+// ============================== concat_mla_absorb_q =============... | 102 | 0 | 5 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/elementwise/concat_mla.cu).
Goal: [1/2] Speed up trtllm_mla attention backend (>10% e2e)
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/CUDADataType.h>
#include <cuda_runti... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
EDIT: 17.5% e2e speedup for short seq len, th... | [{"scope": "e2e", "hardware": null, "model": null, "shape": null, "batch_size": 1, "metric": "latency", "unit": "s", "higher_is_better": false, "baseline": 0.010048, "optimized": 0.006176, "stated_speedup": null, "speedup": 1.6269}, {"scope": "e2e", "hardware": null, "model": null, "shape": null, "batch_size": 16, "met... | null | 5.4031 |
sglang | 3b25dc127ae7df0680cfe0fd7800ba5aa280c64c | 2025-09-16T02:53:21+08:00 | 10,473 | https://github.com/sgl-project/sglang/pull/10473 | [1/2] Speed up trtllm_mla attention backend (>10% e2e) (#10473) | [1/2] Speed up trtllm_mla attention backend (>10% e2e) (#10473) | sgl-kernel/include/sgl_kernel_ops.h | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/include/sgl_kernel_ops.h
+++ sgl-kernel/include/sgl_kernel_ops.h
@@ -172,6 +172,7 @@
void copy_to_gpu_no_ce(const at::Tensor& input, at::Tensor& output);
void concat_mla_k(torch::Tensor k, torch::Tensor k_nope, torch::Tensor k_rope);
+void concat_mla_absorb_q(at::Tensor a, at::Tensor b, at::Tensor ou... | 1 | 0 | 5 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/include/sgl_kernel_ops.h).
Goal: [1/2] Speed up trtllm_mla attention backend (>10% e2e)
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
EDIT: 17.5% e2e speedup for short seq len, th... | [{"scope": "e2e", "hardware": null, "model": null, "shape": null, "batch_size": 1, "metric": "latency", "unit": "s", "higher_is_better": false, "baseline": 0.010048, "optimized": 0.006176, "stated_speedup": null, "speedup": 1.6269}, {"scope": "e2e", "hardware": null, "model": null, "shape": null, "batch_size": 16, "met... | null | 5.4031 |
sglang | a360511d7b9f92fde1f54a45f2a4d85831386d36 | 2025-09-13T10:57:06-07:00 | 9,748 | https://github.com/sgl-project/sglang/pull/9748 | [Generative Score API] Scoring(Prefill-only) optimizations. (#9748) | [Generative Score API] Scoring(Prefill-only) optimizations. (#9748) | python/sglang/srt/layers/logits_processor.py | triton | Apache-2.0 | # Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writin... | # Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writin... | --- python/sglang/srt/layers/logits_processor.py
+++ python/sglang/srt/layers/logits_processor.py
@@ -72,7 +72,10 @@
next_token_top_logprobs_val: Optional[List] = None
next_token_top_logprobs_idx: Optional[List] = None
# The logprobs and ids of the requested token ids in output positions. shape: [#seq, n... | 4 | 1 | 2 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/logits_processor.py).
Goal: [Generative Score API] Scoring(Prefill-only) optimizations.
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version ... | ## 🚀 Motivation
- Follow-up to [PR #8840](https://github.com/sgl-project/sglang/pull/8840) to reduce latency and increase throughput for the generative score API.
- Scoring only needs the next-token distribution after the full prompt, not per-token logprobs inside the prompt or any sampling.
**Performance Impact:**
... | [{"scope": "e2e", "hardware": "H100", "model": "Qwen3-0.6B", "shape": "prompt=300", "batch_size": 10, "metric": "P99 latency", "unit": "ms", "higher_is_better": false, "baseline": 6220, "optimized": 454, "stated_speedup": 13.7, "speedup": 13.7004}, {"scope": "e2e", "hardware": "H100", "model": "Qwen3-0.6B", "shape": "p... | null | 1.8151 |
sglang | c5d2b01cea6cc9a9ad0450222c7a4f19f5487b3d | 2025-09-12T05:56:25+08:00 | 10,303 | https://github.com/sgl-project/sglang/pull/10303 | [LongCat] Optimize zero_experts_compute_triton by changing mask (#10303) | [LongCat] Optimize zero_experts_compute_triton by changing mask (#10303) | python/sglang/srt/layers/moe/ep_moe/kernels.py | triton | Apache-2.0 | import logging
from typing import List, Optional
import torch
import triton
from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
from sglang.srt.utils import ceil_div, dispose_tensor, is_cuda
from sglang.utils import is_in_ci
logger = logging.getLogger(__name__)
_is_cuda = is_cuda()
if _i... | import logging
from typing import List, Optional
import torch
import triton
from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
from sglang.srt.utils import ceil_div, dispose_tensor, is_cuda
from sglang.utils import is_in_ci
logger = logging.getLogger(__name__)
_is_cuda = is_cuda()
if _i... | --- python/sglang/srt/layers/moe/ep_moe/kernels.py
+++ python/sglang/srt/layers/moe/ep_moe/kernels.py
@@ -1416,7 +1416,7 @@
zero_expert_scales[zero_expert_mask] = 0.0
normal_expert_mask = expert_indices >= num_experts
- expert_indices[normal_expert_mask] = 0
+ expert_indices[normal_expert_mask] = ... | 1 | 1 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/ep_moe/kernels.py).
Goal: [LongCat] Optimize zero_experts_compute_triton by changing mask
Target hardware: unspecified
Target models: unspecified
Current file:
```python
import logging
from typing import List, Optional
import torch
import tri... | update moe expert mask @Orchard-DT . Sincerely invite you to help check.
<!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss ... | [{"scope": "e2e", "hardware": null, "model": null, "shape": null, "batch_size": null, "metric": "throughput", "unit": null, "higher_is_better": true, "baseline": null, "optimized": null, "stated_speedup": 1.1, "speedup": 1.1}] | null | 1.1 |
sglang | 0096798ed60b9eadce468c2d206cd2982e97b978 | 2025-09-09T00:00:33+08:00 | 10,156 | https://github.com/sgl-project/sglang/pull/10156 | [1/2] Speed up prefill mla attention (#10156) | [1/2] Speed up prefill mla attention (#10156) | sgl-kernel/csrc/common_extension.cc | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/common_extension.cc
+++ sgl-kernel/csrc/common_extension.cc
@@ -436,6 +436,8 @@
m.def("copy_to_gpu_no_ce(Tensor input, Tensor! output) -> ()");
m.impl("copy_to_gpu_no_ce", torch::kCUDA, ©_to_gpu_no_ce);
+ m.def("concat_mla_k(Tensor! k, Tensor k_nope, Tensor k_rope) -> ()");
+ m.impl("c... | 2 | 0 | 4 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/common_extension.cc).
Goal: [1/2] Speed up prefill mla attention
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
in nsys I see bandwidth is ~90%, and it is no... | [{"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": 2048, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 0.109568, "optimized": 0.028672, "stated_speedup": null, "speedup": 3.8214}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size"... | 3.8106 | 1.0196 |
sglang | 0096798ed60b9eadce468c2d206cd2982e97b978 | 2025-09-09T00:00:33+08:00 | 10,156 | https://github.com/sgl-project/sglang/pull/10156 | [1/2] Speed up prefill mla attention (#10156) | [1/2] Speed up prefill mla attention (#10156) | sgl-kernel/include/sgl_kernel_ops.h | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/include/sgl_kernel_ops.h
+++ sgl-kernel/include/sgl_kernel_ops.h
@@ -723,3 +723,4 @@
void store_kv_cache(at::Tensor k_cache, at::Tensor v_cache, at::Tensor out_loc, at::Tensor k, at::Tensor v);
void copy_to_gpu_no_ce(const at::Tensor& input, at::Tensor& output);
+void concat_mla_k(torch::Tensor k, to... | 1 | 0 | 4 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/include/sgl_kernel_ops.h).
Goal: [1/2] Speed up prefill mla attention
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
in nsys I see bandwidth is ~90%, and it is no... | [{"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": 2048, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 0.109568, "optimized": 0.028672, "stated_speedup": null, "speedup": 3.8214}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size"... | 3.8106 | 1.0196 |
sglang | 85ed8e0a5ec9ed47a5d1ac93ecc8bee2159468e0 | 2025-09-07T13:31:00+08:00 | 10,101 | https://github.com/sgl-project/sglang/pull/10101 | Optimize nvfp4 block scaled gemm kernel when M is small. (#10101) | Optimize nvfp4 block scaled gemm kernel when M is small. (#10101) | sgl-kernel/csrc/gemm/nvfp4_scaled_mm_kernels.cu | cuda | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/gemm/nvfp4_scaled_mm_kernels.cu
+++ sgl-kernel/csrc/gemm/nvfp4_scaled_mm_kernels.cu
@@ -38,27 +38,74 @@
using namespace cute;
#if defined(CUTLASS_ARCH_MMA_SM100_SUPPORTED)
-// Kernel Perf config
-template <typename T>
-struct KernelTraits {
+// Config(half_t/bfloat16_t) for M <= 128
+template <t... | 118 | 34 | 1 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/gemm/nvfp4_scaled_mm_kernels.cu).
Goal: Optimize nvfp4 block scaled gemm kernel when M is small.
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache Lice... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
Until now, the CUTLASS-based nvfp4 block-scale... | [] | null | null |
sglang | cb3918a09127a1da8cc0976f86e7425285a1dca6 | 2025-09-07T09:16:18+08:00 | 9,477 | https://github.com/sgl-project/sglang/pull/9477 | Optimize moe_sum_reduce_kernel (#9477) | Optimize moe_sum_reduce_kernel (#9477)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
Co-authored-by: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com> | python/sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py | triton | Apache-2.0 | from __future__ import annotations
import os
from typing import Any, Dict, List, Optional
import torch
import triton
import triton.language as tl
from sglang.srt.layers.quantization.fp8_kernel import (
per_token_group_quant_fp8,
scaled_fp8_quant,
sglang_per_token_group_quant_fp8,
)
from sglang.srt.layers... | from __future__ import annotations
import os
from typing import Any, Dict, List, Optional
import torch
import triton
import triton.language as tl
from sglang.srt.layers.quantization.fp8_kernel import (
per_token_group_quant_fp8,
scaled_fp8_quant,
sglang_per_token_group_quant_fp8,
)
from sglang.srt.layers... | --- python/sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py
+++ python/sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py
@@ -735,29 +735,32 @@
token_block_id = tl.program_id(0)
dim_block_id = tl.program_id(1)
- token_start = token_block_id * BLOCK_M
- token_end = min... | 27 | 24 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py).
Goal: Optimize moe_sum_reduce_kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from __future__ import annotations
import os
from typing import Any, Dict, Lis... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
During the Prefill stage of TP MoE, the moe_s... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "num_tokens=1.0", "batch_size": 1, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 9.92, "optimized": 9.536, "stated_speedup": null, "speedup": 1.0403}, {"scope": "kernel", "hardware": null, "model": null, "shape": "num_tokens=2.0"... | 1.0343 | 1.1283 |
sglang | 339f8eef09e1cf94044dbe1fdb556dfc01430428 | 2025-09-05T18:45:08+08:00 | 9,534 | https://github.com/sgl-project/sglang/pull/9534 | [1/2] Optimizations and refactors about quant kernel (#9534) | [1/2] Optimizations and refactors about quant kernel (#9534) | python/sglang/srt/layers/quantization/fp8_kernel.py | triton | Apache-2.0 | # Copyright 2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, so... | # Copyright 2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, so... | --- python/sglang/srt/layers/quantization/fp8_kernel.py
+++ python/sglang/srt/layers/quantization/fp8_kernel.py
@@ -43,11 +43,17 @@
_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
if _is_cuda:
- from sgl_kernel import (
- sgl_per_tensor_quant_fp8,
- sgl_per_token_group_quant_fp8,
- ... | 29 | 8 | 7 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/quantization/fp8_kernel.py).
Goal: [1/2] Optimizations and refactors about quant kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
to review this, just look at https://github.c... | [] | null | null |
sglang | 339f8eef09e1cf94044dbe1fdb556dfc01430428 | 2025-09-05T18:45:08+08:00 | 9,534 | https://github.com/sgl-project/sglang/pull/9534 | [1/2] Optimizations and refactors about quant kernel (#9534) | [1/2] Optimizations and refactors about quant kernel (#9534) | python/sglang/srt/layers/quantization/int8_kernel.py | triton | Apache-2.0 | import functools
import json
import logging
import os
from typing import Any, Dict, List, Optional, Tuple
import torch
import triton
import triton.language as tl
from sglang.srt.utils import get_device_name, is_cuda
_is_cuda = is_cuda()
if _is_cuda:
from sgl_kernel import sgl_per_token_group_quant_int8
logger =... | import functools
import json
import logging
import os
from typing import Any, Dict, List, Optional, Tuple
import torch
import triton
import triton.language as tl
from sglang.srt.utils import get_device_name, is_cuda
_is_cuda = is_cuda()
if _is_cuda:
# Temporary
try:
from sgl_kernel import sgl_per_tok... | --- python/sglang/srt/layers/quantization/int8_kernel.py
+++ python/sglang/srt/layers/quantization/int8_kernel.py
@@ -12,7 +12,13 @@
_is_cuda = is_cuda()
if _is_cuda:
- from sgl_kernel import sgl_per_token_group_quant_int8
+ # Temporary
+ try:
+ from sgl_kernel import sgl_per_token_group_quant_8bit
... | 8 | 2 | 7 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/quantization/int8_kernel.py).
Goal: [1/2] Optimizations and refactors about quant kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```python
import functools
import json
import logging
import os
from typing import Any, ... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
to review this, just look at https://github.c... | [] | null | null |
sglang | 339f8eef09e1cf94044dbe1fdb556dfc01430428 | 2025-09-05T18:45:08+08:00 | 9,534 | https://github.com/sgl-project/sglang/pull/9534 | [1/2] Optimizations and refactors about quant kernel (#9534) | [1/2] Optimizations and refactors about quant kernel (#9534) | sgl-kernel/csrc/common_extension.cc | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/common_extension.cc
+++ sgl-kernel/csrc/common_extension.cc
@@ -121,14 +121,9 @@
m.impl("fp8_blockwise_scaled_mm", torch::kCUDA, &fp8_blockwise_scaled_mm);
m.def(
- "sgl_per_token_group_quant_fp8(Tensor input, Tensor output_q, Tensor output_s, int group_size,"
- " float eps, float ... | 3 | 8 | 7 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/common_extension.cc).
Goal: [1/2] Optimizations and refactors about quant kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
to review this, just look at https://github.c... | [] | null | null |
sglang | 339f8eef09e1cf94044dbe1fdb556dfc01430428 | 2025-09-05T18:45:08+08:00 | 9,534 | https://github.com/sgl-project/sglang/pull/9534 | [1/2] Optimizations and refactors about quant kernel (#9534) | [1/2] Optimizations and refactors about quant kernel (#9534) | sgl-kernel/csrc/gemm/per_token_group_quant_8bit.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <cuda_fp8.h>
#include <cmath>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
__device__ __forceinline__ float GroupReduceMax(float val, const int tid) {
unsigned mask = 0xffff;
val = fmaxf(val, __shfl_xor_sync(mask, val, 8));
val = fmaxf(val, __shfl_xor_syn... | #include <ATen/cuda/CUDAContext.h>
#include <c10/util/Float8_e4m3fn.h>
#include <cmath>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
template <int THREADS_PER_SUBWARP>
__device__ __forceinline__ float GroupReduceMax(float val, const int tid) {
unsigned mask = 0xffff;
static_assert(
(THREADS_PER... | --- sgl-kernel/csrc/gemm/per_token_group_quant_8bit.cu
+++ sgl-kernel/csrc/gemm/per_token_group_quant_8bit.cu
@@ -1,119 +1,396 @@
#include <ATen/cuda/CUDAContext.h>
-#include <cuda_fp8.h>
+#include <c10/util/Float8_e4m3fn.h>
#include <cmath>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
+template <i... | 453 | 183 | 7 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/gemm/per_token_group_quant_8bit.cu).
Goal: [1/2] Optimizations and refactors about quant kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#include <cuda_fp8.h>
#include <cmath>
#include ... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
to review this, just look at https://github.c... | [] | null | null |
sglang | 339f8eef09e1cf94044dbe1fdb556dfc01430428 | 2025-09-05T18:45:08+08:00 | 9,534 | https://github.com/sgl-project/sglang/pull/9534 | [1/2] Optimizations and refactors about quant kernel (#9534) | [1/2] Optimizations and refactors about quant kernel (#9534) | sgl-kernel/include/sgl_kernel_ops.h | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/include/sgl_kernel_ops.h
+++ sgl-kernel/include/sgl_kernel_ops.h
@@ -207,23 +207,17 @@
const torch::Dtype& out_dtype);
void scaled_fp4_quant(
torch::Tensor& output, torch::Tensor const& input, torch::Tensor& output_scale, torch::Tensor const& input_scale);
-void sgl_per_token_group_quant_fp8(
... | 6 | 12 | 7 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/include/sgl_kernel_ops.h).
Goal: [1/2] Optimizations and refactors about quant kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
to review this, just look at https://github.c... | [] | null | null |
sglang | 5c34b4f1c769fc1150a39b96a4673559d6657a8e | 2025-08-29T17:17:03-07:00 | 9,556 | https://github.com/sgl-project/sglang/pull/9556 | [NVIDIA] [2/N] Optimize `silu_and_mul_scaled_fp4_grouped_quant` perf (#9556) | [NVIDIA] [2/N] Optimize `silu_and_mul_scaled_fp4_grouped_quant` perf (#9556) | sgl-kernel/csrc/common_extension.cc | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/common_extension.cc
+++ sgl-kernel/csrc/common_extension.cc
@@ -159,8 +159,7 @@
m.def(
"silu_and_mul_scaled_fp4_experts_quant(Tensor! output, Tensor! output_scale,"
- "Tensor input, Tensor input_global_scale, Tensor input_offset_by_experts,"
- "Tensor output_scale_offset_by_exp... | 1 | 2 | 5 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/common_extension.cc).
Goal: [NVIDIA] [2/N] Optimize `silu_and_mul_scaled_fp4_grouped_quant` perf
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache Licens... | This PR further optimizes `silu_and_mul_scaled_fp4_grouped_quant` for DeepEP low-latency scenarios. When the mask contains small values, the kernel achieves up to 82× speedup over the previous version. The change also improves performance for non-masked NVFP4 expert quantization (`scaled_fp4_experts_quant`) and NVFP4 g... | [{"scope": "kernel", "hardware": null, "model": "DeepEP", "shape": null, "batch_size": null, "metric": "performance", "unit": null, "higher_is_better": true, "baseline": null, "optimized": null, "stated_speedup": 82, "speedup": 82.0}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": nul... | 41.75 | null |
sglang | 5c34b4f1c769fc1150a39b96a4673559d6657a8e | 2025-08-29T17:17:03-07:00 | 9,556 | https://github.com/sgl-project/sglang/pull/9556 | [NVIDIA] [2/N] Optimize `silu_and_mul_scaled_fp4_grouped_quant` perf (#9556) | [NVIDIA] [2/N] Optimize `silu_and_mul_scaled_fp4_grouped_quant` perf (#9556) | sgl-kernel/csrc/gemm/nvfp4_expert_quant.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cuda_fp8.h>
#include <cuda_runtime.h>
#include <torch/all.h>
template <typename T>
struct TypeConverter {
using Type = half2;
}; // keep for generality
template <>
struct TypeConverter<half2> {
using Type = half;
};
template <>
struct ... | #include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cuda_fp8.h>
#include <cuda_runtime.h>
#include <torch/all.h>
template <typename T>
struct TypeConverter {
using Type = half2;
}; // keep for generality
template <>
struct TypeConverter<half2> {
using Type = half;
};
template <>
struct ... | --- sgl-kernel/csrc/gemm/nvfp4_expert_quant.cu
+++ sgl-kernel/csrc/gemm/nvfp4_expert_quant.cu
@@ -347,7 +347,7 @@
}
}
- // Eerly exit when using masks.
+ // Early exit when using masks.
if (use_mask && rowIdx_in_expert >= mask[expert_idx]) {
continue;
}
@@ -374,6 +374,107 @@
in... | 134 | 19 | 5 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/gemm/nvfp4_expert_quant.cu).
Goal: [NVIDIA] [2/N] Optimize `silu_and_mul_scaled_fp4_grouped_quant` perf
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include ... | This PR further optimizes `silu_and_mul_scaled_fp4_grouped_quant` for DeepEP low-latency scenarios. When the mask contains small values, the kernel achieves up to 82× speedup over the previous version. The change also improves performance for non-masked NVFP4 expert quantization (`scaled_fp4_experts_quant`) and NVFP4 g... | [{"scope": "kernel", "hardware": null, "model": "DeepEP", "shape": null, "batch_size": null, "metric": "performance", "unit": null, "higher_is_better": true, "baseline": null, "optimized": null, "stated_speedup": 82, "speedup": 82.0}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": nul... | 41.75 | null |
sglang | 5c34b4f1c769fc1150a39b96a4673559d6657a8e | 2025-08-29T17:17:03-07:00 | 9,556 | https://github.com/sgl-project/sglang/pull/9556 | [NVIDIA] [2/N] Optimize `silu_and_mul_scaled_fp4_grouped_quant` perf (#9556) | [NVIDIA] [2/N] Optimize `silu_and_mul_scaled_fp4_grouped_quant` perf (#9556) | sgl-kernel/csrc/gemm/nvfp4_quant_entry.cu | cuda | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/gemm/nvfp4_quant_entry.cu
+++ sgl-kernel/csrc/gemm/nvfp4_quant_entry.cu
@@ -32,9 +32,8 @@
torch::Tensor& output_scale,
torch::Tensor const& input,
torch::Tensor const& input_global_scale,
- torch::Tensor const& input_offset_by_experts,
- torch::Tensor const& output_scale_offset_... | 5 | 7 | 5 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/gemm/nvfp4_quant_entry.cu).
Goal: [NVIDIA] [2/N] Optimize `silu_and_mul_scaled_fp4_grouped_quant` perf
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apach... | This PR further optimizes `silu_and_mul_scaled_fp4_grouped_quant` for DeepEP low-latency scenarios. When the mask contains small values, the kernel achieves up to 82× speedup over the previous version. The change also improves performance for non-masked NVFP4 expert quantization (`scaled_fp4_experts_quant`) and NVFP4 g... | [{"scope": "kernel", "hardware": null, "model": "DeepEP", "shape": null, "batch_size": null, "metric": "performance", "unit": null, "higher_is_better": true, "baseline": null, "optimized": null, "stated_speedup": 82, "speedup": 82.0}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": nul... | 41.75 | null |
sglang | 5c34b4f1c769fc1150a39b96a4673559d6657a8e | 2025-08-29T17:17:03-07:00 | 9,556 | https://github.com/sgl-project/sglang/pull/9556 | [NVIDIA] [2/N] Optimize `silu_and_mul_scaled_fp4_grouped_quant` perf (#9556) | [NVIDIA] [2/N] Optimize `silu_and_mul_scaled_fp4_grouped_quant` perf (#9556) | sgl-kernel/include/sgl_kernel_ops.h | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/include/sgl_kernel_ops.h
+++ sgl-kernel/include/sgl_kernel_ops.h
@@ -394,9 +394,8 @@
torch::Tensor& output_scale,
torch::Tensor const& input,
torch::Tensor const& input_global_scale,
- torch::Tensor const& input_offset_by_experts,
- torch::Tensor const& output_scale_offset_by_experts... | 2 | 3 | 5 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/include/sgl_kernel_ops.h).
Goal: [NVIDIA] [2/N] Optimize `silu_and_mul_scaled_fp4_grouped_quant` perf
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache Licens... | This PR further optimizes `silu_and_mul_scaled_fp4_grouped_quant` for DeepEP low-latency scenarios. When the mask contains small values, the kernel achieves up to 82× speedup over the previous version. The change also improves performance for non-masked NVFP4 expert quantization (`scaled_fp4_experts_quant`) and NVFP4 g... | [{"scope": "kernel", "hardware": null, "model": "DeepEP", "shape": null, "batch_size": null, "metric": "performance", "unit": null, "higher_is_better": true, "baseline": null, "optimized": null, "stated_speedup": 82, "speedup": 82.0}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": nul... | 41.75 | null |
sglang | 5ad296bda140117e4eb85ce781955d94fb221973 | 2025-08-29T08:21:55+08:00 | 8,750 | https://github.com/sgl-project/sglang/pull/8750 | Optimize prefill performance on cpu backend (#8750) | Optimize prefill performance on cpu backend (#8750) | sgl-kernel/csrc/cpu/common.h | cpp | Apache-2.0 | #pragma once
#include <ATen/ATen.h>
#include <ATen/Parallel.h>
#include <ATen/record_function.h>
#if defined(_OPENMP)
#include <omp.h>
#endif
namespace {
// dispatch bool
#define AT_DISPATCH_BOOL(BOOL_V, BOOL_NAME, ...) \
[&] { \
if (BOOL_V) { ... | #pragma once
#include <ATen/ATen.h>
#include <ATen/Parallel.h>
#include <ATen/record_function.h>
#if defined(_OPENMP)
#include <omp.h>
#endif
namespace {
// dispatch bool
#define AT_DISPATCH_BOOL(BOOL_V, BOOL_NAME, ...) \
[&] { \
if (BOOL_V) { ... | --- sgl-kernel/csrc/cpu/common.h
+++ sgl-kernel/csrc/cpu/common.h
@@ -105,7 +105,19 @@
#define CHECK_EQ(a, b) TORCH_CHECK((a) == (b), "CHECK_EQ(" #a ", " #b ") failed. ", a, " vs ", b)
-// parallel routines
+// [NB] Parallel Routines
+//
+// * at::parallel_for - applies for most of generic use cases, this will be... | 118 | 1 | 9 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/common.h).
Goal: Optimize prefill performance on cpu backend
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#pragma once
#include <ATen/ATen.h>
#include <ATen/Parallel.h>
#include <ATen/record_function.h>
#if defined(_OPENMP)
#incl... | ## Motivation
This PR aims at improving prefill performance for CPU backend, it will improve `bfloat16`, `int8_w8a8` and `fp8_w8a8` paths. This one has no effective on decoding performance.
## Modifications
* **amx-int8**: enable **amx-int8** for GEMM and MoE kernels. As SGLang pops up with PyTorch 2.7, we are free ... | [{"scope": "kernel", "hardware": "Intel 4th gen Xeon (SPR)", "model": null, "shape": "M=1024,K=4096,N=14336", "batch_size": 1024, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 4.375, "optimized": 3.483, "stated_speedup": null, "speedup": 1.2561}, {"scope": "kernel", "hardware": "Intel 4th ge... | 1.3585 | null |
sglang | 5ad296bda140117e4eb85ce781955d94fb221973 | 2025-08-29T08:21:55+08:00 | 8,750 | https://github.com/sgl-project/sglang/pull/8750 | Optimize prefill performance on cpu backend (#8750) | Optimize prefill performance on cpu backend (#8750) | sgl-kernel/csrc/cpu/gemm.cpp | cpp | Apache-2.0 | #include "gemm.h"
#include "common.h"
#include "vec.h"
namespace {
// packed layout:
// quants {N, K} int8_t
// comp {N} int32_t
template <int BLOCK_N>
inline void s8s8_compensation(int8_t* __restrict__ packed, int K) {
#if defined(CPU_CAPABILITY_AVX512)
constexpr int COLS = BLOCK_N / 16;
__m512i vc... | #include "gemm.h"
#include "common.h"
#include "vec.h"
namespace {
// packed layout:
// quants {N, K} int8_t
// comp {N} int32_t
template <int BLOCK_N>
inline void s8s8_compensation(int8_t* __restrict__ packed, int K) {
#if defined(CPU_CAPABILITY_AVX512)
constexpr int COLS = BLOCK_N / 16;
__m512i vc... | --- sgl-kernel/csrc/cpu/gemm.cpp
+++ sgl-kernel/csrc/cpu/gemm.cpp
@@ -254,7 +254,7 @@
return;
}
- // pattern: 1-4-16
+ // pattern: 1-4-16, N = 16, 32, 48, 64
constexpr int64_t BLOCK_M = 4;
constexpr int64_t BLOCK_N = 64;
const int64_t MB = div_up(M, BLOCK_M);
@@ -268,35 +268,59 @@
switch (... | 30 | 14 | 9 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/gemm.cpp).
Goal: Optimize prefill performance on cpu backend
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "gemm.h"
#include "common.h"
#include "vec.h"
namespace {
// packed layout:
// quants {N, K} int8_t
// com... | ## Motivation
This PR aims at improving prefill performance for CPU backend, it will improve `bfloat16`, `int8_w8a8` and `fp8_w8a8` paths. This one has no effective on decoding performance.
## Modifications
* **amx-int8**: enable **amx-int8** for GEMM and MoE kernels. As SGLang pops up with PyTorch 2.7, we are free ... | [{"scope": "kernel", "hardware": "Intel 4th gen Xeon (SPR)", "model": null, "shape": "M=1024,K=4096,N=14336", "batch_size": 1024, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 4.375, "optimized": 3.483, "stated_speedup": null, "speedup": 1.2561}, {"scope": "kernel", "hardware": "Intel 4th ge... | 1.3585 | null |
sglang | 5ad296bda140117e4eb85ce781955d94fb221973 | 2025-08-29T08:21:55+08:00 | 8,750 | https://github.com/sgl-project/sglang/pull/8750 | Optimize prefill performance on cpu backend (#8750) | Optimize prefill performance on cpu backend (#8750) | sgl-kernel/csrc/cpu/gemm.h | cpp | Apache-2.0 | #pragma once
#include <ATen/native/CPUBlas.h>
#include "common.h"
// amx-bf16
#define TILE_M 16
#define TILE_N 16
#define TILE_K 32
// block size for AMX gemm
constexpr int block_size_m() {
return 2 * TILE_M;
}
constexpr int block_size_n() {
return 2 * TILE_N;
}
// define threshold using brgemm (intel AMX)
temp... | #pragma once
#include <ATen/native/CPUBlas.h>
#include "common.h"
// amx-bf16
#define TILE_M 16
#define TILE_N 16
#define TILE_K 32
// block size for AMX gemm
constexpr int block_size_m() {
return 2 * TILE_M;
}
constexpr int block_size_n() {
return 2 * TILE_N;
}
// define threshold using brgemm (intel AMX)
temp... | --- sgl-kernel/csrc/cpu/gemm.h
+++ sgl-kernel/csrc/cpu/gemm.h
@@ -27,10 +27,10 @@
inline bool can_use_brgemm<at::Half>(int M) {
return true;
}
-// TODO: add u8s8 brgemm, this requires PyTorch 2.7
+// this requires PyTorch 2.7 or above
template <>
inline bool can_use_brgemm<int8_t>(int M) {
- return false;
+ re... | 4 | 3 | 9 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/gemm.h).
Goal: Optimize prefill performance on cpu backend
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#pragma once
#include <ATen/native/CPUBlas.h>
#include "common.h"
// amx-bf16
#define TILE_M 16
#define TILE_N 16
#define TIL... | ## Motivation
This PR aims at improving prefill performance for CPU backend, it will improve `bfloat16`, `int8_w8a8` and `fp8_w8a8` paths. This one has no effective on decoding performance.
## Modifications
* **amx-int8**: enable **amx-int8** for GEMM and MoE kernels. As SGLang pops up with PyTorch 2.7, we are free ... | [{"scope": "kernel", "hardware": "Intel 4th gen Xeon (SPR)", "model": null, "shape": "M=1024,K=4096,N=14336", "batch_size": 1024, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 4.375, "optimized": 3.483, "stated_speedup": null, "speedup": 1.2561}, {"scope": "kernel", "hardware": "Intel 4th ge... | 1.3585 | null |
sglang | 5ad296bda140117e4eb85ce781955d94fb221973 | 2025-08-29T08:21:55+08:00 | 8,750 | https://github.com/sgl-project/sglang/pull/8750 | Optimize prefill performance on cpu backend (#8750) | Optimize prefill performance on cpu backend (#8750) | sgl-kernel/csrc/cpu/gemm_fp8.cpp | cpp | Apache-2.0 | #include "common.h"
#include "gemm.h"
#include "vec.h"
// we use 4x32 for BLOCK_M
#define BLOCK_SIZE_M_SCALE 4
namespace {
template <typename scalar_t>
inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ input, int64_t size) {
using bVec = at::vec::Vectorized<scalar_t>;
using fVec = at::v... | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
template <typename scalar_t>
inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ input, int64_t size) {
using bVec = at::vec::Vectorized<scalar_t>;
using fVec = at::vec::Vectorized<float>;
constexpr int kVecSize = bVec::s... | --- sgl-kernel/csrc/cpu/gemm_fp8.cpp
+++ sgl-kernel/csrc/cpu/gemm_fp8.cpp
@@ -1,9 +1,6 @@
#include "common.h"
#include "gemm.h"
#include "vec.h"
-
-// we use 4x32 for BLOCK_M
-#define BLOCK_SIZE_M_SCALE 4
namespace {
@@ -250,7 +247,8 @@
int K,
int lda,
int ldb,
- int ldc) {
+ int ... | 36 | 33 | 9 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/gemm_fp8.cpp).
Goal: Optimize prefill performance on cpu backend
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "common.h"
#include "gemm.h"
#include "vec.h"
// we use 4x32 for BLOCK_M
#define BLOCK_SIZE_M_SCALE 4
namespac... | ## Motivation
This PR aims at improving prefill performance for CPU backend, it will improve `bfloat16`, `int8_w8a8` and `fp8_w8a8` paths. This one has no effective on decoding performance.
## Modifications
* **amx-int8**: enable **amx-int8** for GEMM and MoE kernels. As SGLang pops up with PyTorch 2.7, we are free ... | [{"scope": "kernel", "hardware": "Intel 4th gen Xeon (SPR)", "model": null, "shape": "M=1024,K=4096,N=14336", "batch_size": 1024, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 4.375, "optimized": 3.483, "stated_speedup": null, "speedup": 1.2561}, {"scope": "kernel", "hardware": "Intel 4th ge... | 1.3585 | null |
sglang | 5ad296bda140117e4eb85ce781955d94fb221973 | 2025-08-29T08:21:55+08:00 | 8,750 | https://github.com/sgl-project/sglang/pull/8750 | Optimize prefill performance on cpu backend (#8750) | Optimize prefill performance on cpu backend (#8750) | sgl-kernel/csrc/cpu/gemm_int8.cpp | cpp | Apache-2.0 | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
template <typename scalar_t, bool has_bias, int BLOCK_M, int BLOCK_N>
struct tinygemm_kernel_nn {
static inline void apply(
const uint8_t* __restrict__ A,
const int8_t* __restrict__ B,
scalar_t* __restrict__ C,
const float* _... | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
template <typename scalar_t, bool has_bias, int BLOCK_N>
struct scale_C {
static inline void apply(
scalar_t* __restrict__ C,
const int32_t* __restrict__ Ctmp,
const int32_t* __restrict__ Bcomp,
const float* __restrict__ bias... | --- sgl-kernel/csrc/cpu/gemm_int8.cpp
+++ sgl-kernel/csrc/cpu/gemm_int8.cpp
@@ -3,6 +3,61 @@
#include "vec.h"
namespace {
+
+template <typename scalar_t, bool has_bias, int BLOCK_N>
+struct scale_C {
+ static inline void apply(
+ scalar_t* __restrict__ C,
+ const int32_t* __restrict__ Ctmp,
+ const ... | 70 | 12 | 9 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/gemm_int8.cpp).
Goal: Optimize prefill performance on cpu backend
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
template <typename scalar_t, bool has_bias, int BLO... | ## Motivation
This PR aims at improving prefill performance for CPU backend, it will improve `bfloat16`, `int8_w8a8` and `fp8_w8a8` paths. This one has no effective on decoding performance.
## Modifications
* **amx-int8**: enable **amx-int8** for GEMM and MoE kernels. As SGLang pops up with PyTorch 2.7, we are free ... | [{"scope": "kernel", "hardware": "Intel 4th gen Xeon (SPR)", "model": null, "shape": "M=1024,K=4096,N=14336", "batch_size": 1024, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 4.375, "optimized": 3.483, "stated_speedup": null, "speedup": 1.2561}, {"scope": "kernel", "hardware": "Intel 4th ge... | 1.3585 | null |
sglang | 5ad296bda140117e4eb85ce781955d94fb221973 | 2025-08-29T08:21:55+08:00 | 8,750 | https://github.com/sgl-project/sglang/pull/8750 | Optimize prefill performance on cpu backend (#8750) | Optimize prefill performance on cpu backend (#8750) | sgl-kernel/csrc/cpu/moe.cpp | cpp | Apache-2.0 | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
// [NOTE]: Fused MoE kernel with AMX
//
// This file contains implementations for
// * `moe_align_block_size`
// * `fused_moe`
//
// The functionality is identical to triton kernel, excepts:
// * fuse silu_and_mul with gemm1, therefore... | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
// [NOTE]: Fused MoE kernel with AMX
//
// This file contains implementations for
// * `moe_align_block_size`
// * `fused_moe`
//
// The functionality is identical to triton kernel, excepts:
// * fuse silu_and_mul with gemm1, therefore... | --- sgl-kernel/csrc/cpu/moe.cpp
+++ sgl-kernel/csrc/cpu/moe.cpp
@@ -579,35 +579,30 @@
const int64_t stride_e = 2 * N * K;
const int64_t stride_n = K;
+ int64_t avg_M = std::max(int64_t(1), M * topk / E);
+ const bool use_brgemm = can_use_brgemm<scalar_t>(avg_M);
+
// here we only parallel on half of 2N to ... | 43 | 73 | 9 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/moe.cpp).
Goal: Optimize prefill performance on cpu backend
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
// [NOTE]: Fused MoE kernel with AMX
//
// This file co... | ## Motivation
This PR aims at improving prefill performance for CPU backend, it will improve `bfloat16`, `int8_w8a8` and `fp8_w8a8` paths. This one has no effective on decoding performance.
## Modifications
* **amx-int8**: enable **amx-int8** for GEMM and MoE kernels. As SGLang pops up with PyTorch 2.7, we are free ... | [{"scope": "kernel", "hardware": "Intel 4th gen Xeon (SPR)", "model": null, "shape": "M=1024,K=4096,N=14336", "batch_size": 1024, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 4.375, "optimized": 3.483, "stated_speedup": null, "speedup": 1.2561}, {"scope": "kernel", "hardware": "Intel 4th ge... | 1.3585 | null |
sglang | 5ad296bda140117e4eb85ce781955d94fb221973 | 2025-08-29T08:21:55+08:00 | 8,750 | https://github.com/sgl-project/sglang/pull/8750 | Optimize prefill performance on cpu backend (#8750) | Optimize prefill performance on cpu backend (#8750) | sgl-kernel/csrc/cpu/moe_fp8.cpp | cpp | Apache-2.0 | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
template <typename scalar_t>
inline void copy_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t size) {
using Vec = at::vec::Vectorized<scalar_t>;
// no remainder
#pragma GCC unroll 4
for (int64_t d = 0; d < size; d += V... | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
template <typename scalar_t>
inline void copy_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t size) {
using Vec = at::vec::Vectorized<scalar_t>;
// no remainder
#pragma GCC unroll 4
for (int64_t d = 0; d < size; d += V... | --- sgl-kernel/csrc/cpu/moe_fp8.cpp
+++ sgl-kernel/csrc/cpu/moe_fp8.cpp
@@ -174,18 +174,18 @@
const int64_t stride_e = 2 * N * K;
const int64_t stride_n = K;
+ int64_t avg_M = std::max(int64_t(1), M * topk / E);
+ const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(avg_M);
+
+ int64_t B_tmp_size_per_th... | 54 | 49 | 9 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/moe_fp8.cpp).
Goal: Optimize prefill performance on cpu backend
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
template <typename scalar_t>
inline void copy_stub(sc... | ## Motivation
This PR aims at improving prefill performance for CPU backend, it will improve `bfloat16`, `int8_w8a8` and `fp8_w8a8` paths. This one has no effective on decoding performance.
## Modifications
* **amx-int8**: enable **amx-int8** for GEMM and MoE kernels. As SGLang pops up with PyTorch 2.7, we are free ... | [{"scope": "kernel", "hardware": "Intel 4th gen Xeon (SPR)", "model": null, "shape": "M=1024,K=4096,N=14336", "batch_size": 1024, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 4.375, "optimized": 3.483, "stated_speedup": null, "speedup": 1.2561}, {"scope": "kernel", "hardware": "Intel 4th ge... | 1.3585 | null |
sglang | 5ad296bda140117e4eb85ce781955d94fb221973 | 2025-08-29T08:21:55+08:00 | 8,750 | https://github.com/sgl-project/sglang/pull/8750 | Optimize prefill performance on cpu backend (#8750) | Optimize prefill performance on cpu backend (#8750) | sgl-kernel/csrc/cpu/moe_int8.cpp | cpp | Apache-2.0 | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
template <typename scalar_t>
inline void copy_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t size) {
using Vec = at::vec::Vectorized<scalar_t>;
// no remainder
#pragma GCC unroll 4
for (int64_t d = 0; d < size; d += V... | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
template <typename scalar_t>
inline void copy_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t size) {
using Vec = at::vec::Vectorized<scalar_t>;
// no remainder
#pragma GCC unroll 4
for (int64_t d = 0; d < size; d += V... | --- sgl-kernel/csrc/cpu/moe_int8.cpp
+++ sgl-kernel/csrc/cpu/moe_int8.cpp
@@ -107,6 +107,120 @@
for (; d < size; ++d) {
out[d] = static_cast<scalar_t>(input[d] + float(input2[d]) * scale);
}
+}
+
+template <typename scalar_t, int BLOCK_N>
+inline void silu_and_mul(
+ scalar_t* __restrict__ C,
+ const i... | 336 | 98 | 9 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/moe_int8.cpp).
Goal: Optimize prefill performance on cpu backend
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
template <typename scalar_t>
inline void copy_stub(s... | ## Motivation
This PR aims at improving prefill performance for CPU backend, it will improve `bfloat16`, `int8_w8a8` and `fp8_w8a8` paths. This one has no effective on decoding performance.
## Modifications
* **amx-int8**: enable **amx-int8** for GEMM and MoE kernels. As SGLang pops up with PyTorch 2.7, we are free ... | [{"scope": "kernel", "hardware": "Intel 4th gen Xeon (SPR)", "model": null, "shape": "M=1024,K=4096,N=14336", "batch_size": 1024, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 4.375, "optimized": 3.483, "stated_speedup": null, "speedup": 1.2561}, {"scope": "kernel", "hardware": "Intel 4th ge... | 1.3585 | null |
sglang | 5ad296bda140117e4eb85ce781955d94fb221973 | 2025-08-29T08:21:55+08:00 | 8,750 | https://github.com/sgl-project/sglang/pull/8750 | Optimize prefill performance on cpu backend (#8750) | Optimize prefill performance on cpu backend (#8750) | sgl-kernel/csrc/cpu/qkv_proj.cpp | cpp | Apache-2.0 | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
// [NOTE]: Fused kernel for QKV projection with weight absorption and RoPE
//
// 1. `q_a_proj` and `kv_a_proj_with_mqa` fused into one gemm,
// otherwise we need to split IC for the 2nd gemm.
// 2. `q_a_layernorm` and `kv_a_layernorm` fused i... | #include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
// [NOTE]: Fused kernel for QKV projection with weight absorption and RoPE
//
// 1. `q_a_proj` and `kv_a_proj_with_mqa` fused into one gemm,
// otherwise we need to split IC for the 2nd gemm.
// 2. `q_a_layernorm` and `kv_a_layernorm` fused i... | --- sgl-kernel/csrc/cpu/qkv_proj.cpp
+++ sgl-kernel/csrc/cpu/qkv_proj.cpp
@@ -100,8 +100,7 @@
const int64_t NB1 = div_up(N1, BLOCK_N);
const int64_t NB = NB0 + NB1;
- // TODO: brgemm u8s8 depends on PyTorch 2.7 release.
- const bool use_brgemm = false;
+ const bool use_brgemm = can_use_brgemm<int8_t>(M);
... | 1 | 2 | 9 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/qkv_proj.cpp).
Goal: Optimize prefill performance on cpu backend
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "common.h"
#include "gemm.h"
#include "vec.h"
namespace {
// [NOTE]: Fused kernel for QKV projection with weig... | ## Motivation
This PR aims at improving prefill performance for CPU backend, it will improve `bfloat16`, `int8_w8a8` and `fp8_w8a8` paths. This one has no effective on decoding performance.
## Modifications
* **amx-int8**: enable **amx-int8** for GEMM and MoE kernels. As SGLang pops up with PyTorch 2.7, we are free ... | [{"scope": "kernel", "hardware": "Intel 4th gen Xeon (SPR)", "model": null, "shape": "M=1024,K=4096,N=14336", "batch_size": 1024, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 4.375, "optimized": 3.483, "stated_speedup": null, "speedup": 1.2561}, {"scope": "kernel", "hardware": "Intel 4th ge... | 1.3585 | null |
sglang | 0fc54b971e14211d54035b72af1a2af0dd93511e | 2025-08-18T04:09:49+08:00 | 9,272 | https://github.com/sgl-project/sglang/pull/9272 | [fix]: fix cutlass moe ut and and Opt H20 cutlass groupGemm performance (#9272) | [fix]: fix cutlass moe ut and and Opt H20 cutlass groupGemm performance (#9272)
Co-authored-by: wanghanpei <wanghanpei@bytedance.com> | sgl-kernel/csrc/moe/fp8_blockwise_moe_kernel.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cutlass/arch/arch.h>
#include <torch/all.h>
#include "cute/tensor.hpp"
#include "cutlass/cutlass.h"
#include "cutlass/epilogue/collective/collective_builder.hpp"
#include "cutlass/epilogue/collective/default_epilogue.hpp"
#include "cutlass/ep... | #include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cutlass/arch/arch.h>
#include <torch/all.h>
#include "cute/tensor.hpp"
#include "cutlass/cutlass.h"
#include "cutlass/epilogue/collective/collective_builder.hpp"
#include "cutlass/epilogue/collective/default_epilogue.hpp"
#include "cutlass/ep... | --- sgl-kernel/csrc/moe/fp8_blockwise_moe_kernel.cu
+++ sgl-kernel/csrc/moe/fp8_blockwise_moe_kernel.cu
@@ -437,6 +437,34 @@
}
}
+#define JOIN_STRUCT_PP_NAME(m, n, k, a, b, c) sm90_fp8_pp_config##_##m##_##n##_##k##_##a##_##b##_##c
+
+#define JOIN_STRUCT_CO_NAME(m, n, k, a, b, c) sm90_fp8_co_config##_##m##_##n##_#... | 109 | 35 | 2 | [] | [
"H20"
] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/moe/fp8_blockwise_moe_kernel.cu).
Goal: [fix]: fix cutlass moe ut and and Opt H20 cutlass groupGemm performance
Target hardware: H20
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
1. fix cutlass moe UT
2. add optimized cutlas... | [] | null | null |
sglang | 0fc54b971e14211d54035b72af1a2af0dd93511e | 2025-08-18T04:09:49+08:00 | 9,272 | https://github.com/sgl-project/sglang/pull/9272 | [fix]: fix cutlass moe ut and and Opt H20 cutlass groupGemm performance (#9272) | [fix]: fix cutlass moe ut and and Opt H20 cutlass groupGemm performance (#9272)
Co-authored-by: wanghanpei <wanghanpei@bytedance.com> | sgl-kernel/include/utils.h | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/include/utils.h
+++ sgl-kernel/include/utils.h
@@ -252,6 +252,25 @@
CHECK_CUDA_SUCCESS(cudaDeviceGetAttribute(&sm_major, cudaDevAttrComputeCapabilityMajor, device));
CHECK_CUDA_SUCCESS(cudaDeviceGetAttribute(&sm_minor, cudaDevAttrComputeCapabilityMinor, device));
return sm_major * 10 + sm_minor;... | 19 | 0 | 2 | [] | [
"H20"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/include/utils.h).
Goal: [fix]: fix cutlass moe ut and and Opt H20 cutlass groupGemm performance
Target hardware: H20
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.... | <!-- Thank you for your contribution! Please follow these guidelines to enhance your pull request. If anything is unclear, submit your PR and reach out to maintainers for assistance. Join our Slack community at https://slack.sglang.ai to discuss further. -->
## Motivation
1. fix cutlass moe UT
2. add optimized cutlas... | [] | null | null |
sglang | eff4eb3fdd81a82ac76ae16250d06c48fd7ab03e | 2025-08-15T22:08:11-07:00 | 7,667 | https://github.com/sgl-project/sglang/pull/7667 | Add fp4 quantize before all-gather for Flashinfer cutlass MoE DP (max throughput) (#7667) | Add fp4 quantize before all-gather for Flashinfer cutlass MoE DP (max throughput) (#7667) | python/sglang/srt/layers/dp_attention.py | triton | Apache-2.0 | from __future__ import annotations
import functools
import logging
from contextlib import contextmanager
from enum import IntEnum, auto
from typing import TYPE_CHECKING, List, Optional, Tuple
import torch
import triton
import triton.language as tl
from sglang.srt.distributed import (
GroupCoordinator,
get_te... | from __future__ import annotations
import functools
import logging
from contextlib import contextmanager
from enum import IntEnum, auto
from typing import TYPE_CHECKING, List, Optional, Tuple
import torch
import triton
import triton.language as tl
from sglang.srt.distributed import (
GroupCoordinator,
get_te... | --- python/sglang/srt/layers/dp_attention.py
+++ python/sglang/srt/layers/dp_attention.py
@@ -72,6 +72,7 @@
_device: torch.device
_global_dp_buffer_len: int
_local_dp_buffer_len: int
+ _global_num_tokens: Optional[List[int]]
@classmethod
def set_metadata(cls, hidden_size: int, dtype: torch... | 23 | 4 | 8 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/dp_attention.py).
Goal: Add fp4 quantize before all-gather for Flashinfer cutlass MoE DP (max throughput)
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from __future__ import annotations
import functools
import l... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
The goal of this PR is to optimize communications f... | [{"scope": "e2e", "hardware": null, "model": "nvidia/DeepSeek-R1-0528-FP4", "shape": null, "batch_size": 1024, "metric": "benchmark duration", "unit": "s", "higher_is_better": false, "baseline": 75.54, "optimized": 69.06, "stated_speedup": null, "speedup": 1.0938}, {"scope": "e2e", "hardware": null, "model": "nvidia/De... | null | 1.0938 |
sglang | eff4eb3fdd81a82ac76ae16250d06c48fd7ab03e | 2025-08-15T22:08:11-07:00 | 7,667 | https://github.com/sgl-project/sglang/pull/7667 | Add fp4 quantize before all-gather for Flashinfer cutlass MoE DP (max throughput) (#7667) | Add fp4 quantize before all-gather for Flashinfer cutlass MoE DP (max throughput) (#7667) | python/sglang/srt/layers/logits_processor.py | triton | Apache-2.0 | # Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writin... | # Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writin... | --- python/sglang/srt/layers/logits_processor.py
+++ python/sglang/srt/layers/logits_processor.py
@@ -191,7 +191,11 @@
else:
self.global_dp_buffer_len = self.global_dp_buffer_len
- set_dp_buffer_len(self.global_dp_buffer_len, self.dp_local_num_tokens)
+ set_dp_buffer_len(
+ ... | 5 | 1 | 8 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/logits_processor.py).
Goal: Add fp4 quantize before all-gather for Flashinfer cutlass MoE DP (max throughput)
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2023-2024 SGLang Team
# Licensed under the Ap... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
The goal of this PR is to optimize communications f... | [{"scope": "e2e", "hardware": null, "model": "nvidia/DeepSeek-R1-0528-FP4", "shape": null, "batch_size": 1024, "metric": "benchmark duration", "unit": "s", "higher_is_better": false, "baseline": 75.54, "optimized": 69.06, "stated_speedup": null, "speedup": 1.0938}, {"scope": "e2e", "hardware": null, "model": "nvidia/De... | null | 1.0938 |
sglang | 4fc09e0df0f018d74bbc99e16e11e9530d198b1e | 2025-08-15T01:46:16-07:00 | 8,777 | https://github.com/sgl-project/sglang/pull/8777 | Fp4 MOE quant kernel optimization (#8777) | Fp4 MOE quant kernel optimization (#8777)
Co-authored-by: Rain Jiang <96632942+rainj-me@users.noreply.github.com> | sgl-kernel/csrc/gemm/nvfp4_expert_quant.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cuda_fp8.h>
#include <cuda_runtime.h>
#include <torch/all.h>
template <typename T>
struct TypeConverter {
using Type = half2;
}; // keep for generality
template <>
struct TypeConverter<half2> {
using Type = half;
};
template <>
struct ... | #include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cuda_fp8.h>
#include <cuda_runtime.h>
#include <torch/all.h>
template <typename T>
struct TypeConverter {
using Type = half2;
}; // keep for generality
template <>
struct TypeConverter<half2> {
using Type = half;
};
template <>
struct ... | --- sgl-kernel/csrc/gemm/nvfp4_expert_quant.cu
+++ sgl-kernel/csrc/gemm/nvfp4_expert_quant.cu
@@ -240,10 +240,111 @@
}
// Use UE4M3 by default.
-template <class Type, bool UE8M0_SF = false>
+template <class Type, bool UE8M0_SF = false, bool SMALL_NUM_EXPERTS = false>
__global__ void
#if defined(__CUDA_ARCH__) && ... | 232 | 51 | 1 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/gemm/nvfp4_expert_quant.cu).
Goal: Fp4 MOE quant kernel optimization
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cuda_fp8.h>
#include <cuda_runtime... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Port vLLM's [FP4 MoE kernel optimization (PR #19500)]... | [{"scope": "kernel", "hardware": "NVIDIA B200", "model": null, "shape": null, "batch\u6e38\u620f\u526f\u672c_size": null, "metric": "Avg Time", "unit": "sec", "higher_is_better": false, "baseline": 68.22, "optimized": 59.17, "stated_speedup": 1.13, "speedup": 1.1529}] | 1.1529 | null |
sglang | 841810f227d5c8c9385749df811272eebc8e5845 | 2025-08-13T21:59:22-07:00 | 8,818 | https://github.com/sgl-project/sglang/pull/8818 | [Perf] Tunings for SM100 FP8 CUTLASS kernel (#8818) | [Perf] Tunings for SM100 FP8 CUTLASS kernel (#8818) | sgl-kernel/csrc/gemm/fp8_gemm_kernel.cu | cuda | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/gemm/fp8_gemm_kernel.cu
+++ sgl-kernel/csrc/gemm/fp8_gemm_kernel.cu
@@ -48,6 +48,7 @@
#include <cutlass/gemm/kernel/gemm_universal.hpp>
#include <cutlass/util/packed_stride.hpp>
+#include "math.hpp"
#include "utils.h"
using namespace cute;
@@ -1019,8 +1020,18 @@
const torch::Tensor& sca... | 126 | 24 | 1 | [] | [
"SM100"
] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/gemm/fp8_gemm_kernel.cu).
Goal: [Perf] Tunings for SM100 FP8 CUTLASS kernel
Target hardware: SM100
Target models: unspecified
Current file:
```cuda
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "Lice... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Port vLLM's
https://github.com/vllm-project/vllm/p... | [] | null | null |
sglang | 5c31b35db28e740c5cf4152e09e8d641057d8c03 | 2025-08-09T17:16:07-07:00 | 8,245 | https://github.com/sgl-project/sglang/pull/8245 | [hicache] Optimization for DMA copy (#8245) | [hicache] Optimization for DMA copy (#8245) | sgl-kernel/csrc/kvcacheio/transfer.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAException.h>
#include <c10/util/irange.h>
#include <cstdint>
#include "pytorch_extension_utils.h"
__device__ __forceinline__ void
transfer_item_warp(int32_t lane_id, const void* src_addr, void* dst_addr, int64_t item_size_bytes) {
// todo, different chunk s... | #include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAException.h>
#include <c10/util/irange.h>
#include <cstdint>
#include "pytorch_extension_utils.h"
__device__ __forceinline__ void
transfer_item_warp(int32_t lane_id, const void* src_addr, void* dst_addr, int64_t item_size_bytes) {
// todo, different chunk s... | --- sgl-kernel/csrc/kvcacheio/transfer.cu
+++ sgl-kernel/csrc/kvcacheio/transfer.cu
@@ -451,15 +451,33 @@
auto src_indices_cpu = src_indices.cpu();
auto dst_indices_cpu = dst_indices.cpu();
- const int64_t num_pages = src_indices_cpu.size(0) / page_size;
+ const auto num_indices = src_indices_cpu.numel();
... | 24 | 6 | 1 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/kvcacheio/transfer.cu).
Goal: [hicache] Optimization for DMA copy
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAException.h>
#include <c10/util/irange.h>
#include <cst... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Using `cudaMemcpy` for KV cache offloading is ineff... | [] | null | null |
sglang | 3f2e315f6e5beb210b981ba07c4da88f703ccc50 | 2025-08-09T15:29:12+08:00 | 8,962 | https://github.com/sgl-project/sglang/pull/8962 | optimize: reduce shulffle and quantization overhead in cutlass_moe sm90 (#8962) | optimize: reduce shulffle and quantization overhead in cutlass_moe sm90 (#8962)
Co-authored-by: 戚余航 <qiyuhang@bytedance.com> | python/sglang/srt/layers/quantization/fp8_kernel.py | triton | Apache-2.0 | # Copyright 2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, so... | # Copyright 2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, so... | --- python/sglang/srt/layers/quantization/fp8_kernel.py
+++ python/sglang/srt/layers/quantization/fp8_kernel.py
@@ -1356,3 +1356,62 @@
expert_tokens_alignment,
)
return a_q, sfa
+
+
+@triton.jit
+def _per_group_transpose(
+ data_ptr: torch.Tensor,
+ trans_data_ptr: torch.Tensor,
+ expert_off... | 59 | 0 | 2 | [] | [
"SM90"
] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/quantization/fp8_kernel.py).
Goal: optimize: reduce shulffle and quantization overhead in cutlass_moe sm90
Target hardware: SM90
Target models: unspecified
Current file:
```python
# Copyright 2024 SGLang Team
# Licensed under the Apache License, V... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
<!-- Explain the purpose of this PR and the goals i... | [] | null | null |
sglang | 0475448ee303a361d00783274d544ae10977a3ff | 2025-08-06T21:37:50+08:00 | 8,860 | https://github.com/sgl-project/sglang/pull/8860 | Optimize triton swa kernel by skipping computation (#8860) | Optimize triton swa kernel by skipping computation (#8860) | python/sglang/srt/layers/attention/triton_ops/extend_attention.py | triton | Apache-2.0 | # Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writin... | # Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writin... | --- python/sglang/srt/layers/attention/triton_ops/extend_attention.py
+++ python/sglang/srt/layers/attention/triton_ops/extend_attention.py
@@ -134,38 +134,6 @@
start_n = tl.multiple_of(start_n, BLOCK_N)
mask_n = (start_n + offs_n) < cur_seq_len_prefix
- offs_kv_loc = tl.load(
- kv... | 122 | 105 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/triton_ops/extend_attention.py).
Goal: Optimize triton swa kernel by skipping computation
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2023-2024 SGLang Team
# Licensed under the Apache Licen... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
For triton sliding window attention, when mask for ... | [] | null | null |
sglang | d9def43dcdfd961c4fb2d8eee72f30681268abe3 | 2025-08-03T12:13:47+08:00 | 8,722 | https://github.com/sgl-project/sglang/pull/8722 | [Perf]Use Cooperative Schedule for H100 & H200 & H800 in fp8_blockwise_scaled_grouped_mm (#8722) | [Perf]Use Cooperative Schedule for H100 & H200 & H800 in fp8_blockwise_scaled_grouped_mm (#8722) | sgl-kernel/csrc/moe/fp8_blockwise_moe_kernel.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cutlass/arch/arch.h>
#include <torch/all.h>
#include "cute/tensor.hpp"
#include "cutlass/cutlass.h"
#include "cutlass/epilogue/collective/collective_builder.hpp"
#include "cutlass/epilogue/collective/default_epilogue.hpp"
#include "cutlass/ep... | #include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cutlass/arch/arch.h>
#include <torch/all.h>
#include "cute/tensor.hpp"
#include "cutlass/cutlass.h"
#include "cutlass/epilogue/collective/collective_builder.hpp"
#include "cutlass/epilogue/collective/default_epilogue.hpp"
#include "cutlass/ep... | --- sgl-kernel/csrc/moe/fp8_blockwise_moe_kernel.cu
+++ sgl-kernel/csrc/moe/fp8_blockwise_moe_kernel.cu
@@ -485,7 +485,8 @@
torch::TensorOptions options_int = torch::TensorOptions().dtype(torch::kInt64).device(a.device());
torch::Tensor problem_sizes_transpose = torch::empty(num_experts * 3, options_int);
- if... | 3 | 2 | 1 | [] | [
"H100",
"H200",
"H800"
] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/moe/fp8_blockwise_moe_kernel.cu).
Goal: [Perf]Use Cooperative Schedule for H100 & H200 & H800 in fp8_blockwise_scaled_grouped_mm
Target hardware: H100, H200, H800
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#include <c... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Fix https://github.com/sgl-project/sglang/pull/7278#... | [] | null | null |
sglang | 7a4309cc8a56e7a2cffba82a5189b51fd5776259 | 2025-07-29T23:31:54+08:00 | 8,499 | https://github.com/sgl-project/sglang/pull/8499 | [sgl-kernel performace] fix fp8 quant kernels dispatch __nv_fp8_e4m3 bug to improve performance 10%-20% (#8499) | [sgl-kernel performace] fix fp8 quant kernels dispatch __nv_fp8_e4m3 bug to improve performance 10%-20% (#8499)
Co-authored-by: Ke Bao <ispobaoke@gmail.com> | sgl-kernel/csrc/gemm/per_tensor_quant_fp8.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <c10/util/Float8_e4m3fn.h>
#include <cmath>
#include <cub/block/block_reduce.cuh>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
template <typename T>
__global__ void
per_tensor_absmax_kernel(const T* __restrict__ input, float* __restrict__ output_s, const int64_t... | #include <ATen/cuda/CUDAContext.h>
#include <c10/util/Float8_e4m3fn.h>
#include <cmath>
#include <cub/block/block_reduce.cuh>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
template <typename T>
__global__ void
per_tensor_absmax_kernel(const T* __restrict__ input, float* __restrict__ output_s, const int64_t... | --- sgl-kernel/csrc/gemm/per_tensor_quant_fp8.cu
+++ sgl-kernel/csrc/gemm/per_tensor_quant_fp8.cu
@@ -44,10 +44,10 @@
}
}
-template <typename T>
+template <typename T, typename DST_DTYPE>
__global__ void per_tensor_quant_fp8_kernel(
const T* __restrict__ input,
- FP8_TYPE* __restrict__ output,
+ DST_D... | 7 | 7 | 3 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/gemm/per_tensor_quant_fp8.cu).
Goal: [sgl-kernel performace] fix fp8 quant kernels dispatch __nv_fp8_e4m3 bug to improve performance 10%-20%
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#inc... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Follow: https://github.com/sgl-project/sglang/pull/... | [{"scope": "kernel", "hardware": "B200", "model": null, "shape": "seq_len=64", "batch_size": 16, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 25.792001, "optimized": 23.968, "stated_speedup": null, "speedup": 1.0761}, {"scope": "kernel", "hardware": "B200", "model": null, "shape": "seq_len=... | 1.2544 | 1.4431 |
sglang | 7a4309cc8a56e7a2cffba82a5189b51fd5776259 | 2025-07-29T23:31:54+08:00 | 8,499 | https://github.com/sgl-project/sglang/pull/8499 | [sgl-kernel performace] fix fp8 quant kernels dispatch __nv_fp8_e4m3 bug to improve performance 10%-20% (#8499) | [sgl-kernel performace] fix fp8 quant kernels dispatch __nv_fp8_e4m3 bug to improve performance 10%-20% (#8499)
Co-authored-by: Ke Bao <ispobaoke@gmail.com> | sgl-kernel/csrc/gemm/per_token_group_quant_8bit.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <c10/util/Float8_e4m3fn.h>
#include <cuda_fp8.h>
#include <cmath>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
__device__ __forceinline__ float GroupReduceMax(float val, const int tid) {
unsigned mask = 0xffff;
val = fmaxf(val, __shfl_xor_sync(mask, val, 8)... | #include <ATen/cuda/CUDAContext.h>
#include <cuda_fp8.h>
#include <cmath>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
__device__ __forceinline__ float GroupReduceMax(float val, const int tid) {
unsigned mask = 0xffff;
val = fmaxf(val, __shfl_xor_sync(mask, val, 8));
val = fmaxf(val, __shfl_xor_syn... | --- sgl-kernel/csrc/gemm/per_token_group_quant_8bit.cu
+++ sgl-kernel/csrc/gemm/per_token_group_quant_8bit.cu
@@ -1,5 +1,4 @@
#include <ATen/cuda/CUDAContext.h>
-#include <c10/util/Float8_e4m3fn.h>
#include <cuda_fp8.h>
#include <cmath> | 0 | 1 | 3 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/gemm/per_token_group_quant_8bit.cu).
Goal: [sgl-kernel performace] fix fp8 quant kernels dispatch __nv_fp8_e4m3 bug to improve performance 10%-20%
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Follow: https://github.com/sgl-project/sglang/pull/... | [{"scope": "kernel", "hardware": "B200", "model": null, "shape": "seq_len=64", "batch_size": 16, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 25.792001, "optimized": 23.968, "stated_speedup": null, "speedup": 1.0761}, {"scope": "kernel", "hardware": "B200", "model": null, "shape": "seq_len=... | 1.2544 | 1.4431 |
sglang | 7a4309cc8a56e7a2cffba82a5189b51fd5776259 | 2025-07-29T23:31:54+08:00 | 8,499 | https://github.com/sgl-project/sglang/pull/8499 | [sgl-kernel performace] fix fp8 quant kernels dispatch __nv_fp8_e4m3 bug to improve performance 10%-20% (#8499) | [sgl-kernel performace] fix fp8 quant kernels dispatch __nv_fp8_e4m3 bug to improve performance 10%-20% (#8499)
Co-authored-by: Ke Bao <ispobaoke@gmail.com> | sgl-kernel/csrc/gemm/per_token_quant_fp8.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <cmath>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
static constexpr int kWarpSize = 32;
// ---------------------------------------------------------------------------
// 1. Warp‑local, no shared memory
// • One warp handles one token.
// • Eight tokens ... | #include <ATen/cuda/CUDAContext.h>
#include <cmath>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
static constexpr int kWarpSize = 32;
// ---------------------------------------------------------------------------
// 1. Warp‑local, no shared memory
// • One warp handles one token.
// • Eight tokens ... | --- sgl-kernel/csrc/gemm/per_token_quant_fp8.cu
+++ sgl-kernel/csrc/gemm/per_token_quant_fp8.cu
@@ -12,10 +12,10 @@
// • One warp handles one token.
// • Eight tokens per 256‑thread CTA.
// ---------------------------------------------------------------------------
-template <typename T, int kTokensPerCTA = 8,... | 14 | 15 | 3 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/gemm/per_token_quant_fp8.cu).
Goal: [sgl-kernel performace] fix fp8 quant kernels dispatch __nv_fp8_e4m3 bug to improve performance 10%-20%
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#inc... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Follow: https://github.com/sgl-project/sglang/pull/... | [{"scope": "kernel", "hardware": "B200", "model": null, "shape": "seq_len=64", "batch_size": 16, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 25.792001, "optimized": 23.968, "stated_speedup": null, "speedup": 1.0761}, {"scope": "kernel", "hardware": "B200", "model": null, "shape": "seq_len=... | 1.2544 | 1.4431 |
sglang | a3398d8478e981659c4d045635e62c8066ca6bae | 2025-07-07T09:20:30+08:00 | 7,794 | https://github.com/sgl-project/sglang/pull/7794 | Optimize moe align block size kernel (#7794) | Optimize moe align block size kernel (#7794) | sgl-kernel/csrc/moe/moe_align_kernel.cu | cuda | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/moe/moe_align_kernel.cu
+++ sgl-kernel/csrc/moe/moe_align_kernel.cu
@@ -21,16 +21,10 @@
#include "utils.h"
-template <typename T, int N, int Alignment = sizeof(T) * N>
-class alignas(Alignment) AlignedArray {
- public:
- T data[N];
-};
-
#define WARP_SIZE 32
#define VEC_SIZE 4
-using Vec =... | 108 | 77 | 2 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/moe/moe_align_kernel.cu).
Goal: Optimize moe align block size kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "Licen... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
- Optimize prefix sum with Blelloch scan (O(logN))
... | [] | null | null |
sglang | a3398d8478e981659c4d045635e62c8066ca6bae | 2025-07-07T09:20:30+08:00 | 7,794 | https://github.com/sgl-project/sglang/pull/7794 | Optimize moe align block size kernel (#7794) | Optimize moe align block size kernel (#7794) | sgl-kernel/include/utils.h | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/include/utils.h
+++ sgl-kernel/include/utils.h
@@ -363,3 +363,9 @@
}
return tensor_padded;
}
+
+// Get the next power of 2 of a number
+inline uint32_t next_pow2(uint32_t x) noexcept {
+ if (x <= 1) return 1;
+ return 1u << (32 - __builtin_clz(x - 1));
+} | 6 | 0 | 2 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/include/utils.h).
Goal: Optimize moe align block size kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may n... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
- Optimize prefix sum with Blelloch scan (O(logN))
... | [] | null | null |
sglang | 8e9fb43d825508dfacd6b970f19337eb2f755a7a | 2025-07-05T13:25:49+08:00 | 7,782 | https://github.com/sgl-project/sglang/pull/7782 | Optimize Hopper CUTLASS FP8 Blockwise Grouped GEMM Kernel in Small K Scenario (#7782) | Optimize Hopper CUTLASS FP8 Blockwise Grouped GEMM Kernel in Small K Scenario (#7782) | sgl-kernel/csrc/moe/fp8_blockwise_moe_kernel.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cutlass/arch/arch.h>
#include <torch/all.h>
#include "cute/tensor.hpp"
#include "cutlass/cutlass.h"
#include "cutlass/epilogue/collective/collective_builder.hpp"
#include "cutlass/epilogue/collective/default_epilogue.hpp"
#include "cutlass/ep... | #include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <cutlass/arch/arch.h>
#include <torch/all.h>
#include "cute/tensor.hpp"
#include "cutlass/cutlass.h"
#include "cutlass/epilogue/collective/collective_builder.hpp"
#include "cutlass/epilogue/collective/default_epilogue.hpp"
#include "cutlass/ep... | --- sgl-kernel/csrc/moe/fp8_blockwise_moe_kernel.cu
+++ sgl-kernel/csrc/moe/fp8_blockwise_moe_kernel.cu
@@ -61,7 +61,12 @@
using ArchTag = cutlass::arch::Sm90;
using OperatorClass = cutlass::arch::OpClassTensorOp;
- using FusionOperation = cutlass::epilogue::fusion::LinearCombination<ElementD, ElementAccumulat... | 86 | 38 | 1 | [] | [
"HOPPER"
] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/moe/fp8_blockwise_moe_kernel.cu).
Goal: Optimize Hopper CUTLASS FP8 Blockwise Grouped GEMM Kernel in Small K Scenario
Target hardware: HOPPER
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Follow https://github.com/sgl-project/sglang/pull/72... | [] | null | null |
sglang | 2998c4bdf4fea00081978b320b60f4b0ce905297 | 2025-07-04T03:42:44+08:00 | 7,744 | https://github.com/sgl-project/sglang/pull/7744 | [optimize] fuse renormalize into moe_topk_softmax (#7744) | [optimize] fuse renormalize into moe_topk_softmax (#7744)
Co-authored-by: ispobock <ispobaoke@gmail.com> | sgl-kernel/csrc/common_extension.cc | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/common_extension.cc
+++ sgl-kernel/csrc/common_extension.cc
@@ -169,9 +169,7 @@
"pad_sorted_token_ids) -> ()");
m.impl("moe_align_block_size", torch::kCUDA, &moe_align_block_size);
- m.def(
- "topk_softmax(Tensor! topk_weights, Tensor! topk_indices, Tensor! "
- "token_expert_i... | 1 | 3 | 5 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/common_extension.cc).
Goal: [optimize] fuse renormalize into moe_topk_softmax
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
fuse normalize into moe took softmax kernel.
This P... | [] | null | null |
sglang | 2998c4bdf4fea00081978b320b60f4b0ce905297 | 2025-07-04T03:42:44+08:00 | 7,744 | https://github.com/sgl-project/sglang/pull/7744 | [optimize] fuse renormalize into moe_topk_softmax (#7744) | [optimize] fuse renormalize into moe_topk_softmax (#7744)
Co-authored-by: ispobock <ispobaoke@gmail.com> | sgl-kernel/csrc/moe/moe_topk_softmax_kernels.cu | cuda | Apache-2.0 | // Adapt from https://github.com/vllm-project/vllm/blob/v0.7.3/csrc/moe/topk_softmax_kernels.cu
// which is originally adapted from
// https://github.com/NVIDIA/TensorRT-LLM/blob/v0.7.1/cpp/tensorrt_llm/kernels/mixtureOfExperts/moe_kernels.cu
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apach... | // Adapt from https://github.com/vllm-project/vllm/blob/v0.7.3/csrc/moe/topk_softmax_kernels.cu
// which is originally adapted from
// https://github.com/NVIDIA/TensorRT-LLM/blob/v0.7.1/cpp/tensorrt_llm/kernels/mixtureOfExperts/moe_kernels.cu
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apach... | --- sgl-kernel/csrc/moe/moe_topk_softmax_kernels.cu
+++ sgl-kernel/csrc/moe/moe_topk_softmax_kernels.cu
@@ -41,15 +41,29 @@
/// Alignment requirement in bytes
int Alignment = sizeof(T) * N>
class alignas(Alignment) AlignedArray {
- float data[N];
+ T data[N];
};
+
+// ========================== Util funct... | 161 | 85 | 5 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/moe/moe_topk_softmax_kernels.cu).
Goal: [optimize] fuse renormalize into moe_topk_softmax
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
// Adapt from https://github.com/vllm-project/vllm/blob/v0.7.3/csrc/moe/topk_softmax_kerne... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
fuse normalize into moe took softmax kernel.
This P... | [] | null | null |
sglang | 2998c4bdf4fea00081978b320b60f4b0ce905297 | 2025-07-04T03:42:44+08:00 | 7,744 | https://github.com/sgl-project/sglang/pull/7744 | [optimize] fuse renormalize into moe_topk_softmax (#7744) | [optimize] fuse renormalize into moe_topk_softmax (#7744)
Co-authored-by: ispobock <ispobaoke@gmail.com> | sgl-kernel/csrc/torch_extension_rocm.cc | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/torch_extension_rocm.cc
+++ sgl-kernel/csrc/torch_extension_rocm.cc
@@ -63,9 +63,7 @@
"pad_sorted_token_ids) -> ()");
m.impl("moe_align_block_size", torch::kCUDA, &moe_align_block_size);
- m.def(
- "topk_softmax(Tensor! topk_weights, Tensor! topk_indices, Tensor! "
- "token_ex... | 1 | 3 | 5 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/torch_extension_rocm.cc).
Goal: [optimize] fuse renormalize into moe_topk_softmax
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 ... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
fuse normalize into moe took softmax kernel.
This P... | [] | null | null |
sglang | 2998c4bdf4fea00081978b320b60f4b0ce905297 | 2025-07-04T03:42:44+08:00 | 7,744 | https://github.com/sgl-project/sglang/pull/7744 | [optimize] fuse renormalize into moe_topk_softmax (#7744) | [optimize] fuse renormalize into moe_topk_softmax (#7744)
Co-authored-by: ispobock <ispobaoke@gmail.com> | sgl-kernel/include/sgl_kernel_ops.h | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/include/sgl_kernel_ops.h
+++ sgl-kernel/include/sgl_kernel_ops.h
@@ -222,10 +222,7 @@
bool pad_sorted_token_ids);
void topk_softmax(
- torch::Tensor& topk_weights,
- torch::Tensor& topk_indices,
- torch::Tensor& token_expert_indices,
- torch::Tensor& gating_output);
+ torch::Tensor... | 1 | 4 | 5 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/include/sgl_kernel_ops.h).
Goal: [optimize] fuse renormalize into moe_topk_softmax
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
fuse normalize into moe took softmax kernel.
This P... | [] | null | null |
sglang | 18efb5e8e0ed467f6dc42680d88787f5ed6c074e | 2025-06-09T10:37:34+08:00 | 6,929 | https://github.com/sgl-project/sglang/pull/6929 | [perf][sgl-kernel] extend cutlass_mla_decode to support num_head < 128 (#6929) | [perf][sgl-kernel] extend cutlass_mla_decode to support num_head < 128 (#6929) | sgl-kernel/csrc/attention/cutlass_mla_kernel.cu | cuda | Apache-2.0 | /*
Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/... | /*
Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/... | --- sgl-kernel/csrc/attention/cutlass_mla_kernel.cu
+++ sgl-kernel/csrc/attention/cutlass_mla_kernel.cu
@@ -22,8 +22,9 @@
#include <torch/all.h>
#include <cute/tensor.hpp>
-#include <device/sm100_mla.hpp>
-#include <kernel/sm100_mla_tile_scheduler.hpp>
+
+#include "cutlass_sm100_mla/device/sm100_mla.hpp"
+#include ... | 54 | 24 | 4 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/attention/cutlass_mla_kernel.cu).
Goal: [perf][sgl-kernel] extend cutlass_mla_decode to support num_head < 128
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
/*
Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
Copy... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
An inelegant workaround, thus the kernel have to be... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "heads=128", "batch_size": 1, "metric": "bandwidth", "unit": "GB/s", "higher_is_better": true, "baseline": null, "optimized": 18.775741, "stated_speedup": 1.8, "speedup": 1.8}, {"scope": "kernel", "hardware": null, "model": null, "shape": "heads=64", "batch... | 1.8 | null |
sglang | 18efb5e8e0ed467f6dc42680d88787f5ed6c074e | 2025-06-09T10:37:34+08:00 | 6,929 | https://github.com/sgl-project/sglang/pull/6929 | [perf][sgl-kernel] extend cutlass_mla_decode to support num_head < 128 (#6929) | [perf][sgl-kernel] extend cutlass_mla_decode to support num_head < 128 (#6929) | sgl-kernel/csrc/common_extension.cc | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/common_extension.cc
+++ sgl-kernel/csrc/common_extension.cc
@@ -60,7 +60,7 @@
m.impl("merge_state_v2", torch::kCUDA, &merge_state_v2);
m.def(
"cutlass_mla_decode(Tensor! out, Tensor q_nope_and_q_pe, Tensor kv_c_and_k_pe_cache, Tensor seq_lens, Tensor "
- "page_table, Tensor workspa... | 1 | 1 | 4 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/common_extension.cc).
Goal: [perf][sgl-kernel] extend cutlass_mla_decode to support num_head < 128
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache Lice... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
An inelegant workaround, thus the kernel have to be... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "heads=128", "batch_size": 1, "metric": "bandwidth", "unit": "GB/s", "higher_is_better": true, "baseline": null, "optimized": 18.775741, "stated_speedup": 1.8, "speedup": 1.8}, {"scope": "kernel", "hardware": null, "model": null, "shape": "heads=64", "batch... | 1.8 | null |
sglang | 18efb5e8e0ed467f6dc42680d88787f5ed6c074e | 2025-06-09T10:37:34+08:00 | 6,929 | https://github.com/sgl-project/sglang/pull/6929 | [perf][sgl-kernel] extend cutlass_mla_decode to support num_head < 128 (#6929) | [perf][sgl-kernel] extend cutlass_mla_decode to support num_head < 128 (#6929) | sgl-kernel/include/sgl_kernel_ops.h | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/include/sgl_kernel_ops.h
+++ sgl-kernel/include/sgl_kernel_ops.h
@@ -109,8 +109,10 @@
torch::Tensor const& kv_c_and_k_pe_cache,
torch::Tensor const& seq_lens,
torch::Tensor const& page_table,
- torch::Tensor const& workspace);
-int64_t cutlass_mla_get_workspace_size(int64_t max_seq_len,... | 4 | 2 | 4 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/include/sgl_kernel_ops.h).
Goal: [perf][sgl-kernel] extend cutlass_mla_decode to support num_head < 128
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache Lice... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
An inelegant workaround, thus the kernel have to be... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "heads=128", "batch_size": 1, "metric": "bandwidth", "unit": "GB/s", "higher_is_better": true, "baseline": null, "optimized": 18.775741, "stated_speedup": 1.8, "speedup": 1.8}, {"scope": "kernel", "hardware": null, "model": null, "shape": "heads=64", "batch... | 1.8 | null |
sglang | 2a413829f42b8e8433a3e7cfd91cc9cb241cfbc0 | 2025-06-07T17:43:50+08:00 | 5,955 | https://github.com/sgl-project/sglang/pull/5955 | Add triton version as a fused_moe_triton config search key to avoid performace decrease in different Triton version (#5955) | Add triton version as a fused_moe_triton config search key to avoid performace decrease in different Triton version (#5955) | python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py | triton | Apache-2.0 | # Adapted from https://github.com/vllm-project/vllm/blob/a6221a144af772fd1a68fe7e627935dc53e81738/vllm/model_executor/layers/fused_moe/fused_moe.py
"""Fused MoE kernel."""
import functools
import json
import logging
import os
from typing import Any, Callable, Dict, List, Optional, Tuple
import torch
import triton
im... | # Adapted from https://github.com/vllm-project/vllm/blob/a6221a144af772fd1a68fe7e627935dc53e81738/vllm/model_executor/layers/fused_moe/fused_moe.py
"""Fused MoE kernel."""
import functools
import json
import logging
import os
from typing import Any, Callable, Dict, List, Optional, Tuple
import torch
import triton
im... | --- python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py
+++ python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py
@@ -936,8 +936,15 @@
# directory
json_file_name = get_config_file_name(E, N, dtype, [block_n, block_k])
+ # We found that using the fused_moe_kernel config from Triton 3.1.0 with ... | 8 | 1 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py).
Goal: Add triton version as a fused_moe_triton config search key to avoid performace decrease in different Triton version
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Adapted... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Follow https://github.com/sgl-project/sglang/pull/5... | [] | null | null |
sglang | b819381feca4467f803dce1cb01533987c270003 | 2025-06-05T23:00:18-07:00 | 6,838 | https://github.com/sgl-project/sglang/pull/6838 | AITER backend extension and workload optimizations (#6838) | AITER backend extension and workload optimizations (#6838)
Co-authored-by: wunhuang <wunhuang@amd.com>
Co-authored-by: Hubert Lu <Hubert.Lu@amd.com> | python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py | triton | Apache-2.0 | # Adapted from https://github.com/vllm-project/vllm/blob/a6221a144af772fd1a68fe7e627935dc53e81738/vllm/model_executor/layers/fused_moe/fused_moe.py
"""Fused MoE kernel."""
import functools
import json
import logging
import os
from typing import Any, Callable, Dict, List, Optional, Tuple
import torch
import triton
im... | # Adapted from https://github.com/vllm-project/vllm/blob/a6221a144af772fd1a68fe7e627935dc53e81738/vllm/model_executor/layers/fused_moe/fused_moe.py
"""Fused MoE kernel."""
import functools
import json
import logging
import os
from typing import Any, Callable, Dict, List, Optional, Tuple
import torch
import triton
im... | --- python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py
+++ python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py
@@ -1332,7 +1332,7 @@
if (
not (use_fp8_w8a8 or use_int8_w8a8)
or block_shape is not None
- or (_is_hip and get_bool_env_var("SGLANG_AITER_MOE"))
+ or (_is_h... | 1 | 1 | 6 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py).
Goal: AITER backend extension and workload optimizations
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Adapted from https://github.com/vllm-project/vllm/blob/a6221a144af772fd1... | `Co-author: @kkHuang-amd `
<!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
- DeepSeek optimization
... | [] | null | null |
sglang | ff00895c46a4549f6c4279b1f8de24c05f1fa7ef | 2025-06-03T08:37:34+08:00 | 6,456 | https://github.com/sgl-project/sglang/pull/6456 | Add CPU optimized kernels for topk and rope fusions (#6456) | Add CPU optimized kernels for topk and rope fusions (#6456) | sgl-kernel/csrc/cpu/norm.cpp | cpp | Apache-2.0 | #include "common.h"
#include "vec.h"
namespace {
// NB: avoid using `at::vec::map<>` on bfloat16 or half
template <typename scalar_t>
void rmsnorm_kernel_impl(
scalar_t* __restrict__ output,
const scalar_t* __restrict__ input,
const scalar_t* __restrict__ weight,
int64_t batch_size,
int64_t hidden... | #include "common.h"
#include "vec.h"
namespace {
// NB: avoid using `at::vec::map<>` on bfloat16 or half
// Llama4TextL2Norm
template <typename scalar_t>
void l2norm_kernel_impl(
scalar_t* __restrict__ output,
const scalar_t* __restrict__ input,
int64_t batch_size,
int64_t hidden_size,
float eps =... | --- sgl-kernel/csrc/cpu/norm.cpp
+++ sgl-kernel/csrc/cpu/norm.cpp
@@ -4,6 +4,67 @@
namespace {
// NB: avoid using `at::vec::map<>` on bfloat16 or half
+// Llama4TextL2Norm
+template <typename scalar_t>
+void l2norm_kernel_impl(
+ scalar_t* __restrict__ output,
+ const scalar_t* __restrict__ input,
+ int64_... | 77 | 0 | 4 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/norm.cpp).
Goal: Add CPU optimized kernels for topk and rope fusions
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "common.h"
#include "vec.h"
namespace {
// NB: avoid using `at::vec::map<>` on bfloat16 or half
template <... | ## Overview:
This PR is adding the following CPU optimized sgl-kernels that are at least used in Qwen3/LLama1-4 models:
```
- TopK fusions:
TopK+sigmoid
softmax+TopK
- Norm fusion:
L2norm
- RoPE fusions:
origin rope fusion (gpt_neox style and gptj style)
```
| [] | null | null |
sglang | ff00895c46a4549f6c4279b1f8de24c05f1fa7ef | 2025-06-03T08:37:34+08:00 | 6,456 | https://github.com/sgl-project/sglang/pull/6456 | Add CPU optimized kernels for topk and rope fusions (#6456) | Add CPU optimized kernels for topk and rope fusions (#6456) | sgl-kernel/csrc/cpu/rope.cpp | cpp | Apache-2.0 | #include "common.h"
#include "vec.h"
namespace {
template <typename scalar_t>
void rope_kernel_impl(
scalar_t* __restrict__ q_pe_out,
scalar_t* __restrict__ k_pe_out,
int64_t* __restrict__ t_pos,
scalar_t* __restrict__ q_pe,
scalar_t* __restrict__ k_pe,
scalar_t* __restrict__ t_emb_pos,
in... | #include "common.h"
#include "vec.h"
namespace {
template <typename scalar_t>
void rotary_embedding_3D_kernel_impl(
scalar_t* __restrict__ query_out,
scalar_t* __restrict__ key_out,
int64_t* __restrict__ positions,
scalar_t* __restrict__ query,
scalar_t* __restrict__ key,
scalar_t* __restrict_... | --- sgl-kernel/csrc/cpu/rope.cpp
+++ sgl-kernel/csrc/cpu/rope.cpp
@@ -4,126 +4,343 @@
namespace {
template <typename scalar_t>
-void rope_kernel_impl(
- scalar_t* __restrict__ q_pe_out,
- scalar_t* __restrict__ k_pe_out,
- int64_t* __restrict__ t_pos,
- scalar_t* __restrict__ q_pe,
- scalar_t* __rest... | 310 | 93 | 4 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/rope.cpp).
Goal: Add CPU optimized kernels for topk and rope fusions
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "common.h"
#include "vec.h"
namespace {
template <typename scalar_t>
void rope_kernel_impl(
scalar_t* ... | ## Overview:
This PR is adding the following CPU optimized sgl-kernels that are at least used in Qwen3/LLama1-4 models:
```
- TopK fusions:
TopK+sigmoid
softmax+TopK
- Norm fusion:
L2norm
- RoPE fusions:
origin rope fusion (gpt_neox style and gptj style)
```
| [] | null | null |
sglang | ff00895c46a4549f6c4279b1f8de24c05f1fa7ef | 2025-06-03T08:37:34+08:00 | 6,456 | https://github.com/sgl-project/sglang/pull/6456 | Add CPU optimized kernels for topk and rope fusions (#6456) | Add CPU optimized kernels for topk and rope fusions (#6456) | sgl-kernel/csrc/cpu/topk.cpp | cpp | Apache-2.0 | #include "common.h"
#include "vec.h"
namespace {
template <typename scalar_t, int SIZE>
inline void softmax(float* __restrict__ out, const scalar_t* __restrict__ input) {
using bVec = at::vec::Vectorized<scalar_t>;
using fVec = at::vec::Vectorized<float>;
constexpr int kVecSize = bVec::size();
// step 1: ge... | #include "common.h"
#include "vec.h"
namespace {
template <typename scalar_t, int SIZE>
inline void softmax(float* __restrict__ out, const scalar_t* __restrict__ input) {
using bVec = at::vec::Vectorized<scalar_t>;
using fVec = at::vec::Vectorized<float>;
constexpr int kVecSize = bVec::size();
// step 1: ge... | --- sgl-kernel/csrc/cpu/topk.cpp
+++ sgl-kernel/csrc/cpu/topk.cpp
@@ -157,6 +157,101 @@
}
}
+template <typename scalar_t, int NUM_EXPERTS>
+void topk_sigmoid_kernel_impl(
+ float* __restrict__ topk_weights,
+ int32_t* __restrict__ topk_ids,
+ const scalar_t* __restrict__ gating_output,
+ int64_t num_t... | 221 | 0 | 4 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/topk.cpp).
Goal: Add CPU optimized kernels for topk and rope fusions
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "common.h"
#include "vec.h"
namespace {
template <typename scalar_t, int SIZE>
inline void softmax(float* ... | ## Overview:
This PR is adding the following CPU optimized sgl-kernels that are at least used in Qwen3/LLama1-4 models:
```
- TopK fusions:
TopK+sigmoid
softmax+TopK
- Norm fusion:
L2norm
- RoPE fusions:
origin rope fusion (gpt_neox style and gptj style)
```
| [] | null | null |
sglang | ff00895c46a4549f6c4279b1f8de24c05f1fa7ef | 2025-06-03T08:37:34+08:00 | 6,456 | https://github.com/sgl-project/sglang/pull/6456 | Add CPU optimized kernels for topk and rope fusions (#6456) | Add CPU optimized kernels for topk and rope fusions (#6456) | sgl-kernel/csrc/cpu/torch_extension_cpu.cpp | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/cpu/torch_extension_cpu.cpp
+++ sgl-kernel/csrc/cpu/torch_extension_cpu.cpp
@@ -23,6 +23,9 @@
// silu_and_mul
at::Tensor silu_and_mul_cpu(at::Tensor& input);
+// l2norm
+at::Tensor l2norm_cpu(at::Tensor& input, double eps);
+
// rmsnorm
at::Tensor rmsnorm_cpu(at::Tensor& input, at::Tensor& wei... | 25 | 4 | 4 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/torch_extension_cpu.cpp).
Goal: Add CPU optimized kernels for topk and rope fusions
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (t... | ## Overview:
This PR is adding the following CPU optimized sgl-kernels that are at least used in Qwen3/LLama1-4 models:
```
- TopK fusions:
TopK+sigmoid
softmax+TopK
- Norm fusion:
L2norm
- RoPE fusions:
origin rope fusion (gpt_neox style and gptj style)
```
| [] | null | null |
sglang | c087ddd6865a52634326a05af66429cb5531cd16 | 2025-05-28T15:15:23+08:00 | 6,627 | https://github.com/sgl-project/sglang/pull/6627 | Refine pre_reorder_triton_kernel slightly to improve performance (#6627) | Refine pre_reorder_triton_kernel slightly to improve performance (#6627)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com> | python/sglang/srt/layers/moe/ep_moe/kernels.py | triton | Apache-2.0 | import logging
from typing import List, Optional
import torch
import triton
from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
from sglang.srt.utils import dispose_tensor, is_cuda
logger = logging.getLogger(__name__)
_is_cuda = is_cuda()
if _is_cuda:
from sglang.srt.layers.quantizat... | import logging
from typing import List, Optional
import torch
import triton
from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
from sglang.srt.utils import dispose_tensor, is_cuda
logger = logging.getLogger(__name__)
_is_cuda = is_cuda()
if _is_cuda:
from sglang.srt.layers.quantizat... | --- python/sglang/srt/layers/moe/ep_moe/kernels.py
+++ python/sglang/srt/layers/moe/ep_moe/kernels.py
@@ -184,8 +184,10 @@
src_idx = tl.program_id(0)
src2dst_ptr = src2dst_ptr + src_idx * topk
topk_ids_ptr = topk_ids_ptr + src_idx * topk
-
src_ptr = input_ptr + src_idx * hidden_size
+
+ vec = tl.... | 9 | 4 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/ep_moe/kernels.py).
Goal: Refine pre_reorder_triton_kernel slightly to improve performance
Target hardware: unspecified
Target models: unspecified
Current file:
```python
import logging
from typing import List, Optional
import torch
import tr... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
In ep_moe kernel _pre_reorder_triton_kernel_ and _p... | [{"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": null, "metric": "performance", "unit": null, "higher_is_better": true, "baseline": null, "optimized": null, "stated_speedup": 1.1, "speedup": 1.1}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": null, "... | 1.125 | null |
sglang | 25c83fff6a80d9e3d2749f2ead122f96fdc127e9 | 2025-05-12T02:36:29-04:00 | 5,558 | https://github.com/sgl-project/sglang/pull/5558 | Performing Vocabulary Parallelism for LM Head across Attention TP Groups (#5558) | Performing Vocabulary Parallelism for LM Head across Attention TP Groups (#5558)
Co-authored-by: liusy58 <liusy58@linux.alibaba.com> | python/sglang/srt/layers/dp_attention.py | triton | Apache-2.0 | from __future__ import annotations
import functools
import logging
from contextlib import contextmanager
from typing import TYPE_CHECKING, List
import torch
import triton
import triton.language as tl
from sglang.srt.distributed import (
GroupCoordinator,
get_tensor_model_parallel_world_size,
get_tp_group... | from __future__ import annotations
import functools
import logging
from contextlib import contextmanager
from typing import TYPE_CHECKING, List
import torch
import triton
import triton.language as tl
from sglang.srt.distributed import (
GroupCoordinator,
get_tensor_model_parallel_world_size,
get_tp_group... | --- python/sglang/srt/layers/dp_attention.py
+++ python/sglang/srt/layers/dp_attention.py
@@ -252,12 +252,12 @@
)
-def tp_reduce_scatter(
+def attn_tp_reduce_scatter(
output: torch.Tensor,
input_list: List[torch.Tensor],
):
return get_attention_tp_group().reduce_scatter(output, input_list)
... | 2 | 2 | 3 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/dp_attention.py).
Goal: Performing Vocabulary Parallelism for LM Head across Attention TP Groups
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from __future__ import annotations
import functools
import logging
fr... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
<!-- Explain the purpose of this PR and the goals i... | [] | null | null |
sglang | 25c83fff6a80d9e3d2749f2ead122f96fdc127e9 | 2025-05-12T02:36:29-04:00 | 5,558 | https://github.com/sgl-project/sglang/pull/5558 | Performing Vocabulary Parallelism for LM Head across Attention TP Groups (#5558) | Performing Vocabulary Parallelism for LM Head across Attention TP Groups (#5558)
Co-authored-by: liusy58 <liusy58@linux.alibaba.com> | python/sglang/srt/layers/logits_processor.py | triton | Apache-2.0 | # Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writin... | # Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writin... | --- python/sglang/srt/layers/logits_processor.py
+++ python/sglang/srt/layers/logits_processor.py
@@ -23,15 +23,16 @@
from torch import nn
from sglang.srt.distributed import (
- get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_gather,
)
from sglang.srt.l... | 29 | 8 | 3 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/logits_processor.py).
Goal: Performing Vocabulary Parallelism for LM Head across Attention TP Groups
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2023-2024 SGLang Team
# Licensed under the Apache Lice... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
<!-- Explain the purpose of this PR and the goals i... | [] | null | null |
sglang | 18ce468d56aa33a8bebcef0bc3f4777b0d70cdce | 2025-04-26T07:24:59+08:00 | 5,740 | https://github.com/sgl-project/sglang/pull/5740 | update triton 3.2.0 h200 fused moe triton config and add warning about triton fused_moe_kernel performance degradation due to different Triton versions. (#5740) | update triton 3.2.0 h200 fused moe triton config and add warning about triton fused_moe_kernel performance degradation due to different Triton versions. (#5740) | python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py | triton | Apache-2.0 | # Adapted from https://github.com/vllm-project/vllm/blob/a6221a144af772fd1a68fe7e627935dc53e81738/vllm/model_executor/layers/fused_moe/fused_moe.py
"""Fused MoE kernel."""
import functools
import json
import logging
import os
from typing import Any, Callable, Dict, List, Optional, Tuple
import torch
import triton
im... | # Adapted from https://github.com/vllm-project/vllm/blob/a6221a144af772fd1a68fe7e627935dc53e81738/vllm/model_executor/layers/fused_moe/fused_moe.py
"""Fused MoE kernel."""
import functools
import json
import logging
import os
from typing import Any, Callable, Dict, List, Optional, Tuple
import torch
import triton
im... | --- python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py
+++ python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py
@@ -940,7 +940,10 @@
)
if os.path.exists(config_file_path):
with open(config_file_path) as f:
- logger.info("Using configuration from %s for MoE layer.", config_fil... | 4 | 1 | 1 | [] | [
"H200"
] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py).
Goal: update triton 3.2.0 h200 fused moe triton config and add warning about triton fused_moe_kernel performance degradation due to different Triton versions.
Target hardware: H200
Target models: unspecified
Cur... | ## Motivation
<img width="839" alt="图片" src="https://github.com/user-attachments/assets/06f8db13-b826-4bc1-8523-ff2894478cb4" />
Reproduction commands :
```shell
SGL_ENABLE_JIT_DEEPGEMM=0 python3 -m sglang.launch_server --model /DeepSeek-V3 --tp 8 --trust-remote-code --enable-dp-attention --dp-size 8
# Random 1k, 2... | [] | null | null |
sglang | 99456bcacb81a1e39c710f44d5921c1e60814cda | 2025-04-20T15:38:27+08:00 | 5,432 | https://github.com/sgl-project/sglang/pull/5432 | [perf] introduce deep gemm group_gemm_masked as bmm (#5432) | [perf] introduce deep gemm group_gemm_masked as bmm (#5432) | python/sglang/srt/layers/quantization/fp8_kernel.py | triton | Apache-2.0 | # Copyright 2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, so... | # Copyright 2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, so... | --- python/sglang/srt/layers/quantization/fp8_kernel.py
+++ python/sglang/srt/layers/quantization/fp8_kernel.py
@@ -44,6 +44,7 @@
fp8_min = -fp8_max
_enable_jit_deepgemm = False
+_enable_jit_deepgemm_bmm = False
if _is_cuda:
import deep_gemm
from sgl_kernel import (
@@ -53,10 +54,11 @@
)
sm_v... | 108 | 4 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/quantization/fp8_kernel.py).
Goal: [perf] introduce deep gemm group_gemm_masked as bmm
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (t... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
per-token-group quant+deep_gemm's grouped_gemm_mask... | [{"scope": "e2e", "hardware": "H200", "model": "DeepSeek-V3", "shape": "input-len=1000, output-len=100", "batch_size": 1, "metric": "Decode median latency", "unit": "s", "higher_is_better": false, "baseline": 0.0172, "optimized": 0.0166, "stated_speedup": null, "speedup": 1.0361}, {"scope": "e2e", "hardware": "H200", "... | null | 1.0361 |
sglang | 388e15c0db6b0381a1a7e08dd5b7a028130578fc | 2025-04-15T12:28:23+08:00 | 5,381 | https://github.com/sgl-project/sglang/pull/5381 | kernel: support slightly faster merge_state_v2 cuda kernel (#5381) | kernel: support slightly faster merge_state_v2 cuda kernel (#5381) | sgl-kernel/csrc/common_extension.cc | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/common_extension.cc
+++ sgl-kernel/csrc/common_extension.cc
@@ -47,6 +47,8 @@
m.impl("lightning_attention_decode", torch::kCUDA, &lightning_attention_decode);
m.def("merge_state(Tensor v_a, Tensor s_a, Tensor v_b, Tensor s_b, Tensor! v_merged, Tensor! s_merged) -> ()");
m.impl("merge_state"... | 2 | 0 | 4 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/common_extension.cc).
Goal: kernel: support slightly faster merge_state_v2 cuda kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version... | support slightly faster merge_state_v2 cuda kernel
- [x] float32
- [x] float16
- [x] bfloat16
- [x] dispatch by scalar_t
- [x] unit tests (performance & correctness)
- [x] support any headsize which is multiples of 4(float) or 8(half/bfloat16), merge_state v1 seems only support 64,128,256
- [x] slightly faster than m... | [{"scope": "kernel", "hardware": "L20", "model": null, "shape": "tokens=256,heads=8,headsize=32", "batch_size": 256, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 0.0391, "optimized": 0.0204, "stated_speedup": 1.9198, "speedup": 1.9167}, {"scope": "kernel", "hardware": "L20", "model": null, ... | 1.0427 | null |
sglang | 388e15c0db6b0381a1a7e08dd5b7a028130578fc | 2025-04-15T12:28:23+08:00 | 5,381 | https://github.com/sgl-project/sglang/pull/5381 | kernel: support slightly faster merge_state_v2 cuda kernel (#5381) | kernel: support slightly faster merge_state_v2 cuda kernel (#5381) | sgl-kernel/include/sgl_kernel_ops.h | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/include/sgl_kernel_ops.h
+++ sgl-kernel/include/sgl_kernel_ops.h
@@ -88,6 +88,8 @@
torch::Tensor output,
torch::Tensor new_kv);
void merge_state(
+ at::Tensor v_a, at::Tensor s_a, at::Tensor v_b, at::Tensor s_b, at::Tensor v_merged, at::Tensor s_merged);
+void merge_state_v2(
at::Tenso... | 2 | 0 | 4 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/include/sgl_kernel_ops.h).
Goal: kernel: support slightly faster merge_state_v2 cuda kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version... | support slightly faster merge_state_v2 cuda kernel
- [x] float32
- [x] float16
- [x] bfloat16
- [x] dispatch by scalar_t
- [x] unit tests (performance & correctness)
- [x] support any headsize which is multiples of 4(float) or 8(half/bfloat16), merge_state v1 seems only support 64,128,256
- [x] slightly faster than m... | [{"scope": "kernel", "hardware": "L20", "model": null, "shape": "tokens=256,heads=8,headsize=32", "batch_size": 256, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 0.0391, "optimized": 0.0204, "stated_speedup": 1.9198, "speedup": 1.9167}, {"scope": "kernel", "hardware": "L20", "model": null, ... | 1.0427 | null |
sglang | bdde237562ebd70944aef94873a96f765578b048 | 2025-04-15T03:35:43+08:00 | 5,370 | https://github.com/sgl-project/sglang/pull/5370 | [perf] experimental enhance fp8 per-tensor quant (#5370) | [perf] experimental enhance fp8 per-tensor quant (#5370) | python/sglang/srt/layers/quantization/fp8_kernel.py | triton | Apache-2.0 | # Copyright 2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, so... | # Copyright 2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, so... | --- python/sglang/srt/layers/quantization/fp8_kernel.py
+++ python/sglang/srt/layers/quantization/fp8_kernel.py
@@ -839,3 +839,103 @@
)
return C
+
+
+@triton.jit
+def _per_tensor_quant_mla_fp8_stage1(
+ x_ptr,
+ x_s_ptr,
+ head_size,
+ x_stride_h,
+ x_stride_s,
+ eps,
+ fp8_max,
+ ... | 100 | 0 | 2 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/quantization/fp8_kernel.py).
Goal: [perf] experimental enhance fp8 per-tensor quant
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the ... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
1. fix `input_to_float8` precision
2. impl mla absr... | [{"scope": "e2e", "hardware": null, "model": "RedHatAI/DeepSeek-Coder-V2-Lite-Instruct-FP8", "shape": "input-len=128,output-len=128", "batch_size": 1, "metric": "Decode. median latency", "unit": "s", "higher_is_better": false, "baseline": 0.00732, "optimized": 0.00628, "stated_speedup": null, "speedup": 1.1656}, {"scop... | null | 1.1658 |
sglang | 3e4794aad8a2349c1ac3d98d2c6369c637fade18 | 2025-04-13T01:01:13+08:00 | 5,294 | https://github.com/sgl-project/sglang/pull/5294 | refine fused_moe tuning docs (#5294) | refine fused_moe tuning docs (#5294) | python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py | triton | Apache-2.0 | # Adapted from https://github.com/vllm-project/vllm/blob/a6221a144af772fd1a68fe7e627935dc53e81738/vllm/model_executor/layers/fused_moe/fused_moe.py
"""Fused MoE kernel."""
import functools
import json
import logging
import os
from typing import Any, Callable, Dict, List, Optional, Tuple
import torch
import triton
im... | # Adapted from https://github.com/vllm-project/vllm/blob/a6221a144af772fd1a68fe7e627935dc53e81738/vllm/model_executor/layers/fused_moe/fused_moe.py
"""Fused MoE kernel."""
import functools
import json
import logging
import os
from typing import Any, Callable, Dict, List, Optional, Tuple
import torch
import triton
im... | --- python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py
+++ python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py
@@ -956,7 +956,7 @@
logger.warning(
(
"Using default MoE config. Performance might be sub-optimal! "
- "Config file not found at %s"
+ "Config fi... | 1 | 1 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py).
Goal: refine fused_moe tuning docs
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Adapted from https://github.com/vllm-project/vllm/blob/a6221a144af772fd1a68fe7e627935dc53e8173... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
<!-- Explain the purpose of this PR and the goals i... | [] | null | null |
sglang | 23c764b18aeb37c42ddedd7468f1a5753df1f232 | 2025-04-02T00:23:25+08:00 | 4,767 | https://github.com/sgl-project/sglang/pull/4767 | [Feature] Support DeepEP Low Latency (#4767) | [Feature] Support DeepEP Low Latency (#4767)
Co-authored-by: sleepcoo <sleepcoo@gmail.com>
Co-authored-by: laixinn <xielx@shanghaitech.edu.cn>
Co-authored-by: ch-wan <cwan39@gatech.edu> | python/sglang/srt/layers/moe/ep_moe/kernels.py | triton | Apache-2.0 | import logging
from typing import List, Optional
import torch
import triton
import triton.language as tl
from sglang.srt.distributed import get_tensor_model_parallel_rank
from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
from sglang.srt.utils import is_cuda
_is_cuda = is_cuda()
if _is_c... | import logging
from typing import List, Optional
import torch
import triton
import triton.language as tl
from sglang.srt.distributed import get_tensor_model_parallel_rank
from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
from sglang.srt.utils import is_cuda
_is_cuda = is_cuda()
if _is_c... | --- python/sglang/srt/layers/moe/ep_moe/kernels.py
+++ python/sglang/srt/layers/moe/ep_moe/kernels.py
@@ -242,6 +242,148 @@
silu_mul_output = gate_output * up_output * scale
silu_mul_output = silu_mul_output.to(OutDtype)
tl.store(down_input_ptr + offset, silu_mul_output, mask=mask... | 142 | 0 | 3 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/ep_moe/kernels.py).
Goal: [Feature] Support DeepEP Low Latency
Target hardware: unspecified
Target models: unspecified
Current file:
```python
import logging
from typing import List, Optional
import torch
import triton
import triton.language ... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
<!-- Explain the purpose of this PR and the goals i... | [] | null | null |
sglang | 65c24c28f97bbcabf69008d14c47a72149398385 | 2025-03-23T23:44:17-07:00 | 4,396 | https://github.com/sgl-project/sglang/pull/4396 | [Quant Kernel] refactored per token group quant fp8 to support int8 up-to 2x faster (#4396) | [Quant Kernel] refactored per token group quant fp8 to support int8 up-to 2x faster (#4396) | sgl-kernel/csrc/torch_extension.cc | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/torch_extension.cc
+++ sgl-kernel/csrc/torch_extension.cc
@@ -98,6 +98,11 @@
" float eps, float fp8_min, float fp8_max) -> ()");
m.impl("sgl_per_token_group_quant_fp8", torch::kCUDA, &sgl_per_token_group_quant_fp8);
+ m.def(
+ "sgl_per_token_group_quant_int8(Tensor input, Tensor ou... | 5 | 0 | 5 | [] | [] | [
"2x"
] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/torch_extension.cc).
Goal: [Quant Kernel] refactored per token group quant fp8 to support int8 up-to 2x faster
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
#2965
The new cuda kernel accelerates 1x-2x compar... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "seq_len=64,group_size=128,dst_dtype=torch.int8", "batch_size": 1, "metric": "latency", "unit": null, "higher_is_better": false, "baseline": 10.848, "optimized": 9.152, "stated_speedup": null, "speedup": 1.1853}, {"scope": "kernel", "hardware": null, "model... | 1.87 | null |
sglang | 65c24c28f97bbcabf69008d14c47a72149398385 | 2025-03-23T23:44:17-07:00 | 4,396 | https://github.com/sgl-project/sglang/pull/4396 | [Quant Kernel] refactored per token group quant fp8 to support int8 up-to 2x faster (#4396) | [Quant Kernel] refactored per token group quant fp8 to support int8 up-to 2x faster (#4396) | sgl-kernel/include/sgl_kernel_ops.h | cpp | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/include/sgl_kernel_ops.h
+++ sgl-kernel/include/sgl_kernel_ops.h
@@ -141,6 +141,14 @@
double eps,
double fp8_min,
double fp8_max);
+void sgl_per_token_group_quant_int8(
+ at::Tensor input,
+ at::Tensor output_q,
+ at::Tensor output_s,
+ int64_t group_size,
+ double eps,
+ ... | 8 | 0 | 5 | [] | [] | [
"2x"
] | Optimize this cpp kernel file from sglang (sgl-kernel/include/sgl_kernel_ops.h).
Goal: [Quant Kernel] refactored per token group quant fp8 to support int8 up-to 2x faster
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under th... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
#2965
The new cuda kernel accelerates 1x-2x compar... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "seq_len=64,group_size=128,dst_dtype=torch.int8", "batch_size": 1, "metric": "latency", "unit": null, "higher_is_better": false, "baseline": 10.848, "optimized": 9.152, "stated_speedup": null, "speedup": 1.1853}, {"scope": "kernel", "hardware": null, "model... | 1.87 | null |
sglang | c2bd094d6eb621a51a19d6595ddb3ddcc9030970 | 2025-03-23T05:30:34+08:00 | 4,643 | https://github.com/sgl-project/sglang/pull/4643 | Optimize Permute Kernel in DeepEP (#4643) | Optimize Permute Kernel in DeepEP (#4643)
Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com> | python/sglang/srt/layers/moe/ep_moe/kernels.py | triton | Apache-2.0 | import logging
from typing import List, Optional
import torch
import triton
import triton.language as tl
from sglang.srt.distributed import get_tensor_model_parallel_rank
from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
from sglang.srt.utils import is_cuda
_is_cuda = is_cuda()
if _is_c... | import logging
from typing import List, Optional
import torch
import triton
import triton.language as tl
from sglang.srt.distributed import get_tensor_model_parallel_rank
from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
from sglang.srt.utils import is_cuda
_is_cuda = is_cuda()
if _is_c... | --- python/sglang/srt/layers/moe/ep_moe/kernels.py
+++ python/sglang/srt/layers/moe/ep_moe/kernels.py
@@ -15,52 +15,6 @@
sglang_per_token_group_quant_fp8,
)
logger = logging.getLogger(__name__)
-
-
-@triton.jit
-def compute_src2dst_triton_kernel(
- reorder_ids, src2dst, num_toks, BLOCK_SIZE: tl.conste... | 47 | 49 | 3 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/ep_moe/kernels.py).
Goal: Optimize Permute Kernel in DeepEP
Target hardware: unspecified
Target models: unspecified
Current file:
```python
import logging
from typing import List, Optional
import torch
import triton
import triton.language as ... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
<!-- Explain the purpose of this PR and the goals i... | [] | null | null |
sglang | 6a384d5c012e424e5baf9891efa5465088e807dc | 2025-03-22T00:37:57-07:00 | 4,639 | https://github.com/sgl-project/sglang/pull/4639 | Speed up per token and per tensor quant by 15% (#4639) | Speed up per token and per tensor quant by 15% (#4639) | sgl-kernel/csrc/gemm/per_tensor_quant_fp8.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <c10/util/Float8_e4m3fn.h>
#include <cmath>
#include <cub/block/block_reduce.cuh>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
template <typename T>
__global__ void
per_tensor_absmax_kernel(const T* __restrict__ input, float* __restrict__ output_s, const int64_t... | #include <ATen/cuda/CUDAContext.h>
#include <c10/util/Float8_e4m3fn.h>
#include <cmath>
#include <cub/block/block_reduce.cuh>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
template <typename T>
__global__ void
per_tensor_absmax_kernel(const T* __restrict__ input, float* __restrict__ output_s, const int64_t... | --- sgl-kernel/csrc/gemm/per_tensor_quant_fp8.cu
+++ sgl-kernel/csrc/gemm/per_tensor_quant_fp8.cu
@@ -54,42 +54,40 @@
const int grid_size = blockDim.x * gridDim.x;
const float scale_val = 1.0f / (*scale);
- constexpr uint32_t vec_size = 16 / sizeof(T);
- using vec_t = flashinfer::vec_t<T, vec_size>;
+ // We ... | 12 | 14 | 2 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/gemm/per_tensor_quant_fp8.cu).
Goal: Speed up per token and per tensor quant by 15%
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#include <c10/util/Float8_e4m3fn.h>
#include <cmath>
#includ... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
#2965
Improvements
1. Store using `uint4` cast to ... | [] | null | null |
sglang | 6a384d5c012e424e5baf9891efa5465088e807dc | 2025-03-22T00:37:57-07:00 | 4,639 | https://github.com/sgl-project/sglang/pull/4639 | Speed up per token and per tensor quant by 15% (#4639) | Speed up per token and per tensor quant by 15% (#4639) | sgl-kernel/csrc/gemm/per_token_quant_fp8.cu | cuda | Apache-2.0 | #include <ATen/cuda/CUDAContext.h>
#include <cmath>
#include <cub/block/block_reduce.cuh>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
template <typename T>
__global__ void per_token_quant_fp8_kernel(
const T* __restrict__ input,
FP8_TYPE* __restrict__ output_q,
float* __restrict__ output_s,
... | #include <ATen/cuda/CUDAContext.h>
#include <cmath>
#include <cub/block/block_reduce.cuh>
#include <flashinfer/vec_dtypes.cuh>
#include "utils.h"
template <typename T>
__global__ void per_token_quant_fp8_kernel(
const T* __restrict__ input,
FP8_TYPE* __restrict__ output_q,
float* __restrict__ output_s,
... | --- sgl-kernel/csrc/gemm/per_token_quant_fp8.cu
+++ sgl-kernel/csrc/gemm/per_token_quant_fp8.cu
@@ -24,17 +24,19 @@
float max_value = 0.0f;
- constexpr uint32_t vec_size = 16 / sizeof(T);
- using vec_t = flashinfer::vec_t<T, vec_size>;
- const int32_t num_vec_elems = hidden_dim / vec_size;
+ // We want to st... | 17 | 18 | 2 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/gemm/per_token_quant_fp8.cu).
Goal: Speed up per token and per tensor quant by 15%
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#include <cmath>
#include <cub/block/block_reduce.cuh>
#inclu... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
#2965
Improvements
1. Store using `uint4` cast to ... | [] | null | null |
sglang | c0e9a36c5f5e8e018663a3b84c401780c5999381 | 2025-03-19T12:25:38+08:00 | 4,553 | https://github.com/sgl-project/sglang/pull/4553 | Optimize Triton decoding kernel for dynamic workload (#4553) | Optimize Triton decoding kernel for dynamic workload (#4553) | python/sglang/srt/layers/attention/triton_backend.py | triton | Apache-2.0 | from __future__ import annotations
from typing import TYPE_CHECKING, Optional, Union
import torch
import triton
from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
from sglang.srt.layers.attention.utils import create_flashinfer_kv_indices_triton
from sglang.srt.layers.dp_attention import get_a... | from __future__ import annotations
from typing import TYPE_CHECKING, Optional, Union
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
from sglang.srt.layers.attention.utils import create_flashinfer_kv_indices_triton
from sglang.srt.lay... | --- python/sglang/srt/layers/attention/triton_backend.py
+++ python/sglang/srt/layers/attention/triton_backend.py
@@ -4,16 +4,63 @@
import torch
import triton
+import triton.language as tl
from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
from sglang.srt.layers.attention.utils import cr... | 142 | 15 | 5 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/triton_backend.py).
Goal: Optimize Triton decoding kernel for dynamic workload
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from __future__ import annotations
from typing import TYPE_CHECKING, Optional... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
The current implementation of triton decode attenti... | [{"scope": "e2e", "hardware": null, "model": null, "shape": null, "batch_size": null, "metric": "total token throughput", "unit": "tok/s", "higher_is_better": true, "baseline": 10566.4, "optimized": 10893.63, "stated_speedup": null, "speedup": 1.031}, {"scope": "e2e", "hardware": null, "model": null, "shape": null, "ba... | 1.0506 | 1.0309 |
sglang | c0e9a36c5f5e8e018663a3b84c401780c5999381 | 2025-03-19T12:25:38+08:00 | 4,553 | https://github.com/sgl-project/sglang/pull/4553 | Optimize Triton decoding kernel for dynamic workload (#4553) | Optimize Triton decoding kernel for dynamic workload (#4553) | python/sglang/srt/layers/attention/triton_ops/decode_attention.py | triton | Apache-2.0 | # Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writin... | # Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writin... | --- python/sglang/srt/layers/attention/triton_ops/decode_attention.py
+++ python/sglang/srt/layers/attention/triton_ops/decode_attention.py
@@ -37,6 +37,9 @@
)
+_MIN_BLOCK_KV = 32
+
+
@triton.jit
def tanh(x):
# Tanh is just a scaled sigmoid
@@ -52,6 +55,8 @@
kv_indptr,
kv_indices,
Att_Out,
+ ... | 95 | 37 | 5 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/triton_ops/decode_attention.py).
Goal: Optimize Triton decoding kernel for dynamic workload
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2023-2024 SGLang Team
# Licensed under the Apache Lic... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
The current implementation of triton decode attenti... | [{"scope": "e2e", "hardware": null, "model": null, "shape": null, "batch_size": null, "metric": "total token throughput", "unit": "tok/s", "higher_is_better": true, "baseline": 10566.4, "optimized": 10893.63, "stated_speedup": null, "speedup": 1.031}, {"scope": "e2e", "hardware": null, "model": null, "shape": null, "ba... | 1.0506 | 1.0309 |
sglang | 25e1816eff104da56f97ce494e255306603fe2f6 | 2025-03-17T03:16:30+08:00 | 4,477 | https://github.com/sgl-project/sglang/pull/4477 | fix custom allreduce performance/accuracy problem (#4477) | fix custom allreduce performance/accuracy problem (#4477) | sgl-kernel/csrc/allreduce/trt_reduce_internal.cu | cuda | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/csrc/allreduce/trt_reduce_internal.cu
+++ sgl-kernel/csrc/allreduce/trt_reduce_internal.cu
@@ -182,8 +182,9 @@
}
}
}
-
- __syncthreads();
+ if constexpr (start || need_fence) {
+ __syncthreads();
+ }
}
template <typename T, int RANKS_PER_NODE, bool COPY_INPUT = true>
@@ -262,6 +2... | 6 | 19 | 2 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/allreduce/trt_reduce_internal.cu).
Goal: fix custom allreduce performance/accuracy problem
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, V... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Fix accuracy/performance problem for custom allreduc... | [] | null | null |
sglang | 25e1816eff104da56f97ce494e255306603fe2f6 | 2025-03-17T03:16:30+08:00 | 4,477 | https://github.com/sgl-project/sglang/pull/4477 | fix custom allreduce performance/accuracy problem (#4477) | fix custom allreduce performance/accuracy problem (#4477) | sgl-kernel/include/trt_reduce_internal.cuh | cuda | Apache-2.0 | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | /* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in... | --- sgl-kernel/include/trt_reduce_internal.cuh
+++ sgl-kernel/include/trt_reduce_internal.cuh
@@ -39,7 +39,7 @@
namespace trt_llm {
constexpr size_t WARP_SIZE = 32;
-constexpr size_t MAX_ALL_REDUCE_BLOCKS = 36;
+constexpr size_t MAX_ALL_REDUCE_BLOCKS = 32;
constexpr size_t MAX_RANKS_PER_NODE = 8;
constexpr size_t... | 1 | 1 | 2 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/include/trt_reduce_internal.cuh).
Goal: fix custom allreduce performance/accuracy problem
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version... | <!-- Thank you for your contribution! We appreciate it. The following guidelines will help improve your pull request and facilitate feedback. If anything is unclear, don't hesitate to submit your pull request and ask the maintainers for assistance. -->
## Motivation
Fix accuracy/performance problem for custom allreduc... | [] | null | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.