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 | daae3acb3642f99f8ef0393c74188eea6d31b60d | 2026-08-12T16:27:44+08:00 | 34,507 | https://github.com/sgl-project/sglang/pull/34507 | [Diffusion][Z-Image] Tune native QK RMSNorm launch for SM103 (#34507) | [Diffusion][Z-Image] Tune native QK RMSNorm launch for SM103 (#34507) | python/sglang/kernels/ops/diffusion/triton/zimage_native_norm.py | triton | Apache-2.0 | # SPDX-License-Identifier: Apache-2.0
"""Z-Image-specific bit-exact per-head RMSNorm kernel."""
from __future__ import annotations
import torch
import triton # type: ignore
import triton.language as tl # type: ignore
@triton.jit
def _qk_rmsnorm_native_kernel(
y_ptr,
x_ptr,
weight_ptr,
token_stride... | # SPDX-License-Identifier: Apache-2.0
"""Z-Image-specific bit-exact per-head RMSNorm kernel."""
from __future__ import annotations
import torch
import triton # type: ignore
import triton.language as tl # type: ignore
from sglang.kernels.jit.utils import get_jit_cuda_arch
@triton.jit
def _qk_rmsnorm_native_kernel... | --- python/sglang/kernels/ops/diffusion/triton/zimage_native_norm.py
+++ python/sglang/kernels/ops/diffusion/triton/zimage_native_norm.py
@@ -6,6 +6,8 @@
import torch
import triton # type: ignore
import triton.language as tl # type: ignore
+
+from sglang.kernels.jit.utils import get_jit_cuda_arch
@triton.jit
... | 10 | 2 | 1 | [] | [
"SM103"
] | [] | Optimize this triton kernel file from sglang (python/sglang/kernels/ops/diffusion/triton/zimage_native_norm.py).
Goal: [Diffusion][Z-Image] Tune native QK RMSNorm launch for SM103
Target hardware: SM103
Target models: unspecified
Current file:
```python
# SPDX-License-Identifier: Apache-2.0
"""Z-Image-specific bit-ex... | ## Motivation
The bit-exact native Z-Image Q/K RMSNorm kernel used a single launch
configuration on every GPU. At the production Z-Image shape, B300 is
launch-bound with `rows_per_prog=8, num_warps=8`.
## Modifications
- Use `rows_per_prog=16, num_warps=4` on SM103.
- Preserve `rows_per_prog=8, num_warps=8` on every... | [{"scope": "kernel", "hardware": "B300", "model": null, "shape": "N=1,K=4096", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 24.58, "optimized": 14.33, "stated_speedup": 1.72, "speedup": 1.7153}, {"scope": "kernel", "hardware": "B300", "model": null, "shape": "N=1,K=4096", "... | 2.0446 | null |
sglang | 45f706333536466479d8977be14768abda11e438 | 2026-08-12T16:24:18+08:00 | 34,503 | https://github.com/sgl-project/sglang/pull/34503 | [Diffusion] Tune QK head LayerNorm for SM103 (#34503) | [Diffusion] Tune QK head LayerNorm for SM103 (#34503) | python/sglang/kernels/ops/diffusion/triton/layernorm_modulate.py | triton | Apache-2.0 | # SPDX-License-Identifier: Apache-2.0
"""Fused LayerNorm + adaLN modulate Triton kernels for bf16 activations.
Two fusions, each replacing an eager multi-kernel chain with a single
launch while reproducing the eager results bit for bit (``torch.equal``),
so callers need no quality gate:
- ``fused_layernorm_modulate``... | # SPDX-License-Identifier: Apache-2.0
"""Fused LayerNorm + adaLN modulate Triton kernels for bf16 activations.
Two fusions, each replacing an eager multi-kernel chain with a single
launch while reproducing the eager results bit for bit (``torch.equal``),
so callers need no quality gate:
- ``fused_layernorm_modulate``... | --- python/sglang/kernels/ops/diffusion/triton/layernorm_modulate.py
+++ python/sglang/kernels/ops/diffusion/triton/layernorm_modulate.py
@@ -341,9 +341,13 @@
return t.is_cuda and t.dtype is torch.bfloat16
-def _is_sm120_or_newer() -> bool:
+def _qk_head_launch_config() -> tuple[int, int]:
arch = get_jit_... | 11 | 9 | 1 | [] | [
"SM103"
] | [] | Optimize this triton kernel file from sglang (python/sglang/kernels/ops/diffusion/triton/layernorm_modulate.py).
Goal: [Diffusion] Tune QK head LayerNorm for SM103
Target hardware: SM103
Target models: unspecified
Current file:
```python
# SPDX-License-Identifier: Apache-2.0
"""Fused LayerNorm + adaLN modulate Triton... | ## Motivation
PR #34349 tuned the exact Q/K head LayerNorm launch for SM120 while preserving
the H100/H200 launch. B300 (SM103) also suffers from register pressure with the
SM90 `ROWS=64, num_warps=2` configuration, but its best launch is different
from SM120.
## Modifications
- Use `ROWS=16, num_warps=1` on SM103.
... | [{"scope": "kernel", "hardware": "B300", "model": null, "shape": "N=4360,K=32", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 44.97, "optimized": 30.78, "stated_speedup": 1.46, "speedup": 1.461}] | 1.461 | null |
sglang | 3f9d184833a4626e06d72bf099cc00a8c4258886 | 2026-08-12T10:41:54+08:00 | 34,349 | https://github.com/sgl-project/sglang/pull/34349 | [Diffusion] Tune QK head LayerNorm for SM120 (#34349) | [Diffusion] Tune QK head LayerNorm for SM120 (#34349) | python/sglang/kernels/ops/diffusion/triton/layernorm_modulate.py | triton | Apache-2.0 | # SPDX-License-Identifier: Apache-2.0
"""Fused LayerNorm + adaLN modulate Triton kernels for bf16 activations.
Two fusions, each replacing an eager multi-kernel chain with a single
launch while reproducing the eager results bit for bit (``torch.equal``),
so callers need no quality gate:
- ``fused_layernorm_modulate``... | # SPDX-License-Identifier: Apache-2.0
"""Fused LayerNorm + adaLN modulate Triton kernels for bf16 activations.
Two fusions, each replacing an eager multi-kernel chain with a single
launch while reproducing the eager results bit for bit (``torch.equal``),
so callers need no quality gate:
- ``fused_layernorm_modulate``... | --- python/sglang/kernels/ops/diffusion/triton/layernorm_modulate.py
+++ python/sglang/kernels/ops/diffusion/triton/layernorm_modulate.py
@@ -46,6 +46,7 @@
import triton # type: ignore
import triton.language as tl # type: ignore
+from sglang.kernels.jit.utils import get_jit_cuda_arch
from sglang.kernels.ops.diff... | 13 | 2 | 1 | [] | [
"SM120"
] | [] | Optimize this triton kernel file from sglang (python/sglang/kernels/ops/diffusion/triton/layernorm_modulate.py).
Goal: [Diffusion] Tune QK head LayerNorm for SM120
Target hardware: SM120
Target models: unspecified
Current file:
```python
# SPDX-License-Identifier: Apache-2.0
"""Fused LayerNorm + adaLN modulate Triton... | ## Motivation
The exact Q/K head LayerNorm launch was tuned for H200 with 64 rows per Triton program. On RTX 5090, this configuration uses 142 registers per thread and limits the kernel to 25% theoretical / 23.11% achieved occupancy, making it materially slower than a smaller SM120 row group.
## Modifications
- Sele... | [{"scope": "kernel", "hardware": "RTX 5090", "model": null, "shape": "1, 4360, 32, 128", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 93.66, "optimized": 83.42, "stated_speedup": null, "speedup": 1.1228}, {"scope": "kernel", "hardware": "H200", "model": null, "shape": "1, 4... | 1.0006 | null |
sglang | ca0f8a0f4ced5fc107f27a3be386d12bc7850711 | 2026-08-11T05:30:15+08:00 | 33,484 | https://github.com/sgl-project/sglang/pull/33484 | perf(hisparse): fuse the DSv4 value and scale swap-in copy on ROCm (#33484) | perf(hisparse): fuse the DSv4 value and scale swap-in copy on ROCm (#33484) | python/sglang/kernels/jit/csrc/hisparse.cuh | cuda | Apache-2.0 | #include <sgl_kernel/tensor.h>
#include <sgl_kernel/utils.h>
#include <sgl_kernel/utils.cuh>
#include <sgl_kernel/deepseek_v4/kvcacheio.cuh>
#include <dlpack/dlpack.h>
#include <tvm/ffi/container/tensor.h>
#include <stdexcept>
#include <stdint.h>
#include <string>
namespace sglang {
#ifdef USE_ROCM
constexpr int ... | #include <sgl_kernel/tensor.h>
#include <sgl_kernel/utils.h>
#include <sgl_kernel/utils.cuh>
#include <sgl_kernel/deepseek_v4/kvcacheio.cuh>
#include <dlpack/dlpack.h>
#include <tvm/ffi/container/tensor.h>
#include <stdexcept>
#include <stdint.h>
#include <string>
namespace sglang {
#ifdef USE_ROCM
constexpr int ... | --- python/sglang/kernels/jit/csrc/hisparse.cuh
+++ python/sglang/kernels/jit/csrc/hisparse.cuh
@@ -84,6 +84,56 @@
const int64_t tail_start = word_count * static_cast<int64_t>(sizeof(uint64_t));
for (int64_t i = tail_start + lane_id; i < item_size_bytes; i += WARP_SIZE) {
dst[i] = src[i];
+ }
+}
+
+// Copie... | 53 | 4 | 1 | [
"deepseek"
] | [
"ROCM"
] | [] | Optimize this cuda kernel file from sglang (python/sglang/kernels/jit/csrc/hisparse.cuh).
Goal: perf(hisparse): fuse the DSv4 value and scale swap-in copy on ROCm
Target hardware: ROCM
Target models: deepseek
Current file:
```cuda
#include <sgl_kernel/tensor.h>
#include <sgl_kernel/utils.h>
#include <sgl_kernel/util... | ## Motivation
A DeepSeek-V4 C4 token is 584 bytes, but it is **not one contiguous range**. A page row holds 64 values of 576B followed by 64 scales of 8B:
```
/// Paged C4 cache layout:
/// VALUE 0, VALUE 1, ..., VALUE 63,
/// SCALE 0, SCALE 1, ..., SCALE 63,
/// [Padding to align to 576 bytes]
```
So the ROCm swap-... | [{"scope": "kernel", "hardware": "MI355X", "model": "DSv4", "shape": "miss_rate=0.2,hot_buffer_size=4096,top_k=2048", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 102.4, "optimized": 87.8, "stated_speedup": null, "speedup": 1.1663}, {"scope": "kernel", "hardware": "MI355X",... | 1.1299 | null |
sglang | 1a8e4876b6ddccc5f2712efdca7f87fdc5514c7b | 2026-08-11T04:27:15+08:00 | 33,085 | https://github.com/sgl-project/sglang/pull/33085 | perf(hisparse): 128-bit non-temporal swap-in copy on ROCm (#33085) | perf(hisparse): 128-bit non-temporal swap-in copy on ROCm (#33085) | python/sglang/kernels/jit/csrc/hisparse.cuh | cuda | Apache-2.0 | #include <sgl_kernel/tensor.h>
#include <sgl_kernel/utils.h>
#include <sgl_kernel/utils.cuh>
#include <sgl_kernel/deepseek_v4/kvcacheio.cuh>
#include <dlpack/dlpack.h>
#include <tvm/ffi/container/tensor.h>
#include <stdexcept>
#include <stdint.h>
#include <string>
namespace sglang {
#ifdef USE_ROCM
constexpr int ... | #include <sgl_kernel/tensor.h>
#include <sgl_kernel/utils.h>
#include <sgl_kernel/utils.cuh>
#include <sgl_kernel/deepseek_v4/kvcacheio.cuh>
#include <dlpack/dlpack.h>
#include <tvm/ffi/container/tensor.h>
#include <stdexcept>
#include <stdint.h>
#include <string>
namespace sglang {
#ifdef USE_ROCM
constexpr int ... | --- python/sglang/kernels/jit/csrc/hisparse.cuh
+++ python/sglang/kernels/jit/csrc/hisparse.cuh
@@ -32,15 +32,52 @@
}
#ifdef USE_ROCM
+// 128-bit vector type used by the wide copy path below.
+using TransferVec4 = __attribute__((__vector_size__(4 * sizeof(uint32_t)))) uint32_t;
+
__device__ __forceinline__ void tr... | 38 | 1 | 1 | [] | [
"ROCM"
] | [] | Optimize this cuda kernel file from sglang (python/sglang/kernels/jit/csrc/hisparse.cuh).
Goal: perf(hisparse): 128-bit non-temporal swap-in copy on ROCm
Target hardware: ROCM
Target models: unspecified
Current file:
```cuda
#include <sgl_kernel/tensor.h>
#include <sgl_kernel/utils.h>
#include <sgl_kernel/utils.cuh>... | ## Motivation
`transfer_item_warp` exists **twice** in the tree:
| file | used by | ROCm today |
|---|---|---|
| `python/sglang/kernels/aot/csrc/kvcacheio/transfer.cu` | HiCache L2→L1 gather (prefill/TTFT) | being widened in #30024 |
| `python/sglang/kernels/jit/csrc/hisparse.cuh` | **HiSparse swap-in miss copy (deco... | [{"scope": "kernel", "hardware": "MI355X", "model": "DeepSeek-V4", "shape": "ITEM_SIZE_BYTES=1024", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 105.9, "optimized": 84.8, "stated_speedup": null, "speedup": 1.2488}, {"scope": "kernel", "hardware": "MI355X", "model": "DeepSee... | 1.0034 | null |
sglang | 18e6c61c21ad39725522c008190d2b540dd6228d | 2026-08-06T14:46:10-07:00 | 29,677 | https://github.com/sgl-project/sglang/pull/29677 | [AMD] perf: compact Triton extend-attention for ragged prefill (AMD/HIP-only) (#29677) | [AMD] perf: compact Triton extend-attention for ragged prefill (AMD/HIP-only) (#29677) | python/sglang/kernels/ops/attention/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/kernels/ops/attention/extend_attention.py
+++ python/sglang/kernels/ops/attention/extend_attention.py
@@ -25,6 +25,7 @@
context_attention_fwd,
)
from sglang.kernels.ops.attention.score_mod import unpack_aux_tensors
+from sglang.srt.environ import envs
from sglang.srt.utils import is_cuda, is_g... | 100 | 4 | 3 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/kernels/ops/attention/extend_attention.py).
Goal: [AMD] perf: compact Triton extend-attention for ragged prefill (AMD/HIP-only)
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2023-2024 SGLang Team
# Licensed under... | ## Motivation
This PR improves Triton extend-attention performance for ragged mixed-prefill batches by adding an optional **compact query-tile launch grid**.
The legacy grid is rectangular and sized by `max_len_extend`:
```python
grid = (batch_size, head_num, ceil(max_len_extend / BLOCK_M))
```
In a mixed batch, ev... | [{"scope": "e2e", "hardware": "MI355X", "model": "Qwen/Qwen2.5-7B-Instruct", "shape": null, "batch_size": 64, "metric": "throughput", "unit": "tok/s", "higher_is_better": true, "baseline": 14581, "optimized": 49877, "stated_speedup": null, "speedup": 3.4207}, {"scope": "e2e", "hardware": "MI355X", "model": "Qwen/Qwen2.... | 6.4 | 2.0246 |
sglang | 18e6c61c21ad39725522c008190d2b540dd6228d | 2026-08-06T14:46:10-07:00 | 29,677 | https://github.com/sgl-project/sglang/pull/29677 | [AMD] perf: compact Triton extend-attention for ragged prefill (AMD/HIP-only) (#29677) | [AMD] perf: compact Triton extend-attention for ragged prefill (AMD/HIP-only) (#29677) | python/sglang/kernels/ops/attention/verify_splitkv.py | triton | Apache-2.0 | """Split-KV (flash-decode) attention for EAGLE speculative *verify*.
Only valid when speculative ``topk == 1`` (the EAGLE tree reduces to a pure
causal chain); the caller gates on that. ``topk > 1`` trees fall back to
``extend_attention_fwd``.
On the Triton backend, EAGLE target-verify runs through the prefill
``exte... | """Split-KV (flash-decode) attention for EAGLE speculative *verify*.
Only valid when speculative ``topk == 1`` (the EAGLE tree reduces to a pure
causal chain); the caller gates on that. ``topk > 1`` trees fall back to
``extend_attention_fwd``.
On the Triton backend, EAGLE target-verify runs through the prefill
``exte... | --- python/sglang/kernels/ops/attention/verify_splitkv.py
+++ python/sglang/kernels/ops/attention/verify_splitkv.py
@@ -707,6 +707,12 @@
return False
if v_extend.shape[2] != v_buffer.shape[2]:
return False
+ # MLA (head_dim != v_head_dim, e.g. DeepSeek 576 vs 512) uses a shared
+ # latent K... | 6 | 0 | 3 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/kernels/ops/attention/verify_splitkv.py).
Goal: [AMD] perf: compact Triton extend-attention for ragged prefill (AMD/HIP-only)
Target hardware: unspecified
Target models: unspecified
Current file:
```python
"""Split-KV (flash-decode) attention for EAGLE specu... | ## Motivation
This PR improves Triton extend-attention performance for ragged mixed-prefill batches by adding an optional **compact query-tile launch grid**.
The legacy grid is rectangular and sized by `max_len_extend`:
```python
grid = (batch_size, head_num, ceil(max_len_extend / BLOCK_M))
```
In a mixed batch, ev... | [{"scope": "e2e", "hardware": "MI355X", "model": "Qwen/Qwen2.5-7B-Instruct", "shape": null, "batch_size": 64, "metric": "throughput", "unit": "tok/s", "higher_is_better": true, "baseline": 14581, "optimized": 49877, "stated_speedup": null, "speedup": 3.4207}, {"scope": "e2e", "hardware": "MI355X", "model": "Qwen/Qwen2.... | 6.4 | 2.0246 |
sglang | afeaeccfa2f1f2fe9e245fabef2b2e58c34d13f3 | 2026-07-31T12:09:44+08:00 | 32,483 | https://github.com/sgl-project/sglang/pull/32483 | perf(hisparse): eliminate redundant swap output fill (#32483) | perf(hisparse): eliminate redundant swap output fill (#32483)
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com> | python/sglang/kernels/jit/csrc/hisparse.cuh | cuda | Apache-2.0 | #include <sgl_kernel/tensor.h>
#include <sgl_kernel/utils.h>
#include <sgl_kernel/utils.cuh>
#include <sgl_kernel/deepseek_v4/kvcacheio.cuh>
#include <dlpack/dlpack.h>
#include <tvm/ffi/container/tensor.h>
#include <stdexcept>
#include <stdint.h>
#include <string>
namespace {
#ifdef USE_ROCM
constexpr int WARP_SI... | #include <sgl_kernel/tensor.h>
#include <sgl_kernel/utils.h>
#include <sgl_kernel/utils.cuh>
#include <sgl_kernel/deepseek_v4/kvcacheio.cuh>
#include <dlpack/dlpack.h>
#include <tvm/ffi/container/tensor.h>
#include <stdexcept>
#include <stdint.h>
#include <string>
namespace {
#ifdef USE_ROCM
constexpr int WARP_SI... | --- python/sglang/kernels/jit/csrc/hisparse.cuh
+++ python/sglang/kernels/jit/csrc/hisparse.cuh
@@ -217,10 +217,18 @@
constexpr int NUM_BUFFER_CHUNKS = (HOT_BUFFER_SIZE + WARP_SIZE - 1) / WARP_SIZE;
const int bid = blockIdx.x;
- // Early exit for padded blocks (CUDA graph pads batch to a captured size)
- if (... | 20 | 9 | 1 | [] | [] | [] | Optimize this cuda kernel file from sglang (python/sglang/kernels/jit/csrc/hisparse.cuh).
Goal: perf(hisparse): eliminate redundant swap output fill
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <sgl_kernel/tensor.h>
#include <sgl_kernel/utils.h>
#include <sgl_kernel/utils.cu... | ## Motivation
`HiSparseCoordinator.swap_in_selected_pages()` currently clears `top_k_device_locs` with `fill_(-1)` before launching `load_cache_to_device_buffer_mla`.
The fill is needed only because the short-sequence path and CUDA Graph padded rows did not overwrite every output entry. It adds one GPU kernel launch ... | [{"scope": "kernel", "hardware": "NVIDIA H20", "model": null, "shape": "top_k=2048, hot buffer=4096/8192", "batch_size": 1, "metric": "GPU-time improvement", "unit": "%", "higher_is_better": true, "baseline": null, "optimized": null, "stated_speedup": 1.4125, "speedup": 1.4125}, {"scope": "kernel", "hardware": "NVIDIA ... | 1.0755 | null |
sglang | 5558dbad003768d01e1c9b8f3e13245bf145d8a6 | 2026-07-28T19:45:46+08:00 | 31,931 | https://github.com/sgl-project/sglang/pull/31931 | [NPU] Optimize DeepSeek-V4 performance (#31931) | [NPU] Optimize DeepSeek-V4 performance (#31931) | python/sglang/kernels/ops/attention/deepseek_v4_rope.py | triton | Apache-2.0 | import logging
import math
from functools import lru_cache
from typing import Optional
import torch
import triton
import triton.language as tl
logger = logging.getLogger(__name__)
# This module is imported during model-registry discovery. Keep it free of
# TileLang imports so discovery does not load TileLang's nativ... | import logging
import math
from functools import lru_cache
from typing import Optional
import torch
import triton
import triton.language as tl
logger = logging.getLogger(__name__)
# This module is imported during model-registry discovery. Keep it free of
# TileLang imports so discovery does not load TileLang's nativ... | --- python/sglang/kernels/ops/attention/deepseek_v4_rope.py
+++ python/sglang/kernels/ops/attention/deepseek_v4_rope.py
@@ -649,123 +649,3 @@
HAS_WEIGHT=(weight is not None),
USE_POS=(positions is not None),
)
-
-
-# Cache contiguous real/imag halves of each freqs_cis (its .real/.imag are
-# stri... | 0 | 120 | 2 | [
"deepseek"
] | [
"NPU"
] | [] | Optimize this triton kernel file from sglang (python/sglang/kernels/ops/attention/deepseek_v4_rope.py).
Goal: [NPU] Optimize DeepSeek-V4 performance
Target hardware: NPU
Target models: deepseek
Current file:
```python
import logging
import math
from functools import lru_cache
from typing import Optional
import torch... | ## Motivation
This PR improves DeepSeek-V4 serving performance on Ascend NPU by adding PD disaggregation and chunked prefill support, together with optimizations for the prefill and decode execution paths.
All hardware-specific behavior is gated to the NPU + DeepSeek-V4 path, leaving CUDA, ROCm, and other model archi... | [] | null | null |
sglang | 1054060ef1192bceab0e844b8a51fd083cab1b39 | 2026-07-25T19:38:11+08:00 | 31,552 | https://github.com/sgl-project/sglang/pull/31552 | perf: speed up marlin moe with occupancy-aware launch specialization (#31552) | perf: speed up marlin moe with occupancy-aware launch specialization (#31552) | python/sglang/kernels/jit/csrc/gemm/marlin_moe/kernel.h | cpp | Apache-2.0 |
#include <sgl_kernel/scalar_type.hpp>
#include "../marlin/marlin.cuh"
#include "../marlin/marlin_dtypes.cuh"
#define MARLIN_KERNEL_PARAMS \
const int4 *__restrict__ A, const int4 *__restrict__ B, int4 *__restrict__ C, int4 *__r... |
#include <sgl_kernel/scalar_type.hpp>
#include "../marlin/marlin.cuh"
#include "../marlin/marlin_dtypes.cuh"
#define MARLIN_KERNEL_PARAMS \
const int4 *__restrict__ A, const int4 *__restrict__ B, int4 *__restrict__ C, int4 *__r... | --- python/sglang/kernels/jit/csrc/gemm/marlin_moe/kernel.h
+++ python/sglang/kernels/jit/csrc/gemm/marlin_moe/kernel.h
@@ -30,7 +30,9 @@
// fetch pipeline
const int group_blocks, // number of consecutive 16x16 blocks
... | 3 | 1 | 4 | [] | [] | [] | Optimize this cpp kernel file from sglang (python/sglang/kernels/jit/csrc/gemm/marlin_moe/kernel.h).
Goal: perf: speed up marlin moe with occupancy-aware launch specialization
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
#include <sgl_kernel/scalar_type.hpp>
#include "../marlin/marli... | ## Summary
- specialize Marlin MoE kernels for expert-parallel and bias launch paths at JIT compile time, removing per-block branches from the hot loop
- choose large-M Marlin launch configs using register/shared-memory occupancy and available problem parallelism instead of accepting the first valid config
- cover bot... | [{"scope": "e2e", "hardware": "H200", "model": "moonshotai/Kimi-K2.7-Code", "shape": null, "batch_size": null, "metric": "prefill", "unit": "ms", "higher_is_better": false, "baseline": 213.285, "optimized": 204.726, "stated_speedup": null, "speedup": 1.0418}, {"scope": "e2e", "hardware": "H200", "model": "moonshotai/Ki... | 1.0595 | 1.0391 |
sglang | 1054060ef1192bceab0e844b8a51fd083cab1b39 | 2026-07-25T19:38:11+08:00 | 31,552 | https://github.com/sgl-project/sglang/pull/31552 | perf: speed up marlin moe with occupancy-aware launch specialization (#31552) | perf: speed up marlin moe with occupancy-aware launch specialization (#31552) | python/sglang/kernels/jit/csrc/gemm/marlin_moe/marlin_template.h | cpp | Apache-2.0 | /*
* Modified by Neural Magic
* Copyright (C) Marlin.2024 Elias Frantar
*
* 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
*
* Un... | /*
* Modified by Neural Magic
* Copyright (C) Marlin.2024 Elias Frantar
*
* 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
*
* Un... | --- python/sglang/kernels/jit/csrc/gemm/marlin_moe/marlin_template.h
+++ python/sglang/kernels/jit/csrc/gemm/marlin_moe/marlin_template.h
@@ -51,7 +51,9 @@
// fetch pipeline
const int group_blocks, // number of consecutive 16x16 blocks
... | 49 | 32 | 4 | [] | [] | [] | Optimize this cpp kernel file from sglang (python/sglang/kernels/jit/csrc/gemm/marlin_moe/marlin_template.h).
Goal: perf: speed up marlin moe with occupancy-aware launch specialization
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/*
* Modified by Neural Magic
* Copyright (C) Marlin.2... | ## Summary
- specialize Marlin MoE kernels for expert-parallel and bias launch paths at JIT compile time, removing per-block branches from the hot loop
- choose large-M Marlin launch configs using register/shared-memory occupancy and available problem parallelism instead of accepting the first valid config
- cover bot... | [{"scope": "e2e", "hardware": "H200", "model": "moonshotai/Kimi-K2.7-Code", "shape": null, "batch_size": null, "metric": "prefill", "unit": "ms", "higher_is_better": false, "baseline": 213.285, "optimized": 204.726, "stated_speedup": null, "speedup": 1.0418}, {"scope": "e2e", "hardware": "H200", "model": "moonshotai/Ki... | 1.0595 | 1.0391 |
sglang | 1054060ef1192bceab0e844b8a51fd083cab1b39 | 2026-07-25T19:38:11+08:00 | 31,552 | https://github.com/sgl-project/sglang/pull/31552 | perf: speed up marlin moe with occupancy-aware launch specialization (#31552) | perf: speed up marlin moe with occupancy-aware launch specialization (#31552) | python/sglang/kernels/jit/csrc/gemm/marlin_moe/moe_wna16_marlin.cuh | cuda | Apache-2.0 | /*
* Modified by Neural Magic
* Copyright (C) Marlin.2024 Elias Frantar
*
* 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
*
* Un... | /*
* Modified by Neural Magic
* Copyright (C) Marlin.2024 Elias Frantar
*
* 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
*
* Un... | --- python/sglang/kernels/jit/csrc/gemm/marlin_moe/moe_wna16_marlin.cuh
+++ python/sglang/kernels/jit/csrc/gemm/marlin_moe/moe_wna16_marlin.cuh
@@ -154,6 +154,9 @@
thread_config_t tb_cfg;
} exec_config_t;
+constexpr int kSharedMemoryValidityMargin = 512;
+constexpr int kSharedMemoryLaunchReserve = 1024;
+
int ge... | 42 | 25 | 4 | [] | [] | [] | Optimize this cuda kernel file from sglang (python/sglang/kernels/jit/csrc/gemm/marlin_moe/moe_wna16_marlin.cuh).
Goal: perf: speed up marlin moe with occupancy-aware launch specialization
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
/*
* Modified by Neural Magic
* Copyright (C) Mar... | ## Summary
- specialize Marlin MoE kernels for expert-parallel and bias launch paths at JIT compile time, removing per-block branches from the hot loop
- choose large-M Marlin launch configs using register/shared-memory occupancy and available problem parallelism instead of accepting the first valid config
- cover bot... | [{"scope": "e2e", "hardware": "H200", "model": "moonshotai/Kimi-K2.7-Code", "shape": null, "batch_size": null, "metric": "prefill", "unit": "ms", "higher_is_better": false, "baseline": 213.285, "optimized": 204.726, "stated_speedup": null, "speedup": 1.0418}, {"scope": "e2e", "hardware": "H200", "model": "moonshotai/Ki... | 1.0595 | 1.0391 |
sglang | 9402012f0f19857ed9a5a0922dafebaf6382a0e3 | 2026-07-25T08:17:12+08:00 | 32,296 | https://github.com/sgl-project/sglang/pull/32296 | [Perf] Halve the non-finite sanitization overhead in per_token_group_quant (#32296) | [Perf] Halve the non-finite sanitization overhead in per_token_group_quant (#32296)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com> | python/sglang/kernels/jit/csrc/gemm/per_token_group_quant.cuh | cuda | Apache-2.0 | #include <sgl_kernel/tensor.h>
#include <sgl_kernel/utils.h>
#include <sgl_kernel/math.cuh>
#include <sgl_kernel/type.cuh>
#include <sgl_kernel/utils.cuh>
#include <sgl_kernel/vec.cuh>
#include <sgl_kernel/warp.cuh>
#include <sgl_kernel/deepseek_v4/fp8_utils.cuh>
#include <dlpack/dlpack.h>
#include <tvm/ffi/containe... | #include <sgl_kernel/tensor.h>
#include <sgl_kernel/utils.h>
#include <sgl_kernel/math.cuh>
#include <sgl_kernel/type.cuh>
#include <sgl_kernel/utils.cuh>
#include <sgl_kernel/vec.cuh>
#include <sgl_kernel/warp.cuh>
#include <sgl_kernel/deepseek_v4/fp8_utils.cuh>
#include <dlpack/dlpack.h>
#include <tvm/ffi/containe... | --- python/sglang/kernels/jit/csrc/gemm/per_token_group_quant.cuh
+++ python/sglang/kernels/jit/csrc/gemm/per_token_group_quant.cuh
@@ -48,14 +48,16 @@
using packed2_t = fp8x2_e4m3_t;
static constexpr float kMaxValue = DTypeTrait<fp8_e4m3_t>::kFloatMax;
// SATFINITE saturates +-inf / out-of-range values, but c... | 17 | 14 | 1 | [] | [] | [] | Optimize this cuda kernel file from sglang (python/sglang/kernels/jit/csrc/gemm/per_token_group_quant.cuh).
Goal: [Perf] Halve the non-finite sanitization overhead in per_token_group_quant
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <sgl_kernel/tensor.h>
#include <sgl_kernel... | ## Motivation
#32188 fixed the H100 deepep TBO CI break by sanitizing non-finite quant inputs with a **two-sided clamp** before the fp8 conversion. That fix is correct, but it costs more than it needs to: the trait-driven quant kernel is latency-bound at decode batch sizes, and the clamp added 16 `HMNMX2` (ue8m0 path)... | [{"scope": "kernel", "hardware": "B200", "model": null, "shape": "gs=32/64/128", "batch_size": null, "metric": "speedup", "unit": null, "higher_is_better": true, "baseline": null, "optimized": null, "stated_speedup": 1.009, "speedup": 1.009}, {"scope": "kernel", "hardware": "B200", "model": null, "shape": "gs=32/64/128... | 1.009 | null |
sglang | 1b63155efead7494843f6db8681851ba94611f73 | 2026-07-22T22:00:25-07:00 | 32,109 | https://github.com/sgl-project/sglang/pull/32109 | [Perf] Skip blocks past per-request live length in full-width Triton kernels (#32109) | [Perf] Skip blocks past per-request live length in full-width Triton kernels (#32109) | python/sglang/kernels/ops/attention/dsa/index_buf_accessor.py | triton | Apache-2.0 | from typing import TYPE_CHECKING
import torch
import triton
import triton.language as tl
from sglang.kernels.ops.quantization.fp8_kernel import is_fp8_fnuz
from sglang.srt.layers.attention.dsa.utils import aiter_can_use_preshuffle_paged_mqa
from sglang.srt.utils import get_bool_env_var, is_hip
_is_hip = is_hip()
_is... | from typing import TYPE_CHECKING
import torch
import triton
import triton.language as tl
from sglang.kernels.ops.quantization.fp8_kernel import is_fp8_fnuz
from sglang.srt.layers.attention.dsa.utils import aiter_can_use_preshuffle_paged_mqa
from sglang.srt.utils import get_bool_env_var, is_hip
_is_hip = is_hip()
_is... | --- python/sglang/kernels/ops/attention/dsa/index_buf_accessor.py
+++ python/sglang/kernels/ops/attention/dsa/index_buf_accessor.py
@@ -639,6 +639,9 @@
k_offsets = thread_idx * BLOCK_SIZE_K + tl.arange(0, BLOCK_SIZE_K)
seq_len = tl.load(seq_len_ptr + batch_id)
+ # Grid axis 1 spans the batch-max seq len;... | 3 | 0 | 6 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/kernels/ops/attention/dsa/index_buf_accessor.py).
Goal: [Perf] Skip blocks past per-request live length in full-width Triton kernels
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from typing import TYPE_CHECKING
import torc... | Extend the per-request early-exit pattern from #31981 to the remaining kernels whose grid or inner loop spans the full captured width: DSA decode / target-verify metadata, FA3 decode metadata, TRTLLM MHA graph metadata, extend attention (regular and unified), the DSA indexer K+S gather, and the sparse prefill index tra... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "width=16384,seq=1024", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 5.87, "optimized": 5.73, "stated_speedup": null, "speedup": 1.0244}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batc... | 1.0227 | null |
sglang | 1b63155efead7494843f6db8681851ba94611f73 | 2026-07-22T22:00:25-07:00 | 32,109 | https://github.com/sgl-project/sglang/pull/32109 | [Perf] Skip blocks past per-request live length in full-width Triton kernels (#32109) | [Perf] Skip blocks past per-request live length in full-width Triton kernels (#32109) | python/sglang/kernels/ops/attention/dsa/transform_index.py | triton | Apache-2.0 | from itertools import accumulate
from typing import List, Optional
import torch
import triton
import triton.language as tl
def transform_index_page_table_prefill(**kwargs):
return transform_index_page_table_prefill_fast(**kwargs)
def transform_index_page_table_decode(**kwargs):
return transform_index_page_... | from itertools import accumulate
from typing import List, Optional
import torch
import triton
import triton.language as tl
def transform_index_page_table_prefill(**kwargs):
return transform_index_page_table_prefill_fast(**kwargs)
def transform_index_page_table_decode(**kwargs):
return transform_index_page_... | --- python/sglang/kernels/ops/attention/dsa/transform_index.py
+++ python/sglang/kernels/ops/attention/dsa/transform_index.py
@@ -87,6 +87,9 @@
query_start = tl.load(cu_seqlens_q_ptr + request_id)
query_end = tl.load(cu_seqlens_q_ptr + request_id + 1)
+ # Grid axis 1 spans the batch-max extend len; fully... | 3 | 0 | 6 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/kernels/ops/attention/dsa/transform_index.py).
Goal: [Perf] Skip blocks past per-request live length in full-width Triton kernels
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from itertools import accumulate
from typing imp... | Extend the per-request early-exit pattern from #31981 to the remaining kernels whose grid or inner loop spans the full captured width: DSA decode / target-verify metadata, FA3 decode metadata, TRTLLM MHA graph metadata, extend attention (regular and unified), the DSA indexer K+S gather, and the sparse prefill index tra... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "width=16384,seq=1024", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 5.87, "optimized": 5.73, "stated_speedup": null, "speedup": 1.0244}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batc... | 1.0227 | null |
sglang | 1b63155efead7494843f6db8681851ba94611f73 | 2026-07-22T22:00:25-07:00 | 32,109 | https://github.com/sgl-project/sglang/pull/32109 | [Perf] Skip blocks past per-request live length in full-width Triton kernels (#32109) | [Perf] Skip blocks past per-request live length in full-width Triton kernels (#32109) | python/sglang/kernels/ops/attention/dsa_metadata.py | triton | Apache-2.0 | from typing import Optional
import torch
import triton
import triton.language as tl
@triton.jit(
do_not_specialize=[
"page_table_stride_0",
"real_page_table_stride_0",
"max_len",
]
)
def _fused_dsa_decode_metadata_kernel(
seq_lens,
req_pool_indices,
req_to_token,
cache... | from typing import Optional
import torch
import triton
import triton.language as tl
@triton.jit(
do_not_specialize=[
"page_table_stride_0",
"real_page_table_stride_0",
"max_len",
]
)
def _fused_dsa_decode_metadata_kernel(
seq_lens,
req_pool_indices,
req_to_token,
cache... | --- python/sglang/kernels/ops/attention/dsa_metadata.py
+++ python/sglang/kernels/ops/attention/dsa_metadata.py
@@ -71,6 +71,16 @@
mask=row < bs,
other=0,
)
+ # Skip column blocks past the request's kv length: no consumer reads there
+ # (attention and the indexer both stay within cache_seq... | 28 | 0 | 6 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/kernels/ops/attention/dsa_metadata.py).
Goal: [Perf] Skip blocks past per-request live length in full-width Triton kernels
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from typing import Optional
import torch
import triton... | Extend the per-request early-exit pattern from #31981 to the remaining kernels whose grid or inner loop spans the full captured width: DSA decode / target-verify metadata, FA3 decode metadata, TRTLLM MHA graph metadata, extend attention (regular and unified), the DSA indexer K+S gather, and the sparse prefill index tra... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "width=16384,seq=1024", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 5.87, "optimized": 5.73, "stated_speedup": null, "speedup": 1.0244}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batc... | 1.0227 | null |
sglang | 1b63155efead7494843f6db8681851ba94611f73 | 2026-07-22T22:00:25-07:00 | 32,109 | https://github.com/sgl-project/sglang/pull/32109 | [Perf] Skip blocks past per-request live length in full-width Triton kernels (#32109) | [Perf] Skip blocks past per-request live length in full-width Triton kernels (#32109) | python/sglang/kernels/ops/attention/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/kernels/ops/attention/extend_attention.py
+++ python/sglang/kernels/ops/attention/extend_attention.py
@@ -313,6 +313,10 @@
cur_seq_len_prefix = tl.load(kv_indptr + cur_seq + 1) - cur_seq_kv_start_idx
cur_seq_len = cur_seq_len_prefix + cur_seq_len_extend
+ # Grid axis 2 spans the batch-m... | 8 | 0 | 6 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/kernels/ops/attention/extend_attention.py).
Goal: [Perf] Skip blocks past per-request live length in full-width Triton kernels
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2023-2024 SGLang Team
# Licensed under ... | Extend the per-request early-exit pattern from #31981 to the remaining kernels whose grid or inner loop spans the full captured width: DSA decode / target-verify metadata, FA3 decode metadata, TRTLLM MHA graph metadata, extend attention (regular and unified), the DSA indexer K+S gather, and the sparse prefill index tra... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "width=16384,seq=1024", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 5.87, "optimized": 5.73, "stated_speedup": null, "speedup": 1.0244}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batc... | 1.0227 | null |
sglang | 1b63155efead7494843f6db8681851ba94611f73 | 2026-07-22T22:00:25-07:00 | 32,109 | https://github.com/sgl-project/sglang/pull/32109 | [Perf] Skip blocks past per-request live length in full-width Triton kernels (#32109) | [Perf] Skip blocks past per-request live length in full-width Triton kernels (#32109) | python/sglang/kernels/ops/attention/metadata.py | triton | Apache-2.0 | from typing import TYPE_CHECKING, Optional
import torch
import triton
import triton.language as tl
if TYPE_CHECKING:
from sglang.srt.mem_cache.swa_memory_pool import SWAKVPool
@triton.jit
def get_num_kv_splits_triton(
num_kv_splits_ptr,
seq_lens_ptr,
num_seq,
num_group,
num_head,
num_kv_... | from typing import TYPE_CHECKING, Optional
import torch
import triton
import triton.language as tl
if TYPE_CHECKING:
from sglang.srt.mem_cache.swa_memory_pool import SWAKVPool
@triton.jit
def get_num_kv_splits_triton(
num_kv_splits_ptr,
seq_lens_ptr,
num_seq,
num_group,
num_head,
num_kv_... | --- python/sglang/kernels/ops/attention/metadata.py
+++ python/sglang/kernels/ops/attention/metadata.py
@@ -213,13 +213,25 @@
return
i = pid_b
+ # Self-guard on the device-side seq_len: skip column chunks past the
+ # request's live pages (tails keep stale values the attention kernels
+ # never... | 28 | 4 | 6 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/kernels/ops/attention/metadata.py).
Goal: [Perf] Skip blocks past per-request live length in full-width Triton kernels
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from typing import TYPE_CHECKING, Optional
import torch
im... | Extend the per-request early-exit pattern from #31981 to the remaining kernels whose grid or inner loop spans the full captured width: DSA decode / target-verify metadata, FA3 decode metadata, TRTLLM MHA graph metadata, extend attention (regular and unified), the DSA indexer K+S gather, and the sparse prefill index tra... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "width=16384,seq=1024", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 5.87, "optimized": 5.73, "stated_speedup": null, "speedup": 1.0244}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batc... | 1.0227 | null |
sglang | 1b63155efead7494843f6db8681851ba94611f73 | 2026-07-22T22:00:25-07:00 | 32,109 | https://github.com/sgl-project/sglang/pull/32109 | [Perf] Skip blocks past per-request live length in full-width Triton kernels (#32109) | [Perf] Skip blocks past per-request live length in full-width Triton kernels (#32109) | python/sglang/kernels/ops/kvcache/trtllm_mha_graph_metadata.py | triton | Apache-2.0 | """Fused CUDA-graph metadata update for the TRTLLM MHA backend.
`TRTLLMHAAttnBackend._apply_cuda_graph_metadata` used to rebuild the
page table(s) and seqlen buffers with ~25 small aten ops per graph
replay (index gathers, floor_divide, cumsum, dtype casts, copies).
On some CPUs that is ~0.7-1.0 ms of pure host dispat... | """Fused CUDA-graph metadata update for the TRTLLM MHA backend.
`TRTLLMHAAttnBackend._apply_cuda_graph_metadata` used to rebuild the
page table(s) and seqlen buffers with ~25 small aten ops per graph
replay (index gathers, floor_divide, cumsum, dtype casts, copies).
On some CPUs that is ~0.7-1.0 ms of pure host dispat... | --- python/sglang/kernels/ops/kvcache/trtllm_mha_graph_metadata.py
+++ python/sglang/kernels/ops/kvcache/trtllm_mha_graph_metadata.py
@@ -75,9 +75,12 @@
row_out = page_table_ptr + pid.to(tl.int64) * page_table_stride
if HAS_SWA:
swa_row_out = swa_page_table_ptr + pid.to(tl.int64) * swa_pa... | 11 | 3 | 6 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/kernels/ops/kvcache/trtllm_mha_graph_metadata.py).
Goal: [Perf] Skip blocks past per-request live length in full-width Triton kernels
Target hardware: unspecified
Target models: unspecified
Current file:
```python
"""Fused CUDA-graph metadata update for the ... | Extend the per-request early-exit pattern from #31981 to the remaining kernels whose grid or inner loop spans the full captured width: DSA decode / target-verify metadata, FA3 decode metadata, TRTLLM MHA graph metadata, extend attention (regular and unified), the DSA indexer K+S gather, and the sparse prefill index tra... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "width=16384,seq=1024", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 5.87, "optimized": 5.73, "stated_speedup": null, "speedup": 1.0244}, {"scope": "kernel", "hardware": null, "model": null, "shape": null, "batc... | 1.0227 | null |
sglang | 08cb081ed0bbb75d61233d343b52fb306801db74 | 2026-07-21T16:20:14-07:00 | 31,981 | https://github.com/sgl-project/sglang/pull/31981 | [Perf] Skip page-table columns past kv length in DSA draft-extend metadata kernel (#31981) | [Perf] Skip page-table columns past kv length in DSA draft-extend metadata kernel (#31981) | python/sglang/kernels/ops/attention/dsa_metadata.py | triton | Apache-2.0 | from typing import Optional
import torch
import triton
import triton.language as tl
@triton.jit(
do_not_specialize=[
"page_table_stride_0",
"real_page_table_stride_0",
"max_len",
]
)
def _fused_dsa_decode_metadata_kernel(
seq_lens,
req_pool_indices,
req_to_token,
cache... | from typing import Optional
import torch
import triton
import triton.language as tl
@triton.jit(
do_not_specialize=[
"page_table_stride_0",
"real_page_table_stride_0",
"max_len",
]
)
def _fused_dsa_decode_metadata_kernel(
seq_lens,
req_pool_indices,
req_to_token,
cache... | --- python/sglang/kernels/ops/attention/dsa_metadata.py
+++ python/sglang/kernels/ops/attention/dsa_metadata.py
@@ -524,6 +524,15 @@
mask=req_row < bs,
other=0,
).to(tl.int32)
+ # Skip column blocks past the request's kv length: no consumer reads there
+ # (attention and the indexer both st... | 9 | 0 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/kernels/ops/attention/dsa_metadata.py).
Goal: [Perf] Skip page-table columns past kv length in DSA draft-extend metadata kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from typing import Optional
import torch
import t... | The DSA draft-extend metadata kernel copies the page table over the full captured context width, but each request's real kv length is far smaller. Add a per-program early-exit: skip whole column blocks whose start is past the request's kv length (downstream attention and the indexer both bound their reads by `cache_seq... | [{"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 74.9, "optimized": 40.9, "stated_speedup": null, "speedup": 1.8313}] | 1.8313 | null |
sglang | 5ce5e1ee3efe934d9737d382e410165381572af8 | 2026-07-10T09:09:38+08:00 | 30,716 | https://github.com/sgl-project/sglang/pull/30716 | [Diffusion] Revert CPU AMX optimizations (#30716) | [Diffusion] Revert CPU AMX optimizations (#30716) | sgl-kernel/csrc/cpu/flash_attn.cpp | cpp | BSD-3-Clause | /*****************************************************************************************
* Copyright (c) 2025 - 2025 Codeplay Software Ltd. All rights reserved.
* Copyright (C) 2025 Intel Corporation, All rights reserved.
* SPDX-License-Identifier: BSD-3-Clause
*
* Redistribution and use in source and binary for... | /*****************************************************************************************
* Copyright (c) 2025 - 2025 Codeplay Software Ltd. All rights reserved.
* Copyright (C) 2025 Intel Corporation, All rights reserved.
* SPDX-License-Identifier: BSD-3-Clause
*
* Redistribution and use in source and binary for... | --- sgl-kernel/csrc/cpu/flash_attn.cpp
+++ sgl-kernel/csrc/cpu/flash_attn.cpp
@@ -443,8 +443,7 @@
const at::Tensor& cu_seqlens_k,
int64_t max_seqlen_q,
int64_t max_seqlen_k,
- bool causal,
- const std::optional<double>& sm_scale) {
+ bool causal) {
CHECK_LAST_DIM_CONTIGUOUS_INPUT(q);
CHECK... | 4 | 5 | 3 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/flash_attn.cpp).
Goal: [Diffusion] Revert CPU AMX optimizations
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
/*****************************************************************************************
* Copyright (c) 2025 - 2025 Co... | Reverts #28527 (`177c048c68`).
That change adds a second unconditional `process_weights_after_loading()` pass in the diffusion FSDP loader. The hook is not idempotent for quantized weights, and the added `device_loading_context` also breaks FSDP inference by attempting to restore CUDA-backed parameters to CPU storage.... | [] | null | null |
sglang | 5ce5e1ee3efe934d9737d382e410165381572af8 | 2026-07-10T09:09:38+08:00 | 30,716 | https://github.com/sgl-project/sglang/pull/30716 | [Diffusion] Revert CPU AMX optimizations (#30716) | [Diffusion] Revert CPU AMX optimizations (#30716) | 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
@@ -235,8 +235,7 @@
const at::Tensor& cu_seqlens_k,
int64_t max_seqlen_q,
int64_t max_seqlen_k,
- bool causal,
- const std::optional<double>& sm_scale);
+ bool causal);
// linear attention
std::tu... | 2 | 3 | 3 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/torch_extension_cpu.cpp).
Goal: [Diffusion] Revert CPU AMX optimizations
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License... | Reverts #28527 (`177c048c68`).
That change adds a second unconditional `process_weights_after_loading()` pass in the diffusion FSDP loader. The hook is not idempotent for quantized weights, and the added `device_loading_context` also breaks FSDP inference by attempting to restore CUDA-backed parameters to CPU storage.... | [] | null | null |
sglang | 177c048c68b38930a675fac1e0b860b4c668b95f | 2026-07-09T10:26:22+08:00 | 28,527 | https://github.com/sgl-project/sglang/pull/28527 | [Diffusion][CPU] Adding AMX optimizations for CPU platform (#28527) | [Diffusion][CPU] Adding AMX optimizations for CPU platform (#28527)
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com> | sgl-kernel/csrc/cpu/flash_attn.cpp | cpp | BSD-3-Clause | /*****************************************************************************************
* Copyright (c) 2025 - 2025 Codeplay Software Ltd. All rights reserved.
* Copyright (C) 2025 Intel Corporation, All rights reserved.
* SPDX-License-Identifier: BSD-3-Clause
*
* Redistribution and use in source and binary for... | /*****************************************************************************************
* Copyright (c) 2025 - 2025 Codeplay Software Ltd. All rights reserved.
* Copyright (C) 2025 Intel Corporation, All rights reserved.
* SPDX-License-Identifier: BSD-3-Clause
*
* Redistribution and use in source and binary for... | --- sgl-kernel/csrc/cpu/flash_attn.cpp
+++ sgl-kernel/csrc/cpu/flash_attn.cpp
@@ -443,7 +443,8 @@
const at::Tensor& cu_seqlens_k,
int64_t max_seqlen_q,
int64_t max_seqlen_k,
- bool causal) {
+ bool causal,
+ const std::optional<double>& sm_scale) {
CHECK_LAST_DIM_CONTIGUOUS_INPUT(q);
CHECK... | 5 | 4 | 3 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/flash_attn.cpp).
Goal: [Diffusion][CPU] Adding AMX optimizations for CPU platform
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
/*****************************************************************************************
* Copyright ... | This pr takes parts of follow-ups (mentioned in https://github.com/sgl-project/sglang/pull/20816) to bring key AMX based optimizations (scoping from LLM models) for CPU platforms, including:
- Adding CPU AMX backend, replace SDPA to bring best attention perf.
- Similar like for LLM models, adding replacement for amx w... | [{"scope": "e2e", "hardware": null, "model": "Wan-AI/Wan2.2-TI2V-5B-Diffusers", "shape": null, "batch_size": null, "metric": "speedup", "unit": null, "higher_is_better": true, "baseline": null, "optimized": null, "stated_speedup": 10.0, "speedup": 10.0}, {"scope": "e2e", "hardware": null, "model": "Wan-AI/Wan2.2-TI2V-5... | null | 2 |
sglang | 177c048c68b38930a675fac1e0b860b4c668b95f | 2026-07-09T10:26:22+08:00 | 28,527 | https://github.com/sgl-project/sglang/pull/28527 | [Diffusion][CPU] Adding AMX optimizations for CPU platform (#28527) | [Diffusion][CPU] Adding AMX optimizations for CPU platform (#28527)
Co-authored-by: Ma Mingfei <mingfei.ma@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
@@ -153,7 +153,8 @@
const at::Tensor& cu_seqlens_k,
int64_t max_seqlen_q,
int64_t max_seqlen_k,
- bool causal);
+ bool causal,
+ const std::optional<double>& sm_scale);
// linear attention
std::tu... | 3 | 2 | 3 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/torch_extension_cpu.cpp).
Goal: [Diffusion][CPU] Adding AMX optimizations for CPU platform
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version... | This pr takes parts of follow-ups (mentioned in https://github.com/sgl-project/sglang/pull/20816) to bring key AMX based optimizations (scoping from LLM models) for CPU platforms, including:
- Adding CPU AMX backend, replace SDPA to bring best attention perf.
- Similar like for LLM models, adding replacement for amx w... | [{"scope": "e2e", "hardware": null, "model": "Wan-AI/Wan2.2-TI2V-5B-Diffusers", "shape": null, "batch_size": null, "metric": "speedup", "unit": null, "higher_is_better": true, "baseline": null, "optimized": null, "stated_speedup": 10.0, "speedup": 10.0}, {"scope": "e2e", "hardware": null, "model": "Wan-AI/Wan2.2-TI2V-5... | null | 2 |
sglang | a2d7eb303eb8fef99276cbf126dded16a0c23bc8 | 2026-07-03T11:18:59+08:00 | 26,771 | https://github.com/sgl-project/sglang/pull/26771 | [MoE] Consolidate ungrouped + grouped gate/topk onto one Triton router (#26771) — faster than AOT on B200/H100/H200, at parity with flashinfer (#29771) | [MoE] Consolidate ungrouped + grouped gate/topk onto one Triton router (#26771) — faster than AOT on B200/H100/H200, at parity with flashinfer (#29771)
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> | python/sglang/srt/layers/moe/topk.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/moe/topk.py
+++ python/sglang/srt/layers/moe/topk.py
@@ -789,6 +789,24 @@
num_token_non_padded=num_token_non_padded,
)
# ===== END TO BE REFACTORED ====
+ elif _is_cuda and envs.SGLANG_OPT_USE_JIT_KERNEL_FUSED_TOPK.get():
+ # Unified ... | 80 | 14 | 1 | [] | [
"B200",
"H100",
"H200"
] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/topk.py).
Goal: [MoE] Consolidate ungrouped + grouped gate/topk onto one Triton router (#26771) — faster than AOT on B200/H100/H200, at parity with flashinfer
Target hardware: B200, H100, H200
Target models: unspecified
Current file:
```python... | [] | null | null | |
sglang | 56f22cd520cbdf58cb89c688d7847c1186df6bc3 | 2026-07-01T10:11:29+08:00 | 28,612 | https://github.com/sgl-project/sglang/pull/28612 | Optimize C128 state pool allocation using request state pool (#28612) | Optimize C128 state pool allocation using request state pool (#28612) | python/sglang/srt/layers/attention/dsv4/compress_hip.py | triton | Apache-2.0 | from __future__ import annotations
import os
from functools import cached_property
from typing import TYPE_CHECKING, Any
import torch
import torch.nn as nn
import triton
import triton.language as tl
from sglang.srt.environ import envs
from sglang.srt.layers.attention.dsa.dsa_indexer import rotate_activation
from sgl... | from __future__ import annotations
import os
from functools import cached_property
from typing import TYPE_CHECKING, Any
import torch
import torch.nn as nn
import triton
import triton.language as tl
from sglang.srt.environ import envs
from sglang.srt.layers.attention.dsa.dsa_indexer import rotate_activation
from sgl... | --- python/sglang/srt/layers/attention/dsv4/compress_hip.py
+++ python/sglang/srt/layers/attention/dsv4/compress_hip.py
@@ -210,13 +210,18 @@
pre_state_indices = self.compute_state_len_indices(
seq_len=prefix_lens[i], ratio=self.ratio
).to(device)
- raw_loc = torch.... | 54 | 31 | 5 | [
"deepseek"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/dsv4/compress_hip.py).
Goal: Optimize C128 state pool allocation using request state pool
Target hardware: unspecified
Target models: deepseek
Current file:
```python
from __future__ import annotations
import os
from functools import ca... | <!-- 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.io to discuss further. -->
## Motivation
Co-authored-by: [shiyu7](https://github.com/s... | [] | null | null |
sglang | 56f22cd520cbdf58cb89c688d7847c1186df6bc3 | 2026-07-01T10:11:29+08:00 | 28,612 | https://github.com/sgl-project/sglang/pull/28612 | Optimize C128 state pool allocation using request state pool (#28612) | Optimize C128 state pool allocation using request state pool (#28612) | python/sglang/srt/layers/attention/dsv4/compressor_v2.py | triton | Apache-2.0 | from __future__ import annotations
from typing import TYPE_CHECKING, List, Literal, Optional, TypeAlias, Union, cast
import torch
from sglang.jit_kernel.dsv4 import (
CompressorDecodePlan,
CompressorPrefillPlan,
compress_forward,
compress_norm_rope_store,
)
from sglang.jit_kernel.utils import is_hip_... | from __future__ import annotations
from typing import TYPE_CHECKING, List, Literal, Optional, TypeAlias, Union, cast
import torch
from sglang.jit_kernel.dsv4 import (
CompressorDecodePlan,
CompressorPrefillPlan,
compress_forward,
compress_norm_rope_store,
)
from sglang.jit_kernel.utils import is_hip_... | --- python/sglang/srt/layers/attention/dsv4/compressor_v2.py
+++ python/sglang/srt/layers/attention/dsv4/compressor_v2.py
@@ -731,7 +731,7 @@
seq_lens=seq_lens_planner,
extend_lens=extend_lens_planner,
req_to_token=req_to_token,
- full_to_swa=full_to_swa,
+ f... | 2 | 8 | 5 | [
"deepseek"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/dsv4/compressor_v2.py).
Goal: Optimize C128 state pool allocation using request state pool
Target hardware: unspecified
Target models: deepseek
Current file:
```python
from __future__ import annotations
from typing import TYPE_CHECKING,... | <!-- 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.io to discuss further. -->
## Motivation
Co-authored-by: [shiyu7](https://github.com/s... | [] | null | null |
sglang | bd3b252e0af0e7bfbcc1f329a672a7ee47a36721 | 2026-06-29T09:38:30+08:00 | 29,382 | https://github.com/sgl-project/sglang/pull/29382 | [CPU] use faster exp in silu_and_mul (#29382) | [CPU] use faster exp in silu_and_mul (#29382) | sgl-kernel/csrc/cpu/activation.cpp | cpp | Apache-2.0 | #include "common.h"
#include "vec.h"
namespace {
template <typename scalar_t, typename func_t, typename vec_func_t>
void act_and_mul_kernel_impl(
scalar_t* __restrict__ output,
const scalar_t* __restrict__ input,
int64_t num_tokens,
int64_t dim,
const func_t& f,
const vec_func_t& vf) {
using... | #include "common.h"
#include "vec.h"
namespace {
template <typename scalar_t, typename func_t, typename vec_func_t>
void act_and_mul_kernel_impl(
scalar_t* __restrict__ output,
const scalar_t* __restrict__ input,
int64_t num_tokens,
int64_t dim,
const func_t& f,
const vec_func_t& vf) {
using... | --- sgl-kernel/csrc/cpu/activation.cpp
+++ sgl-kernel/csrc/cpu/activation.cpp
@@ -121,7 +121,7 @@
num_tokens,
d,
[](float x) { return x / (1.f + std::exp(-x)); },
- [](Vec x) { return x / (Vec(1.f) + x.neg().exp()); });
+ [](Vec x) { return x / (Vec(1.f) + x.neg().exp_u20()); })... | 1 | 1 | 1 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/activation.cpp).
Goal: [CPU] use faster exp in silu_and_mul
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "common.h"
#include "vec.h"
namespace {
template <typename scalar_t, typename func_t, typename vec_func_t>
void act... | <!-- 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.io to discuss further. -->
## Motivation
This PR speeds up the CPU `silu_and_mul_cpu` ... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "N=1,K=18432", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 10.311, "optimized": 9.103, "stated_speedup": 1.133, "speedup": 1.1327}, {"scope": "kernel", "hardware": null, "model": null, "shape": "N=17,K=18432", ... | 1.1635 | null |
sglang | 2c3f007a656b8c9414d15320f9a49f15420950b9 | 2026-06-25T09:04:34+08:00 | 29,117 | https://github.com/sgl-project/sglang/pull/29117 | [CPU] optimize GDN prefill performance (#29117) | [CPU] optimize GDN prefill performance (#29117) | 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
@@ -287,6 +287,77 @@
// transpose utils
// taken from my PR in ggml: https://github.com/ggml-org/llama.cpp/pull/8998
#if defined(CPU_CAPABILITY_AVX512)
+inline void transpose_16x16_16bit(__m256i* v) {
+ __m256i v1[16];
+ v1[0] = _mm256_unpacklo_epi16(v[0]... | 71 | 0 | 2 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/vec.h).
Goal: [CPU] optimize GDN prefill performance
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#pragma once
#if defined(__AVX512F__) && defined(__AVX512BF16__) && defined(__AMX_BF16__)
#define CPU_CAPABILITY_AVX512
#endif
#inc... | <!-- 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.io to discuss further. -->
## Motivation
Optimize the CPU GDN (Gated Delta Network) ke... | [] | null | null |
sglang | d5e9176f6581e3c9274dd33a95315033aa194df5 | 2026-06-24T22:13:35+02:00 | 27,053 | https://github.com/sgl-project/sglang/pull/27053 | [BCG][GLM5] perf: BCG support and prefill enhancements (#27053) | [BCG][GLM5] perf: BCG support and prefill enhancements (#27053) | python/sglang/srt/layers/attention/dsa/utils.py | triton | Apache-2.0 | from functools import lru_cache
from typing import TYPE_CHECKING, List, Tuple, Union
import torch
import triton
import triton.language as tl
from sglang.srt.environ import envs
from sglang.srt.layers.dp_attention import (
DpPaddingMode,
)
from sglang.srt.runtime_context import get_parallel
from sglang.srt.server_... | from functools import lru_cache
from typing import TYPE_CHECKING, List, Tuple, Union
import torch
import triton
import triton.language as tl
from sglang.srt.environ import envs
from sglang.srt.layers.dp_attention import (
DpPaddingMode,
)
from sglang.srt.model_executor.runner_backend_utils.breakable_cuda_graph im... | --- python/sglang/srt/layers/attention/dsa/utils.py
+++ python/sglang/srt/layers/attention/dsa/utils.py
@@ -9,9 +9,15 @@
from sglang.srt.layers.dp_attention import (
DpPaddingMode,
)
+from sglang.srt.model_executor.runner_backend_utils.breakable_cuda_graph import (
+ is_in_breakable_cuda_graph,
+)
+from sglan... | 20 | 1 | 3 | [
"glm"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/dsa/utils.py).
Goal: [BCG][GLM5] perf: BCG support and prefill enhancements
Target hardware: unspecified
Target models: glm
Current file:
```python
from functools import lru_cache
from typing import TYPE_CHECKING, List, Tuple, Union
imp... | ## Motivation
https://github.com/sgl-project/sglang/pull/23351 introduced PCG for GLM5 (`nsa_indexer`). As the current implementation runs the full indexer path, which is not fully captured by a CUDA graph, we found that running the fast indexer in eager mode yields better performance.
Moreover, the current PCG split... | [{"scope": "e2e", "hardware": "B200", "model": null, "shape": "1k1k", "batch_size": 4, "metric": "TTFT", "unit": "ms", "higher_is_better": false, "baseline": 88.74, "optimized": 78.18, "stated_speedup": null, "speedup": 1.1351}, {"scope": "e2e", "hardware": "B200", "model": null, "shape": "1k1k", "batch_size": 8, "metr... | null | 1.011 |
sglang | af9027f6c938b07995e507ac78496ceef3014d31 | 2026-06-23T17:20:48+08:00 | 28,938 | https://github.com/sgl-project/sglang/pull/28938 | [AMD] Improve performance of dsv4 in high concurrency (#28938) | [AMD] Improve performance of dsv4 in high concurrency (#28938)
Co-authored-by: wunhuang <wunhuang@amd.com> | python/sglang/srt/layers/deepseek_v4_rope.py | triton | Apache-2.0 | import logging
import math
from functools import lru_cache
from typing import Optional
import torch
import triton
import triton.language as tl
logger = logging.getLogger(__name__)
# tilelang isn't shipped on every platform (e.g. Ascend NPU images) and the
# only tilelang artifacts in this file are pass_configs that ... | import logging
import math
from functools import lru_cache
from typing import Optional
import torch
import triton
import triton.language as tl
logger = logging.getLogger(__name__)
# tilelang isn't shipped on every platform (e.g. Ascend NPU images) and the
# only tilelang artifacts in this file are pass_configs that ... | --- python/sglang/srt/layers/deepseek_v4_rope.py
+++ python/sglang/srt/layers/deepseek_v4_rope.py
@@ -161,7 +161,7 @@
BLOCK_M: tl.constexpr,
BLOCK_P: tl.constexpr,
):
- # Batched variant: BLOCK_M tokens per program (mirrors ATOM's inverse_rope_gptj
+ # Batched variant: BLOCK_M tokens per program
#... | 43 | 32 | 2 | [
"deepseek"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/deepseek_v4_rope.py).
Goal: [AMD] Improve performance of dsv4 in high concurrency
Target hardware: unspecified
Target models: deepseek
Current file:
```python
import logging
import math
from functools import lru_cache
from typing import Optional
... | ## Motivation
DeepSeek-V4 on AMD (MI35X, HIP) under high concurrency (TP8 + DP-attention, c512) trails. Per-layer trace analysis (pure-prefill + a c512/ISL8192 mixed trace) showed the gap was **not** the heavy GEMM/MoE kernels but a few avoidable overheads in the prefill / dp-attention path:
- the attention-output (i... | [{"scope": "e2e", "hardware": "MI35X", "model": "DeepSeek-V4-Pro", "shape": null, "batch_size": 512, "metric": "throughput", "unit": "tok/s", "higher_is_better": true, "baseline": 17800, "optimized": 18088, "stated_speedup": null, "speedup": 1.0162}, {"scope": "e2e", "hardware": "MI35X", "model": "DeepSeek-V4-Pro", "sh... | 2.8475 | 1.0589 |
sglang | af9027f6c938b07995e507ac78496ceef3014d31 | 2026-06-23T17:20:48+08:00 | 28,938 | https://github.com/sgl-project/sglang/pull/28938 | [AMD] Improve performance of dsv4 in high concurrency (#28938) | [AMD] Improve performance of dsv4 in high concurrency (#28938)
Co-authored-by: wunhuang <wunhuang@amd.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 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_at... | 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_at... | --- python/sglang/srt/layers/dp_attention.py
+++ python/sglang/srt/layers/dp_attention.py
@@ -580,13 +580,11 @@
else:
local_real = local_tokens.new_zeros((local_rows, *local_tokens.shape[1:]))
local_real[: local_tokens.shape[0]].copy_(local_tokens)
- gathered = get_tp_group().all_gatherv(local... | 5 | 7 | 2 | [
"deepseek"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/dp_attention.py).
Goal: [AMD] Improve performance of dsv4 in high concurrency
Target hardware: unspecified
Target models: deepseek
Current file:
```python
from __future__ import annotations
import functools
import logging
from contextlib import c... | ## Motivation
DeepSeek-V4 on AMD (MI35X, HIP) under high concurrency (TP8 + DP-attention, c512) trails. Per-layer trace analysis (pure-prefill + a c512/ISL8192 mixed trace) showed the gap was **not** the heavy GEMM/MoE kernels but a few avoidable overheads in the prefill / dp-attention path:
- the attention-output (i... | [{"scope": "e2e", "hardware": "MI35X", "model": "DeepSeek-V4-Pro", "shape": null, "batch_size": 512, "metric": "throughput", "unit": "tok/s", "higher_is_better": true, "baseline": 17800, "optimized": 18088, "stated_speedup": null, "speedup": 1.0162}, {"scope": "e2e", "hardware": "MI35X", "model": "DeepSeek-V4-Pro", "sh... | 2.8475 | 1.0589 |
sglang | 47cad39f345a28c4b70069c02c27d08b603c7aa7 | 2026-06-20T17:11:01+08:00 | 28,722 | https://github.com/sgl-project/sglang/pull/28722 | [AMD] Optimize o_proj gemm and attn output rope performance (#28722) | [AMD] Optimize o_proj gemm and attn output rope performance (#28722)
Co-authored-by: wunhuang <wunhuang@amd.com> | python/sglang/srt/layers/deepseek_v4_rope.py | triton | Apache-2.0 | import logging
import math
from functools import lru_cache
from typing import Optional
import torch
import triton
import triton.language as tl
logger = logging.getLogger(__name__)
# tilelang isn't shipped on every platform (e.g. Ascend NPU images) and the
# only tilelang artifacts in this file are pass_configs that ... | import logging
import math
from functools import lru_cache
from typing import Optional
import torch
import triton
import triton.language as tl
logger = logging.getLogger(__name__)
# tilelang isn't shipped on every platform (e.g. Ascend NPU images) and the
# only tilelang artifacts in this file are pass_configs that ... | --- python/sglang/srt/layers/deepseek_v4_rope.py
+++ python/sglang/srt/layers/deepseek_v4_rope.py
@@ -143,12 +143,209 @@
tl.store(x_ptr + offs_x_imag, out_imag, mask=mask)
+@triton.jit
+def apply_rotary_emb_triton_kernel_batched(
+ x_ptr,
+ freqs_ptr,
+ positions_ptr,
+ rope_dim,
+ n_tokens,
+ ... | 197 | 0 | 2 | [
"deepseek"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/deepseek_v4_rope.py).
Goal: [AMD] Optimize o_proj gemm and attn output rope performance
Target hardware: unspecified
Target models: deepseek
Current file:
```python
import logging
import math
from functools import lru_cache
from typing import Opti... | ## Motivation
1. **The MLA output / q / kv projection w8a8-block FP8 GEMM uses the Triton kernel.** `apply_w8a8_block_fp8_linear` selects `triton_gemm_a8w8_blockscale` for the MLA projection shapes (they are in the hard-coded `use_aiter_triton_gemm_w8a8_tuned_gfx950` list). On gfx950 + HIP≥7.2 the CK bpreshuffle kerne... | [{"scope": "e2e", "hardware": "MI355X", "model": "DeepSeek-V4", "shape": null, "batch_size": 512, "metric": "throughput", "unit": "tok/s", "higher_is_better": true, "baseline": 48291, "optimized": 52560, "stated_speedup": null, "speedup": 1.0884}, {"scope": "e2e", "hardware": "MI355X", "model": "DeepSeek-V4", "shape": ... | null | 1.0692 |
sglang | c1067f88d6970562525d0385b80c7916aa8ea2a8 | 2026-06-19T05:16:52+07:00 | 27,793 | https://github.com/sgl-project/sglang/pull/27793 | [AMD][Perf] Tune extend attention block sizes for gfx950 (head_dim > 128) (#27793) | [AMD][Perf] Tune extend attention block sizes for gfx950 (head_dim > 128) (#27793) | 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
@@ -23,13 +23,14 @@
from sglang.srt.layers.attention.triton_ops.prefill_attention import (
context_attention_fwd,
)
-from sglang.srt.utils import is_cuda, is_hip
+from sglan... | 13 | 3 | 1 | [] | [
"GFX950"
] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/triton_ops/extend_attention.py).
Goal: [AMD][Perf] Tune extend attention block sizes for gfx950 (head_dim > 128)
Target hardware: GFX950
Target models: unspecified
Current file:
```python
# Copyright 2023-2024 SGLang Team
# Licensed unde... | ## Motivation
While profiling long-context prefill on MI350X (Qwen3.5-style GQA models, head_dim 256), the Triton extend kernel turned out to be the dominant cost: about 62% of GPU time at roughly 28% MFU, far below what the memory system allows.
The reason is that `_get_block_sizes_for_extend_attention` uses one har... | [{"scope": "kernel", "hardware": "MI350X", "model": null, "shape": "M=32768, prefix=67k", "batch\u6e38\u620f\u526f\u672c_size": null, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 137.2, "optimized": 87.5, "stated_speedup": null, "speedup": 1.568}, {"scope": "kernel", "hardware": "MI350X", "... | 1.5851 | 1.3178 |
sglang | c80d8fe78ae22411d660c58dd4024fbfa2edb8f2 | 2026-06-13T00:01:53+08:00 | 27,896 | https://github.com/sgl-project/sglang/pull/27896 | [Perf] Skip per-call mat_a/scales_a padding in cutlass FP8 blockwise GEMM (#27896) | [Perf] Skip per-call mat_a/scales_a padding in cutlass FP8 blockwise GEMM (#27896)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.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
@@ -565,6 +565,59 @@
return x_q, x_s
+def sglang_per_token_group_quant_fp8_row_padded(
+ x: torch.Tensor,
+ group_size: int,
+ eps: float = 1e-10,
+ row_alignment: int = 4,
+) -> Tuple[to... | 53 | 0 | 2 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/quantization/fp8_kernel.py).
Goal: [Perf] Skip per-call mat_a/scales_a padding in cutlass FP8 blockwise GEMM
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2024 SGLang Team
# Licensed under the Apache L... | ## Summary
The cutlass FP8 blockwise linear path (`cutlass_w8a8_block_fp8_linear_with_fallback`) quantizes the activation, then calls `fp8_blockwise_scaled_mm`, whose sgl-kernel wrapper pads `mat_a` and `scales_a` up to a multiple of 4 rows on every invocation. Each pad is a `torch.zeros` fill plus a `torch.cat`, so e... | [{"scope": "kernel", "hardware": "H20-3e", "model": "Qwen3.5-122B-A10B-FP8", "shape": null, "batch_size": null, "metric": "GPU time", "unit": "ms", "higher_is_better": false, "baseline": 69, "optimized": 0.4, "stated_speedup": null, "speedup": 172.5}, {"scope": "kernel", "hardware": "H20-3e", "model": "Qwen3.5-122B-A10... | 149.8214 | 1.0206 |
sglang | f57f8a8afd845ca5f03e229f41d891b727890fd3 | 2026-06-06T14:14:25+08:00 | 26,588 | https://github.com/sgl-project/sglang/pull/26588 | Optimize Gemma4 H200 MoE and extend attention (#26588) | Optimize Gemma4 H200 MoE and extend attention (#26588)
Co-authored-by: BBuf <xiaoyu.zhang@radixark.net> | 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
@@ -82,8 +82,10 @@
BLOCK_M, BLOCK_N = (16, 64)
elif _is_cuda and CUDA_CAPABILITY[0] >= 9:
# Hopper architecture (H100, etc.)
- if ... | 3 | 1 | 2 | [
"gemma"
] | [
"H200"
] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/triton_ops/extend_attention.py).
Goal: Optimize Gemma4 H200 MoE and extend attention
Target hardware: H200
Target models: gemma
Current file:
```python
# Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (t... | ## Summary
- Add H200 Triton fused MoE configs for Gemma4 `E=128,N=704` normal and down projections.
- Tune Hopper extend-attention block sizes for `Lq=129..256` to reduce TTFT/TPOT on Gemma4 prefill-heavy serving.
- Add a small-batch (M ≤ 256) per-head Triton kernel for Gemma4 QKV RMSNorm to lower MTP draft latency.
... | [{"scope": "e2e", "hardware": "NVIDIA H200", "model": "google/gemma-4-26B-A4B-it", "shape": null, "batch_size": 80, "metric": "output tok/s", "unit": "tok/s", "higher_is_better": true, "baseline": 1670.72, "optimized": 1725.5, "stated_speedup": 1.0328, "speedup": 1.0328}, {"scope": "e2e", "hardware": "NVIDIA H200", "mo... | null | 1.1267 |
sglang | f57f8a8afd845ca5f03e229f41d891b727890fd3 | 2026-06-06T14:14:25+08:00 | 26,588 | https://github.com/sgl-project/sglang/pull/26588 | Optimize Gemma4 H200 MoE and extend attention (#26588) | Optimize Gemma4 H200 MoE and extend attention (#26588)
Co-authored-by: BBuf <xiaoyu.zhang@radixark.net> | python/sglang/srt/layers/gemma4_fused_ops.py | triton | Apache-2.0 | """Fused triton kernels for Gemma4 decoder layer operations.
Fuses standard RMSNorm + residual-add (+ optional scalar multiply) into
a single kernel pass to reduce kernel launch overhead.
"""
from typing import Optional
import torch
import triton
import triton.language as tl
@triton.jit
def _gemma_rmsnorm_residual... | """Fused triton kernels for Gemma4 decoder layer operations.
Fuses standard RMSNorm + residual-add (+ optional scalar multiply) into
a single kernel pass to reduce kernel launch overhead.
"""
from typing import Optional
import torch
import triton
import triton.language as tl
@triton.jit
def _gemma_rmsnorm_residual... | --- python/sglang/srt/layers/gemma4_fused_ops.py
+++ python/sglang/srt/layers/gemma4_fused_ops.py
@@ -133,6 +133,29 @@
@triton.jit
+def _gemma_qkv_rmsnorm_store(
+ X_ptr,
+ W_ptr,
+ stride_m,
+ m,
+ h,
+ cols,
+ mask,
+ HEAD_DIM: tl.constexpr,
+ eps,
+ HAS_WEIGHT: tl.constexpr,
+):
+ ... | 107 | 35 | 2 | [
"gemma"
] | [
"H200"
] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/gemma4_fused_ops.py).
Goal: Optimize Gemma4 H200 MoE and extend attention
Target hardware: H200
Target models: gemma
Current file:
```python
"""Fused triton kernels for Gemma4 decoder layer operations.
Fuses standard RMSNorm + residual-add (+ opt... | ## Summary
- Add H200 Triton fused MoE configs for Gemma4 `E=128,N=704` normal and down projections.
- Tune Hopper extend-attention block sizes for `Lq=129..256` to reduce TTFT/TPOT on Gemma4 prefill-heavy serving.
- Add a small-batch (M ≤ 256) per-head Triton kernel for Gemma4 QKV RMSNorm to lower MTP draft latency.
... | [{"scope": "e2e", "hardware": "NVIDIA H200", "model": "google/gemma-4-26B-A4B-it", "shape": null, "batch_size": 80, "metric": "output tok/s", "unit": "tok/s", "higher_is_better": true, "baseline": 1670.72, "optimized": 1725.5, "stated_speedup": 1.0328, "speedup": 1.0328}, {"scope": "e2e", "hardware": "NVIDIA H200", "mo... | null | 1.1267 |
sglang | 38ae22e08c73ae549245eebac229f6888089cc35 | 2026-06-05T22:31:46-07:00 | 26,733 | https://github.com/sgl-project/sglang/pull/26733 | Nemotron perf changes (#26733) | Nemotron perf changes (#26733)
Co-authored-by: Brayden Zhong <brayden@radixark.ai> | 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
@@ -14,6 +14,7 @@
import triton.language as tl
from einops import rearrange
+from sglang.jit_kernel.utils import is_arch_support_pdl
from sglang.srt.server_args import get_global_server_args
... | 10 | 0 | 9 | [
"nemotron"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/fla/layernorm_gated.py).
Goal: Nemotron perf changes
Target hardware: unspecified
Target models: nemotron
Current file:
```python
# Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/modules/layernorm_gated.py
# C... | <!-- 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.io to discuss further. -->
## TODOs:
This PR aims to address the very obvious issues. T... | [{"scope": "e2e", "hardware": null, "model": "NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "shape": null, "batch_size": 1, "metric": "throughput", "unit": "TPS", "higher_is_better": true, "baseline": 127.67, "optimized": 130.61, "stated_speedup": null, "speedup": 1.023}, {"scope": "e2e", "hardware": null, "model": "NVIDIA-... | null | 1.0135 |
sglang | 38ae22e08c73ae549245eebac229f6888089cc35 | 2026-06-05T22:31:46-07:00 | 26,733 | https://github.com/sgl-project/sglang/pull/26733 | Nemotron perf changes (#26733) | Nemotron perf changes (#26733)
Co-authored-by: Brayden Zhong <brayden@radixark.ai> | python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py | triton | Apache-2.0 | import logging
from typing import 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.mamba.causal_conv1d_triton import PAD_SLOT_ID
from sglang.srt.layers.attention.mamba.mamba import MambaM... | import logging
from typing import 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.mamba.causal_conv1d_triton import PAD_SLOT_ID
from sglang.srt.layers.attention.mamba.mamba import MambaM... | --- python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
+++ python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
@@ -710,7 +710,7 @@
self,
mixer: MambaMixer2,
hidden_states: torch.Tensor,
- output: torch.Tensor,
+ output: Optional[torch.Tensor],
... | 4 | 2 | 9 | [
"nemotron"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py).
Goal: Nemotron perf changes
Target hardware: unspecified
Target models: nemotron
Current file:
```python
import logging
from typing import Optional, Union
import torch
import triton
import triton.language... | <!-- 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.io to discuss further. -->
## TODOs:
This PR aims to address the very obvious issues. T... | [{"scope": "e2e", "hardware": null, "model": "NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "shape": null, "batch_size": 1, "metric": "throughput", "unit": "TPS", "higher_is_better": true, "baseline": 127.67, "optimized": 130.61, "stated_speedup": null, "speedup": 1.023}, {"scope": "e2e", "hardware": null, "model": "NVIDIA-... | null | 1.0135 |
sglang | 38ae22e08c73ae549245eebac229f6888089cc35 | 2026-06-05T22:31:46-07:00 | 26,733 | https://github.com/sgl-project/sglang/pull/26733 | Nemotron perf changes (#26733) | Nemotron perf changes (#26733)
Co-authored-by: Brayden Zhong <brayden@radixark.ai> | python/sglang/srt/layers/attention/mamba/causal_conv1d_triton.py | triton | Apache-2.0 | # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# 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/... | # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# 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/... | --- python/sglang/srt/layers/attention/mamba/causal_conv1d_triton.py
+++ python/sglang/srt/layers/attention/mamba/causal_conv1d_triton.py
@@ -9,6 +9,8 @@
import torch
import triton
import triton.language as tl
+
+from sglang.jit_kernel.utils import is_arch_support_pdl
PAD_SLOT_ID = -1
@@ -629,8 +631,12 @@
... | 12 | 0 | 9 | [
"mamba",
"nemotron"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/mamba/causal_conv1d_triton.py).
Goal: Nemotron perf changes
Target hardware: unspecified
Target models: mamba, nemotron
Current file:
```python
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the... | <!-- 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.io to discuss further. -->
## TODOs:
This PR aims to address the very obvious issues. T... | [{"scope": "e2e", "hardware": null, "model": "NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "shape": null, "batch_size": 1, "metric": "throughput", "unit": "TPS", "higher_is_better": true, "baseline": 127.67, "optimized": 130.61, "stated_speedup": null, "speedup": 1.023}, {"scope": "e2e", "hardware": null, "model": "NVIDIA-... | null | 1.0135 |
sglang | 38ae22e08c73ae549245eebac229f6888089cc35 | 2026-06-05T22:31:46-07:00 | 26,733 | https://github.com/sgl-project/sglang/pull/26733 | Nemotron perf changes (#26733) | Nemotron perf changes (#26733)
Co-authored-by: Brayden Zhong <brayden@radixark.ai> | python/sglang/srt/layers/attention/mamba/ops/mamba_ssm.py | triton | Apache-2.0 | # Adapted from: https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/mamba/ops/mamba_ssm.py
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# Copyright (c) 2024, Tri Dao, Albert Gu.
# Adapted from https://github.com/state-spaces/mamba/blo... | # Adapted from: https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/mamba/ops/mamba_ssm.py
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# Copyright (c) 2024, Tri Dao, Albert Gu.
# Adapted from https://github.com/state-spaces/mamba/blo... | --- python/sglang/srt/layers/attention/mamba/ops/mamba_ssm.py
+++ python/sglang/srt/layers/attention/mamba/ops/mamba_ssm.py
@@ -10,6 +10,8 @@
import triton
import triton.language as tl
from packaging import version
+
+from sglang.jit_kernel.utils import is_arch_support_pdl
PAD_SLOT_ID = -1
@@ -141,7 +143,11 @@
... | 12 | 0 | 9 | [
"mamba",
"nemotron"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/mamba/ops/mamba_ssm.py).
Goal: Nemotron perf changes
Target hardware: unspecified
Target models: mamba, nemotron
Current file:
```python
# Adapted from: https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/mamba/ops/... | <!-- 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.io to discuss further. -->
## TODOs:
This PR aims to address the very obvious issues. T... | [{"scope": "e2e", "hardware": null, "model": "NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "shape": null, "batch_size": 1, "metric": "throughput", "unit": "TPS", "higher_is_better": true, "baseline": 127.67, "optimized": 130.61, "stated_speedup": null, "speedup": 1.023}, {"scope": "e2e", "hardware": null, "model": "NVIDIA-... | null | 1.0135 |
sglang | bd47869ba4121ca18cf60acc946985a68bf75eb4 | 2026-06-04T22:11:43-07:00 | 27,320 | https://github.com/sgl-project/sglang/pull/27320 | [perf] parallelize create_flashmla_kv_indices over page-blocks (#27320) | [perf] parallelize create_flashmla_kv_indices over page-blocks (#27320) | python/sglang/srt/layers/attention/triton_ops/kv_indices.py | triton | Apache-2.0 | import triton
import triton.language as tl
_FLASHMLA_CREATE_KV_BLOCK_SIZE = 4096
FLASHMLA_CREATE_KV_BLOCK_SIZE_TRITON = tl.constexpr(_FLASHMLA_CREATE_KV_BLOCK_SIZE)
@triton.jit
def create_flashinfer_kv_indices_triton(
req_to_token_ptr, # [max_batch, max_context_len]
req_pool_indices_ptr,
page_kernel_len... | import triton
import triton.language as tl
_FLASHMLA_CREATE_KV_BLOCK_SIZE = 4096
FLASHMLA_CREATE_KV_BLOCK_SIZE_TRITON = tl.constexpr(_FLASHMLA_CREATE_KV_BLOCK_SIZE)
@triton.jit
def create_flashinfer_kv_indices_triton(
req_to_token_ptr, # [max_batch, max_context_len]
req_pool_indices_ptr,
page_kernel_len... | --- python/sglang/srt/layers/attention/triton_ops/kv_indices.py
+++ python/sglang/srt/layers/attention/triton_ops/kv_indices.py
@@ -49,6 +49,15 @@
return num_page_per_block
+def get_num_kv_index_blocks_flashmla(kv_indices_width: int, page_size: int) -> int:
+ """Grid axis-1 size for create_flashmla_kv_indic... | 13 | 1 | 6 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/triton_ops/kv_indices.py).
Goal: [perf] parallelize create_flashmla_kv_indices over page-blocks
Target hardware: unspecified
Target models: unspecified
Current file:
```python
import triton
import triton.language as tl
_FLASHMLA_CREATE_... | Launch one CTA per page-block (grid axis 1) instead of a single CTA looping over every block, so building the kv-index page table scales with context length instead of serializing it (a large win at long context, where bs=1 left one CTA to walk the whole table).
Add get_num_kv_index_blocks_flashmla to size grid axis 1... | [{"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 15, "optimized": 1.5, "stated_speedup": null, "speedup": 10.0}] | 10 | null |
sglang | bd47869ba4121ca18cf60acc946985a68bf75eb4 | 2026-06-04T22:11:43-07:00 | 27,320 | https://github.com/sgl-project/sglang/pull/27320 | [perf] parallelize create_flashmla_kv_indices over page-blocks (#27320) | [perf] parallelize create_flashmla_kv_indices over page-blocks (#27320) | 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 logging
import math
from dataclasses import dataclass
from typing import TYPE_CHECKING, Optional, Union
import torch
import triton
import triton.language as tl
from sglang.jit_kernel.fixup_zero_kv imp... | from __future__ import annotations
"""
Support attention backend for TRTLLM MLA kernels from flashinfer.
"""
import logging
import math
from dataclasses import dataclass
from typing import TYPE_CHECKING, Optional, Union
import torch
import triton
import triton.language as tl
from sglang.jit_kernel.fixup_zero_kv imp... | --- python/sglang/srt/layers/attention/trtllm_mla_backend.py
+++ python/sglang/srt/layers/attention/trtllm_mla_backend.py
@@ -23,6 +23,7 @@
from sglang.srt.layers.attention.utils import (
concat_mla_absorb_q_general,
create_flashmla_kv_indices_triton,
+ get_num_kv_index_blocks_flashmla,
get_num_page_... | 15 | 2 | 6 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/trtllm_mla_backend.py).
Goal: [perf] parallelize create_flashmla_kv_indices over page-blocks
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from __future__ import annotations
"""
Support attention backen... | Launch one CTA per page-block (grid axis 1) instead of a single CTA looping over every block, so building the kv-index page table scales with context length instead of serializing it (a large win at long context, where bs=1 left one CTA to walk the whole table).
Add get_num_kv_index_blocks_flashmla to size grid axis 1... | [{"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 15, "optimized": 1.5, "stated_speedup": null, "speedup": 10.0}] | 10 | null |
sglang | 3ea1ba5b1531daaaedd6ae55bb32a2a7fe0fd7b6 | 2026-06-02T16:48:31+08:00 | 26,206 | https://github.com/sgl-project/sglang/pull/26206 | [GDN] Optimize prefill QKV split dispatch (#26206) | [GDN] Optimize prefill QKV split dispatch (#26206)
Co-authored-by: BBuf <xiaoyu.zhang@radixark.net> | python/sglang/srt/layers/attention/fla/chunk_delta_h.py | triton | Apache-2.0 | # Adapted from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/common/chunk_delta_h.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.fla.index im... | # Adapted from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/common/chunk_delta_h.py
# -*- coding: utf-8 -*-
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
import os
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/chunk_delta_h.py
+++ python/sglang/srt/layers/attention/fla/chunk_delta_h.py
@@ -2,6 +2,7 @@
# -*- coding: utf-8 -*-
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+import os
from typing import Optional, Tuple
import torch
@@ -20,6 +21,9 @@
NUM_WARPS = [2, 4] if ... | 14 | 2 | 2 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/fla/chunk_delta_h.py).
Goal: [GDN] Optimize prefill QKV split dispatch
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Adapted from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/com... | ## Summary
This patch reduces Qwen3.6 GDN prefill QKV split overhead observed during profiling:
- add a Triton `fused_qkv_split_gdn_prefill_kernel` to split post-conv GDN Q/K/V into contiguous FLA prefill tensors in one launch;
- route CUDA GDN prefill tensors, including strided post-conv views, through that fused sp... | [{"scope": "e2e", "hardware": "B200", "model": "Qwen/Qwen3.6-35B-A3B-FP8", "shape": null, "batch_size": null, "metric": "throughput", "unit": "tok/s", "higher_is_better": true, "baseline": 7205.38, "optimized": 7396.73, "stated_speedup": null, "speedup": 1.0266}, {"scope": "e2e", "hardware": "B200", "model": "Qwen/Qwen... | 1.8029 | 1.0262 |
sglang | 5ae8d286d2b851a188aa753b46efc45e5322949e | 2026-06-02T01:00:17-07:00 | 26,502 | https://github.com/sgl-project/sglang/pull/26502 | perf(gemma4): single-launch fused router (topk + softmax + scale) (#26502) | perf(gemma4): single-launch fused router (topk + softmax + scale) (#26502) | python/sglang/srt/layers/gemma4_fused_ops.py | triton | Apache-2.0 | """Fused triton kernels for Gemma4 decoder layer operations.
Fuses standard RMSNorm + residual-add (+ optional scalar multiply) into
a single kernel pass to reduce kernel launch overhead.
"""
from typing import Optional
import torch
import triton
import triton.language as tl
@triton.jit
def _gemma_rmsnorm_residual... | """Fused triton kernels for Gemma4 decoder layer operations.
Fuses standard RMSNorm + residual-add (+ optional scalar multiply) into
a single kernel pass to reduce kernel launch overhead.
"""
from typing import Optional
import torch
import triton
import triton.language as tl
@triton.jit
def _gemma_rmsnorm_residual... | --- python/sglang/srt/layers/gemma4_fused_ops.py
+++ python/sglang/srt/layers/gemma4_fused_ops.py
@@ -283,3 +283,117 @@
BLOCK_SIZE=BLOCK_SIZE,
)
return out
+
+
+@triton.jit
+def _gemma4_routing_kernel(
+ gating_ptr, # [T, E] router logits, any float dtype
+ per_expert_scale_ptr, # [E] per-exp... | 114 | 0 | 1 | [
"gemma"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/gemma4_fused_ops.py).
Goal: perf(gemma4): single-launch fused router (topk + softmax + scale)
Target hardware: unspecified
Target models: gemma
Current file:
```python
"""Fused triton kernels for Gemma4 decoder layer operations.
Fuses standard RM... | ## Motivation
`Gemma4MoE.routing_function` emits four per-layer GPU kernels for every MoE
forward pass (one decode step touches ~30 routed-MoE layers):
| Kernel | Decode share (Gemma-4-26B-A4B-IT + MTP, B200) |
|---|---|
| `at::native::sbtopk::gatherTopK<bf16,uint,2,false>` (torch.topk) | 2.4% |
| `at::native::bitoni... | [{"scope": "e2e", "hardware": "B200", "model": "Gemma-4-26B-A4B-IT", "shape": null, "batch_size": null, "metric": "throughput", "unit": "tok/s", "higher_is_better": true, "baseline": 2729.3, "optimized": 2880.94, "stated_speedup": null, "speedup": 1.0556}, {"scope": "e2e", "hardware": "B200", "model": "Gemma-4-26B-A4B-... | null | 1.0447 |
sglang | deaba74745d7f98485211ce8db406ecefb63512d | 2026-05-28T06:23:35+08:00 | 26,383 | https://github.com/sgl-project/sglang/pull/26383 | [AMD][DSV4] DSV4 MTP graph + sparse triton attn optimizations (#26383) | [AMD][DSV4] DSV4 MTP graph + sparse triton attn optimizations (#26383)
Co-authored-by: wunhuang <wunhuang@amd.com>
Co-authored-by: Thomas Wang <1am9trash@gmail.com>
Co-authored-by: Xinyi Song <86638975+RolaoDenthu@users.noreply.github.com>
Co-authored-by: HaiShaw <hixiao@gmail.com>
Co-authored-by: amd-danli103 <danli1... | python/sglang/srt/layers/attention/nsa/triton_decode/triton_mla_kernels_decode_fused.py | triton | Apache-2.0 | """
Fused Gather+Dequant+Attention Kernel for DSV4 (d_qk=512)
This module implements a fused kernel that combines:
1. Gather: Load KV from sparse indices
2. Dequant: FP8 to BF16 dequantization
3. Attention: Compute attention scores and output
Benefits for workloads without extra scope:
- Eliminates intermediate buffe... | """
Fused Gather+Dequant+Attention Kernel for DSV4 (d_qk=512)
This module implements a fused kernel that combines:
1. Gather: Load KV from sparse indices
2. Dequant: FP8 to BF16 dequantization
3. Attention: Compute attention scores and output
Benefits for workloads without extra scope:
- Eliminates intermediate buffe... | --- python/sglang/srt/layers/attention/nsa/triton_decode/triton_mla_kernels_decode_fused.py
+++ python/sglang/srt/layers/attention/nsa/triton_decode/triton_mla_kernels_decode_fused.py
@@ -1208,6 +1208,9 @@
triton.Config({"BLOCK_H": 64, "BLOCK_N": 128}, num_warps=4, num_stages=1),
triton.Config({"BLOCK... | 12 | 1 | 4 | [
"deepseek"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/nsa/triton_decode/triton_mla_kernels_decode_fused.py).
Goal: [AMD][DSV4] DSV4 MTP graph + sparse triton attn optimizations
Target hardware: unspecified
Target models: deepseek
Current file:
```python
"""
Fused Gather+Dequant+Attention Ke... | Co-authored-by: @1am9trash @RolaoDenthu @yichiche @hubertlu-tw @sogalin @HaiShaw @yctseng0211 @bingxche @amd-danli103 @Raiden-Makoto
## Motivation
The long-standing draft #25552 collected 85 commits of AMD / DSV4 work over several rebases. Most of the optimization PRs it tracked (compressor opts, fused topk, sparse M... | [] | null | null |
sglang | e27d4fb70f33d6a86533a5dc3e14114c480b4d2f | 2026-05-25T16:08:21+08:00 | 25,775 | https://github.com/sgl-project/sglang/pull/25775 | [Perf][Qwen3.5] Add case 512 to topkGatingSoftmaxKernelLauncher, (#25775) | [Perf][Qwen3.5] Add case 512 to topkGatingSoftmaxKernelLauncher, (#25775) | 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
@@ -696,6 +696,9 @@
case 256:
LAUNCH_SOFTMAX(T, 256, WARPS_PER_TB);
break;
+ case 512:
+ LAUNCH_SOFTMAX(T, 512, WARPS_PER_TB);
+ break;
default: {
TORCH_CHECK(
s... | 4 | 1 | 1 | [
"qwen"
] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/moe/moe_topk_softmax_kernels.cu).
Goal: [Perf][Qwen3.5] Add case 512 to topkGatingSoftmaxKernelLauncher,
Target hardware: unspecified
Target models: qwen
Current file:
```cuda
// Adapt from https://github.com/vllm-project/vllm/blob/v0.7.3/csrc/moe/topk_softm... | ## Motivation
<!-- Describe the purpose and goals of this pull request. -->
## Modifications
Add case 512 to topkGatingSoftmaxKernelLauncher, enabling the fused single-kernel path for models with 512 experts (e.g. Qwen3.5-397B-A17B).
## Accuracy Tests
<!-- If this pull request affects model outputs (e.g., changes t... | [{"scope": "kernel", "hardware": "H200", "model": "Qwen3.5-397B-A17B", "shape": "num_experts=512,topk=10,dtype=bf16", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 18.63, "optimized": 10.39, "stated_speedup": 1.79, "speedup": 1.7931}, {"scope": "kernel", "hardware": "H200", ... | 1.8183 | 1.0144 |
sglang | ec6fcb93cb8b100e26db2bcf6a43613e9b05c7d1 | 2026-05-24T21:36:16-07:00 | 26,241 | https://github.com/sgl-project/sglang/pull/26241 | [perf][spec decoding] Skip common_template in TRTLLMMLAMultiStepDraftBackend init (#26241) | [perf][spec decoding] Skip common_template in TRTLLMMLAMultiStepDraftBackend init (#26241) | 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 logging
import math
from dataclasses import dataclass
from typing import TYPE_CHECKING, Optional, Union
import torch
import triton
import triton.language as tl
from sglang.jit_kernel.fixup_zero_kv imp... | from __future__ import annotations
"""
Support attention backend for TRTLLM MLA kernels from flashinfer.
"""
import logging
import math
from dataclasses import dataclass
from typing import TYPE_CHECKING, Optional, Union
import torch
import triton
import triton.language as tl
from sglang.jit_kernel.fixup_zero_kv imp... | --- python/sglang/srt/layers/attention/trtllm_mla_backend.py
+++ python/sglang/srt/layers/attention/trtllm_mla_backend.py
@@ -1235,3 +1235,22 @@
kv_indptr_buf=self.kv_indptr[i],
q_indptr_decode_buf=self.q_indptr_decode,
)
+
+ def init_forward_metadata(self, forward_batc... | 19 | 0 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/trtllm_mla_backend.py).
Goal: [perf][spec decoding] Skip common_template in TRTLLMMLAMultiStepDraftBackend init
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from __future__ import annotations
"""
Suppo... | `FlashInferMLAMultiStepDraftBackend.{init_forward_metadata, init_forward_metadata_replay_cuda_graph}` go through `common_template`, which launches the `generate_draft_decode_kv_indices` triton kernel and slices `spec_info.kv_indptr/kv_indices` for each draft step. Both are unnecessary for the trtllm_mla path used by EA... | [{"scope": "e2e", "hardware": "Kimi-K2.5-NVFP4", "model": null, "shape": "TP=4, 80K ctx, EAGLE3 3-step, topk=1", "batch_size": null, "metric": "Mean TPOT", "unit": "ms", "higher_is_better": false, "baseline": 2.47, "optimized": 2.41, "stated_speedup": null, "speedup": 1.0249}, {"scope": "e2e", "hardware": "Kimi-K2.5-NV... | null | 1.025 |
sglang | 821d5f4a5b2550bd6a9fa60fded40568d2c7344c | 2026-05-25T10:28:52+08:00 | 25,874 | https://github.com/sgl-project/sglang/pull/25874 | [CPU] add faster KV-cache writes (#25874) | [CPU] add faster KV-cache writes (#25874) | sgl-kernel/csrc/cpu/common.h | cpp | Apache-2.0 | #pragma once
#include <ATen/ATen.h>
#include <ATen/Parallel.h>
#if defined(_OPENMP)
#include <omp.h>
#endif
namespace {
// dispatch bool
#define AT_DISPATCH_BOOL(BOOL_V, BOOL_NAME, ...) \
[&] { \
if (BOOL_V) { \
constexpr bool BOOL_... | #pragma once
#include <ATen/ATen.h>
#include <ATen/Dispatch.h>
#include <ATen/Parallel.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
@@ -1,6 +1,7 @@
#pragma once
#include <ATen/ATen.h>
+#include <ATen/Dispatch.h>
#include <ATen/Parallel.h>
#if defined(_OPENMP)
@@ -43,6 +44,14 @@
} \
} ... | 9 | 0 | 2 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/common.h).
Goal: [CPU] add faster KV-cache writes
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#pragma once
#include <ATen/ATen.h>
#include <ATen/Parallel.h>
#if defined(_OPENMP)
#include <omp.h>
#endif
namespace {
// dispatch ... | ## Motivation
Faster KV-cache writes on CPU: replace slow `index_put_` scatter for MHA with a dedicated row-copy kernel, matching the CUDA `store_cache` idea.
## Modifications
- **sgl-kernel**: Add `store_cache_cpu` (OpenMP parallel row copy); 3D tensors `[B, H, D]`; `int32`/`int64` indices; dtypes `fp16`, `bf16`, `... | [] | null | null |
sglang | 821d5f4a5b2550bd6a9fa60fded40568d2c7344c | 2026-05-25T10:28:52+08:00 | 25,874 | https://github.com/sgl-project/sglang/pull/25874 | [CPU] add faster KV-cache writes (#25874) | [CPU] add faster KV-cache writes (#25874) | 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
@@ -410,6 +410,15 @@
bool disable_grouping,
at::ScalarType out_dtype);
+// kvcache
+void store_cache_cpu(
+ const at::Tensor& k,
+ const at::Tensor& v,
+ const at::Tensor& k_cache,
+ const at::Tensor&... | 15 | 0 | 2 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/torch_extension_cpu.cpp).
Goal: [CPU] add faster KV-cache writes
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you ... | ## Motivation
Faster KV-cache writes on CPU: replace slow `index_put_` scatter for MHA with a dedicated row-copy kernel, matching the CUDA `store_cache` idea.
## Modifications
- **sgl-kernel**: Add `store_cache_cpu` (OpenMP parallel row copy); 3D tensors `[B, H, D]`; `int32`/`int64` indices; dtypes `fp16`, `bf16`, `... | [] | null | null |
sglang | ff70aeac3051b5f85559f382d0622716831895c8 | 2026-05-14T00:57:46+08:00 | 24,491 | https://github.com/sgl-project/sglang/pull/24491 | [diffusion] feat: add performance mode server args (#24491) | [diffusion] feat: add performance mode server args (#24491) | python/sglang/srt/layers/attention/triton_ops/rocm_mla_decode_rope.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/rocm_mla_decode_rope.py
+++ python/sglang/srt/layers/attention/triton_ops/rocm_mla_decode_rope.py
@@ -395,7 +395,7 @@
IS_NEOX_STYLE=is_neox_style,
num_warps=4,
num_stages=num_stages,
- **extra_kargs
+ **extra_kargs,
)
| 1 | 1 | 2 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/triton_ops/rocm_mla_decode_rope.py).
Goal: [diffusion] feat: add performance mode server args
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2023-2024 SGLang Team
# Licensed under the Apache L... | ## Summary
- Add `--performance-mode` / `--mode` for diffusion server defaults: `auto`, `throughput`, `memory`, and `balanced`, with `aggressive`, `conservative`, and `balance` aliases.
- Auto-select FSDP+CFG only for high-confidence multi-GPU Qwen/Wan CFG cases, gated by the least available memory across selected GPU... | [] | null | null |
sglang | bfc2eda42d478affdce6dbc642f5b14cd17164fb | 2026-05-12T08:55:27+08:00 | 23,633 | https://github.com/sgl-project/sglang/pull/23633 | [MUSA] Use MUSA-optimized operators in piecewise CUDA graph (#23633) | [MUSA] Use MUSA-optimized operators in piecewise CUDA graph (#23633)
Signed-off-by: popsiclexu <zhenxuexu@gmail.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
@@ -93,6 +93,24 @@
except ImportError:
# Fallback: vllm not available, will use native PyTorch implementation
_has_vllm = False
+
+if _is_musa:
+
+ @register_fake_if_exis... | 18 | 0 | 4 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/quantization/fp8_kernel.py).
Goal: [MUSA] Use MUSA-optimized operators in piecewise CUDA graph
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Copyright 2024 SGLang Team
# Licensed under the Apache License, Versio... | <!-- 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.io to discuss further. -->
## Motivation
MUSA devices previously could not use piecewis... | [] | null | null |
sglang | d3fd91ed9726cc87d3f2e67e758202d3982cde32 | 2026-05-10T15:24:12+08:00 | 24,696 | https://github.com/sgl-project/sglang/pull/24696 | [Gemma4] Optimize Gemm4 with fused Q/K/V RMSNorm + per-expert FP8 ckpt loader (#24696) | [Gemma4] Optimize Gemm4 with fused Q/K/V RMSNorm + per-expert FP8 ckpt loader (#24696)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com> | python/sglang/srt/layers/gemma4_fused_ops.py | triton | Apache-2.0 | """Fused triton kernels for Gemma4 decoder layer operations.
Fuses standard RMSNorm + residual-add (+ optional scalar multiply) into
a single kernel pass to reduce kernel launch overhead.
"""
import torch
import triton
import triton.language as tl
@triton.jit
def _gemma_rmsnorm_residual_kernel(
X_ptr,
W_ptr... | """Fused triton kernels for Gemma4 decoder layer operations.
Fuses standard RMSNorm + residual-add (+ optional scalar multiply) into
a single kernel pass to reduce kernel launch overhead.
"""
from typing import Optional
import torch
import triton
import triton.language as tl
@triton.jit
def _gemma_rmsnorm_residual... | --- python/sglang/srt/layers/gemma4_fused_ops.py
+++ python/sglang/srt/layers/gemma4_fused_ops.py
@@ -3,6 +3,8 @@
Fuses standard RMSNorm + residual-add (+ optional scalar multiply) into
a single kernel pass to reduce kernel launch overhead.
"""
+
+from typing import Optional
import torch
import triton
@@ -130,6 ... | 115 | 0 | 1 | [
"gemma"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/gemma4_fused_ops.py).
Goal: [Gemma4] Optimize Gemm4 with fused Q/K/V RMSNorm + per-expert FP8 ckpt loader
Target hardware: unspecified
Target models: gemma
Current file:
```python
"""Fused triton kernels for Gemma4 decoder layer operations.
Fuses... | ## Motivation
The original motivation is to speedup Gemm4 model. Two unrelated issues on the Gemma4 FP8-MoE path:
1. **Attention RMSNorm is launch-bound.** Each Gemma4 attention layer issues three separate sgl-kernel `rmsnorm` launches (q/k/v). At small token counts the per-launch host overhead dominates — three back... | [{"scope": "kernel", "hardware": "H100", "model": null, "shape": "Q=8h,KV=2h,head_dim=128", "batch_size": 1, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 44.4, "optimized": 19.8, "stated_speedup": 2.25, "speedup": 2.2424}, {"scope": "kernel", "hardware": "H100", "model": null, "shape": "Q=8... | 2.9189 | 1.0077 |
sglang | a61a14f416c4003809a82112b4a591eec38a0a10 | 2026-05-09T08:52:51+08:00 | 24,271 | https://github.com/sgl-project/sglang/pull/24271 | [KDA] Optimize prefill kernels with diagonal and recompute fuse (#24271) | [KDA] Optimize prefill kernels with diagonal and recompute fuse (#24271)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com> | python/sglang/srt/layers/attention/fla/chunk_delta_h.py | triton | Apache-2.0 | # Adapted from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/common/chunk_delta_h.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.fla.index im... | # Adapted from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/common/chunk_delta_h.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.fla.index im... | --- python/sglang/srt/layers/attention/fla/chunk_delta_h.py
+++ python/sglang/srt/layers/attention/fla/chunk_delta_h.py
@@ -13,22 +13,30 @@
prepare_chunk_offsets,
)
from sglang.srt.layers.attention.fla.op import exp, safe_exp
-from sglang.srt.layers.attention.fla.utils import is_nvidia_hopper
+from sglang.srt.la... | 21 | 14 | 4 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/fla/chunk_delta_h.py).
Goal: [KDA] Optimize prefill kernels with diagonal and recompute fuse
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Adapted from https://github.com/fla-org/flash-linear-attention... | <!-- 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.io to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [{"scope": "e2e", "hardware": null, "model": "Kimi-Linear-48B-A3B-Instruct", "shape": "B=1,T=8192", "batch_size": 1, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 5.0, "optimized": 3.3, "stated_speedup": null, "speedup": 1.5152}, {"scope": "e2e", "hardware": null, "model": "Kimi-Linear-48B-A... | 1.3088 | 1.1744 |
sglang | a61a14f416c4003809a82112b4a591eec38a0a10 | 2026-05-09T08:52:51+08:00 | 24,271 | https://github.com/sgl-project/sglang/pull/24271 | [KDA] Optimize prefill kernels with diagonal and recompute fuse (#24271) | [KDA] Optimize prefill kernels with diagonal and recompute fuse (#24271)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com> | python/sglang/srt/layers/attention/fla/chunk_intra.py | triton | Apache-2.0 | # Adapted from flash-linear-attention project.
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.fla.chunk_intra_token_parallel import (
chunk_kda_fwd_intra_token_parallel,
)
from sglang.srt.layers.attention.fla.index import ... | # Adapted from flash-linear-attention project.
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.fla.chunk_intra_token_parallel import (
chunk_kda_fwd_intra_token_parallel,
)
from sglang.srt.layers.attention.fla.index import ... | --- python/sglang/srt/layers/attention/fla/chunk_intra.py
+++ python/sglang/srt/layers/attention/fla/chunk_intra.py
@@ -11,7 +11,7 @@
from sglang.srt.layers.attention.fla.index import (
prepare_chunk_indices,
)
-from sglang.srt.layers.attention.fla.op import exp2, gather
+from sglang.srt.layers.attention.fla.op ... | 495 | 104 | 4 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/fla/chunk_intra.py).
Goal: [KDA] Optimize prefill kernels with diagonal and recompute fuse
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Adapted from flash-linear-attention project.
# Copyright (c) 202... | <!-- 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.io to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [{"scope": "e2e", "hardware": null, "model": "Kimi-Linear-48B-A3B-Instruct", "shape": "B=1,T=8192", "batch_size": 1, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 5.0, "optimized": 3.3, "stated_speedup": null, "speedup": 1.5152}, {"scope": "e2e", "hardware": null, "model": "Kimi-Linear-48B-A... | 1.3088 | 1.1744 |
sglang | a61a14f416c4003809a82112b4a591eec38a0a10 | 2026-05-09T08:52:51+08:00 | 24,271 | https://github.com/sgl-project/sglang/pull/24271 | [KDA] Optimize prefill kernels with diagonal and recompute fuse (#24271) | [KDA] Optimize prefill kernels with diagonal and recompute fuse (#24271)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com> | python/sglang/srt/layers/attention/fla/cumsum.py | triton | Apache-2.0 | # Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/utils/cumsum.py
# -*- coding: utf-8 -*-
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
from typing import Optional
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.fla.index import prepare_chun... | # Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/utils/cumsum.py
# -*- coding: utf-8 -*-
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
from typing import Optional
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.fla.index import prepare_chun... | --- python/sglang/srt/layers/attention/fla/cumsum.py
+++ python/sglang/srt/layers/attention/fla/cumsum.py
@@ -70,9 +70,10 @@
@triton.autotune(
configs=[
- triton.Config({"BS": BS}, num_warps=num_warps)
+ triton.Config({"BS": BS}, num_warps=num_warps, num_stages=num_stages)
for BS in BS_LI... | 2 | 1 | 4 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/fla/cumsum.py).
Goal: [KDA] Optimize prefill kernels with diagonal and recompute fuse
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Adapt from https://github.com/fla-org/flash-linear-attention/blob/mai... | <!-- 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.io to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [{"scope": "e2e", "hardware": null, "model": "Kimi-Linear-48B-A3B-Instruct", "shape": "B=1,T=8192", "batch_size": 1, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 5.0, "optimized": 3.3, "stated_speedup": null, "speedup": 1.5152}, {"scope": "e2e", "hardware": null, "model": "Kimi-Linear-48B-A... | 1.3088 | 1.1744 |
sglang | a61a14f416c4003809a82112b4a591eec38a0a10 | 2026-05-09T08:52:51+08:00 | 24,271 | https://github.com/sgl-project/sglang/pull/24271 | [KDA] Optimize prefill kernels with diagonal and recompute fuse (#24271) | [KDA] Optimize prefill kernels with diagonal and recompute fuse (#24271)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com> | python/sglang/srt/layers/attention/fla/kda.py | triton | Apache-2.0 | # Adapted from https://github.com/vllm-project/vllm/blob/0384aa7150c4c9778efca041ffd1beb3ad2bd694/vllm/model_executor/layers/fla/ops/kda.py
# This file contains code copied from the flash-linear-attention project.
# The original source code was licensed under the MIT license and included
# the following copyright notic... | # Adapted from https://github.com/vllm-project/vllm/blob/0384aa7150c4c9778efca041ffd1beb3ad2bd694/vllm/model_executor/layers/fla/ops/kda.py
# This file contains code copied from the flash-linear-attention project.
# The original source code was licensed under the MIT license and included
# the following copyright notic... | --- python/sglang/srt/layers/attention/fla/kda.py
+++ python/sglang/srt/layers/attention/fla/kda.py
@@ -17,10 +17,14 @@
from sglang.srt.layers.attention.fla.fused_recurrent import (
fused_recurrent_gated_delta_rule_fwd_kernel,
)
-from sglang.srt.layers.attention.fla.index import prepare_chunk_indices
+from sglan... | 36 | 14 | 4 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/fla/kda.py).
Goal: [KDA] Optimize prefill kernels with diagonal and recompute fuse
Target hardware: unspecified
Target models: unspecified
Current file:
```python
# Adapted from https://github.com/vllm-project/vllm/blob/0384aa7150c4c9778... | <!-- 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.io to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [{"scope": "e2e", "hardware": null, "model": "Kimi-Linear-48B-A3B-Instruct", "shape": "B=1,T=8192", "batch_size": 1, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 5.0, "optimized": 3.3, "stated_speedup": null, "speedup": 1.5152}, {"scope": "e2e", "hardware": null, "model": "Kimi-Linear-48B-A... | 1.3088 | 1.1744 |
sglang | 15e6572f21980e906c568fa82f9677edec601eaa | 2026-05-08T11:38:33+08:00 | 23,255 | https://github.com/sgl-project/sglang/pull/23255 | [MUSA][18/N] Add MUSA-optimized kernel implementations for hot ops (#23255) | [MUSA][18/N] Add MUSA-optimized kernel implementations for hot ops (#23255)
Signed-off-by: Joey-gvwal <joey_gvwal@yeah.net>
Co-authored-by: R0CKSTAR <yeahdongcn@gmail.com> | sgl-kernel/csrc/common_extension_musa.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_musa.cc
+++ sgl-kernel/csrc/common_extension_musa.cc
@@ -72,6 +72,12 @@
m.def("concat_mla_k(Tensor! k, Tensor k_nope, Tensor k_rope) -> ()");
m.impl("concat_mla_k", torch::kMUSA, &concat_mla_k);
+ m.def(
+ "rotary_embedding(Tensor positions, Tensor! query,"
+ " ... | 42 | 0 | 4 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/common_extension_musa.cc).
Goal: [MUSA][18/N] Add MUSA-optimized kernel implementations for hot ops
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache Lic... | <!-- 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.io to discuss further. -->
## Motivation
This PR continues the ongoing effort (tracked... | [] | null | null |
sglang | 15e6572f21980e906c568fa82f9677edec601eaa | 2026-05-08T11:38:33+08:00 | 23,255 | https://github.com/sgl-project/sglang/pull/23255 | [MUSA][18/N] Add MUSA-optimized kernel implementations for hot ops (#23255) | [MUSA][18/N] Add MUSA-optimized kernel implementations for hot ops (#23255)
Signed-off-by: Joey-gvwal <joey_gvwal@yeah.net>
Co-authored-by: R0CKSTAR <yeahdongcn@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
@@ -26,6 +26,10 @@
#include <vector>
#include "scalar_type.hpp"
+
+#ifdef USE_MUSA
+#include "sgl_kernel_musa_ops.h"
+#endif
#define _CONCAT(A, B) A##B
#define CONCAT(A, B) _CONCAT(A, B) | 4 | 0 | 4 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/include/sgl_kernel_ops.h).
Goal: [MUSA][18/N] Add MUSA-optimized kernel implementations for hot ops
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* 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.io to discuss further. -->
## Motivation
This PR continues the ongoing effort (tracked... | [] | null | null |
sglang | e5c58eb9d627421236f78053137562b97c14984e | 2026-05-05T00:36:26+08:00 | 24,048 | https://github.com/sgl-project/sglang/pull/24048 | [VLM] Optimize Gemma4 VLM with PCG and fuse RMSNorm + residual add + scalar (#24048) | [VLM] Optimize Gemma4 VLM with PCG and fuse RMSNorm + residual add + scalar (#24048)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com> | python/sglang/srt/layers/gemma4_fused_ops.py | triton | Apache-2.0 | """Fused triton kernels for Gemma4 decoder layer operations.
Fuses standard RMSNorm + residual-add (+ optional scalar multiply) into
a single kernel pass to reduce kernel launch overhead.
"""
import torch
import triton
import triton.language as tl
@triton.jit
def _gemma_rmsnorm_residual_kernel(
X_ptr,
W_ptr... | """Fused triton kernels for Gemma4 decoder layer operations.
Fuses standard RMSNorm + residual-add (+ optional scalar multiply) into
a single kernel pass to reduce kernel launch overhead.
"""
import torch
import triton
import triton.language as tl
@triton.jit
def _gemma_rmsnorm_residual_kernel(
X_ptr,
W_ptr... | --- python/sglang/srt/layers/gemma4_fused_ops.py
+++ python/sglang/srt/layers/gemma4_fused_ops.py
@@ -77,3 +77,94 @@
BLOCK_SIZE=BLOCK_SIZE,
)
return out
+
+
+@triton.jit
+def _gemma_dual_rmsnorm_residual_kernel(
+ X1_ptr,
+ W1_ptr,
+ X2_ptr,
+ W2_ptr,
+ W3_ptr,
+ Residual_ptr,
+ ... | 91 | 0 | 1 | [
"gemma",
"llava/vlm"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/gemma4_fused_ops.py).
Goal: [VLM] Optimize Gemma4 VLM with PCG and fuse RMSNorm + residual add + scalar
Target hardware: unspecified
Target models: gemma, llava/vlm
Current file:
```python
"""Fused triton kernels for Gemma4 decoder layer operation... | <!-- 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.io to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [{"scope": "e2e", "hardware": null, "model": "Gemma4", "shape": null, "batch_size": 93, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 35.6, "optimized": 16.5, "stated_speedup": null, "speedup": 2.1576}, {"scope": "e2e", "hardware": null, "model": "Gemma4", "shape": null, "batch_size": 189, "... | null | 1.3865 |
sglang | 587fd15bd272d03f78dcc58a3e771a88d0c40f64 | 2026-04-24T12:05:16-07:00 | 21,985 | https://github.com/sgl-project/sglang/pull/21985 | perf: eliminate attention DtoD copy by passing pre-allocated output to FA (#21985) | perf: eliminate attention DtoD copy by passing pre-allocated output to FA (#21985)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> | sgl-kernel/csrc/flash_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/flash_extension.cc
+++ sgl-kernel/csrc/flash_extension.cc
@@ -29,7 +29,7 @@
" Tensor? k_new," // (b, s_k_new, h_k, d) or (total_k_new, h_k, d)
" Tensor? v_new," // (b, s_k_new, h_k, dv) or (total_k_new, h_k, dv)
" Tensor? q_v," /... | 2 | 2 | 4 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/flash_extension.cc).
Goal: perf: eliminate attention DtoD copy by passing pre-allocated output to FA
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache Li... | ## Summary
Eliminate a ~14µs DtoD memcpy per decoder layer by passing the pre-allocated output tensor directly to Flash Attention, rather than letting FA allocate internally and then copying.
### Root Cause
`unified_attention_with_output` pre-allocates an output tensor, calls `attn_backend.forward()` which invokes F... | [{"scope": "e2e", "hardware": "H200", "model": "Qwen3-0.6B FP8", "shape": null, "batch_size": null, "metric": "throughput", "unit": "items/sec", "higher_is_better": true, "baseline": 30.77, "optimized": 37.83, "stated_speedup": null, "speedup": 1.2294}, {"scope": "e2e", "hardware": "H200", "model": "Qwen3-0.6B FP8", "s... | 6.6 | 1.122 |
sglang | b2af34be5404c3ca8f69642ea464470901871711 | 2026-04-14T23:50:32-07:00 | 22,844 | https://github.com/sgl-project/sglang/pull/22844 | [AMD] Optimize _append_shared_to_topk_output by a single fused Triton kernel for Qwen3.5 (#22844) | [AMD] Optimize _append_shared_to_topk_output by a single fused Triton kernel for Qwen3.5 (#22844)
Co-authored-by: HaiShaw <hixiao@gmail.com> | python/sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py | triton | Apache-2.0 | from __future__ import annotations
import functools
from collections import OrderedDict
from typing import Any, Dict, List, Optional
import torch
import triton
import triton.language as tl
from sglang.srt.batch_invariant_ops import is_batch_invariant_mode_enabled
from sglang.srt.layers.moe.utils import get_moe_paddi... | from __future__ import annotations
import functools
from collections import OrderedDict
from typing import Any, Dict, List, Optional
import torch
import triton
import triton.language as tl
from sglang.srt.batch_invariant_ops import is_batch_invariant_mode_enabled
from sglang.srt.layers.moe.utils import get_moe_paddi... | --- 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
@@ -1194,3 +1194,79 @@
num_warps=1,
)
return out_ids, out_weights
+
+
+@triton.jit
+def _fused_append_shared_experts_with_weights_kernel(
+ to... | 76 | 0 | 1 | [
"qwen"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py).
Goal: [AMD] Optimize _append_shared_to_topk_output by a single fused Triton kernel for Qwen3.5
Target hardware: unspecified
Target models: qwen
Current file:
```python
from __future__ import annot... | ## Motivation
- `_append_shared_to_topk_output` previously used 4 separate kernel launches (2 elementwise + 2 concat) to append shared expert IDs and weights.
- This path is on the critical MoE routing path and adds avoidable launch overhead.
- Qwen3.5 requires per-token shared weights from learned `shared_expert_gate... | [{"scope": "e2e", "hardware": null, "model": "Qwen3.5", "shape": "ilen=8192,olen=8192", "batch_size": 4, "metric": "TTT", "unit": "tok/s", "higher_is_better": true, "baseline": 834.29, "optimized": 868.99, "stated_speedup": null, "speedup": 1.0416}, {"scope": "e2e", "hardware": null, "model": "Qwen3.5", "shape": "ilen=... | null | 1.0195 |
sglang | 18f41ac427461b75ba769d8589196fe2a9a2fcb5 | 2026-04-10T14:56:05+08:00 | 22,316 | https://github.com/sgl-project/sglang/pull/22316 | [Reland] DeepSeek-R1-0528-w4a8: DeepEP Low Latency Dispatch Adopts FP8 Communication (#22316) | [Reland] DeepSeek-R1-0528-w4a8: DeepEP Low Latency Dispatch Adopts FP8 Communication (#22316)
Co-authored-by: undefined <zhouchen.arrebol@jd.com>
Co-authored-by: xq25478 <xq25478@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
@@ -1381,3 +1381,76 @@
NUM_STAGE=NUM_STAGES,
)
return output
+
+
+@triton.jit
+def _fp8_per_token_quant_to_per_tensor_quant_kernel(
+ x_ptr,
+ x_scale_ptr,
+ x_scale_stride0,
+ x_scale_st... | 73 | 0 | 5 | [
"deepseek"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/ep_moe/kernels.py).
Goal: [Reland] DeepSeek-R1-0528-w4a8: DeepEP Low Latency Dispatch Adopts FP8 Communication
Target hardware: unspecified
Target models: deepseek
Current file:
```python
import logging
import torch
import triton
from sglang... | <!-- 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.io to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [{"scope": "e2e", "hardware": "H20", "model": "DeepSeek-R1-0528-W4AFP8", "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 | 7119d5974798d157bc3a16cc9074e84c2b9c4c9a | 2026-03-30T22:27:28+08:00 | 14,162 | https://github.com/sgl-project/sglang/pull/14162 | DeepSeek-R1-0528-w4a8: DeepEP Low Latency Dispatch Adopts FP8 Communication (#14162) | DeepSeek-R1-0528-w4a8: DeepEP Low Latency Dispatch Adopts FP8 Communication (#14162)
Co-authored-by: undefined <zhouchen.arrebol@jd.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
@@ -1381,3 +1381,76 @@
NUM_STAGE=NUM_STAGES,
)
return output
+
+
+@triton.jit
+def _fp8_per_token_quant_to_per_tensor_quant_kernel(
+ x_ptr,
+ x_scale_ptr,
+ x_scale_stride0,
+ x_scale_st... | 73 | 0 | 5 | [
"deepseek"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/ep_moe/kernels.py).
Goal: DeepSeek-R1-0528-w4a8: DeepEP Low Latency Dispatch Adopts FP8 Communication
Target hardware: unspecified
Target models: deepseek
Current file:
```python
import logging
import torch
import triton
from sglang.srt.util... |
## Motivation
profiling:
deepseek-R1-0528:
<img width="3170" height="510" alt="ME1764483666809" src="https://github.com/user-attachments/assets/d2a2c921-7c41-4308-8420-a75bdf40cb83" />
deepseek-R1-0528-w4afp8:
<img width="3092" height="576" alt="ME1764483641467" src="https://github.com/user-attachments/assets/6489b43... | [{"scope": "e2e", "hardware": "H20", "model": "DeepSeek-R1-0528-W4AFP8", "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 | 6da8f5f69e0f74fa4555234f8bd016b092af9592 | 2026-03-30T14:43:16+08:00 | 14,702 | https://github.com/sgl-project/sglang/pull/14702 | fix topk softmax performance issue (#14702) | fix topk softmax performance issue (#14702) | 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
@@ -227,11 +227,9 @@
queue[e] = {scores[e], e};
}
- std::partial_sort(
- queue.begin(),
- queue.begin() + num_experts_per_group,
- queue.end(),
- [](const elem_t& x, const elem_t& y) -> bool { re... | 3 | 5 | 1 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/topk.cpp).
Goal: fix topk softmax performance issue
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* __restrict__ out,... | This is a minor change to fix performance issue of topk softmax.
orginal code uses full sort but we only need topk.
## Checklist
- [ ] Format your code according to the [Format code with pre-commit](https://docs.sglang.io/developer_guide/contribution_guide.html#format-code-with-pre-commit).
- [ ] Add unit tests acco... | [] | null | null |
sglang | a12fea21ed08f1a5a808de2ed21bf8e43165e159 | 2026-03-25T13:17:27-07:00 | 21,103 | https://github.com/sgl-project/sglang/pull/21103 | perf(sgl-kernel): expose get_scheduler_metadata for FA3 decode optimization (#21103) | perf(sgl-kernel): expose get_scheduler_metadata for FA3 decode optimization (#21103) | sgl-kernel/csrc/flash_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/flash_extension.cc
+++ sgl-kernel/csrc/flash_extension.cc
@@ -61,6 +61,40 @@
") -> (Tensor, Tensor, Tensor, Tensor)"); // NEW return type: tuple of 4 tensors
m.impl("fwd", torch::kCUDA, make_pytorch_shim(&mha_fwd));
+
+ /*
+ * From flash-attention: get_scheduler_metadata
+ * Precom... | 34 | 0 | 3 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/flash_extension.cc).
Goal: perf(sgl-kernel): expose get_scheduler_metadata for FA3 decode optimization
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache ... | ## Summary
Register `mha_fwd_get_scheduler_metadata` as a torch op in sgl_kernel namespace. This function precomputes FA3 tile scheduling metadata so that the `prepare_varlen_num_blocks` kernel does not need to run per-layer during decode.
The C++ symbol already exists in `flash_ops.so` (compiled from sgl-attn) but w... | [] | null | null |
sglang | a12fea21ed08f1a5a808de2ed21bf8e43165e159 | 2026-03-25T13:17:27-07:00 | 21,103 | https://github.com/sgl-project/sglang/pull/21103 | perf(sgl-kernel): expose get_scheduler_metadata for FA3 decode optimization (#21103) | perf(sgl-kernel): expose get_scheduler_metadata for FA3 decode optimization (#21103) | sgl-kernel/include/sgl_flash_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_flash_kernel_ops.h
+++ sgl-kernel/include/sgl_flash_kernel_ops.h
@@ -83,3 +83,34 @@
std::optional<bool> pack_gqa_,
int64_t sm_margin,
std::optional<const at::Tensor>& sinks_); // (h)
+
+/*
+ * From flash-attention: get_scheduler_metadata
+ * Precomputes tile scheduling metadat... | 31 | 0 | 3 | [] | [] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/include/sgl_flash_kernel_ops.h).
Goal: perf(sgl-kernel): expose get_scheduler_metadata for FA3 decode optimization
Target hardware: unspecified
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the ... | ## Summary
Register `mha_fwd_get_scheduler_metadata` as a torch op in sgl_kernel namespace. This function precomputes FA3 tile scheduling metadata so that the `prepare_varlen_num_blocks` kernel does not need to run per-layer during decode.
The C++ symbol already exists in `flash_ops.so` (compiled from sgl-attn) but w... | [] | null | null |
sglang | 126cd5cfae7afbc9c7efb2b00dd52807ff17b842 | 2026-03-19T13:30:03+08:00 | 20,392 | https://github.com/sgl-project/sglang/pull/20392 | gpt-oss decode performance optimization (#20392) | gpt-oss decode performance optimization (#20392)
Co-authored-by: wunhuang <wunhuang@amd.com> | python/sglang/srt/layers/attention/utils.py | triton | Apache-2.0 | import torch
import triton
import triton.language as tl
from sglang.srt.utils import is_cuda
_FLASHMLA_CREATE_KV_BLOCK_SIZE = 4096
FLASHMLA_CREATE_KV_BLOCK_SIZE_TRITON = tl.constexpr(_FLASHMLA_CREATE_KV_BLOCK_SIZE)
_is_cuda = is_cuda()
if _is_cuda:
from sgl_kernel import concat_mla_absorb_q
@triton.jit
def cr... | import torch
import triton
import triton.language as tl
from sglang.srt.utils import is_cuda
_FLASHMLA_CREATE_KV_BLOCK_SIZE = 4096
FLASHMLA_CREATE_KV_BLOCK_SIZE_TRITON = tl.constexpr(_FLASHMLA_CREATE_KV_BLOCK_SIZE)
_is_cuda = is_cuda()
if _is_cuda:
from sgl_kernel import concat_mla_absorb_q
@triton.jit
def cr... | --- python/sglang/srt/layers/attention/utils.py
+++ python/sglang/srt/layers/attention/utils.py
@@ -472,3 +472,189 @@
return concat_mla_absorb_q(q_nope, q_rope)
else:
return torch.cat([q_nope, q_rope], dim=-1)
+
+
+@triton.jit
+def reshape_and_cache_flash(
+ key_ptr,
+ value_ptr,
+ key_c... | 186 | 0 | 3 | [
"gpt-oss"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/utils.py).
Goal: gpt-oss decode performance optimization
Target hardware: unspecified
Target models: gpt-oss
Current file:
```python
import torch
import triton
import triton.language as tl
from sglang.srt.utils import is_cuda
_FLASHMLA... | <!-- 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.io to discuss further. -->
## Motivation
Improve the performance for gpt-oss model run... | [] | null | null |
sglang | 687d9eb66f4c853f7eb772b830f1181a102a8364 | 2026-03-19T13:18:15+08:00 | 15,168 | https://github.com/sgl-project/sglang/pull/15168 | [CPU] Optimize image preprocessor performance for Qwen2VLImageProcessorFast (#15168) | [CPU] Optimize image preprocessor performance for Qwen2VLImageProcessorFast (#15168) | 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
@@ -127,6 +127,8 @@
#define CHECK_DIM(d, x) TORCH_CHECK(x.dim() == d, #x " must be a " #d "D tensor")
#define CHECK_EQ(a, b) TORCH_CHECK((a) == (b), "CHECK_EQ(" #a ", " #b ") failed. ", a, " vs ", b)
+#define CHECK_GT(a, b) TORCH_CHECK((a) > (b), "CH... | 2 | 1 | 2 | [
"qwen"
] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/common.h).
Goal: [CPU] Optimize image preprocessor performance for Qwen2VLImageProcessorFast
Target hardware: CPU
Target models: qwen
Current file:
```cpp
#pragma once
#include <ATen/ATen.h>
#include <ATen/Parallel.h>
#include <ATen/record_function.h>
#... | ## Motivation
Optimize the performance image preprocessor on intel Xeon CPU with AMX support.
## Modifications
Add kernel implementation of
```
std::tuple<at::Tensor, at::Tensor> image_preprocess_cpu(
at::TensorList images,
bool do_convert_rgb,
bool do_resize,
int64_t shortest_edge,
int64_t ... | [] | null | null |
sglang | 687d9eb66f4c853f7eb772b830f1181a102a8364 | 2026-03-19T13:18:15+08:00 | 15,168 | https://github.com/sgl-project/sglang/pull/15168 | [CPU] Optimize image preprocessor performance for Qwen2VLImageProcessorFast (#15168) | [CPU] Optimize image preprocessor performance for Qwen2VLImageProcessorFast (#15168) | 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
@@ -359,6 +359,25 @@
int64_t head_qk,
int64_t head_v);
+// image preprocessor
+std::tuple<at::Tensor, at::Tensor> image_preprocess_cpu(
+ at::TensorList images,
+ bool do_convert_rgb,
+ bool do_resize,
+... | 27 | 0 | 2 | [
"qwen"
] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/torch_extension_cpu.cpp).
Goal: [CPU] Optimize image preprocessor performance for Qwen2VLImageProcessorFast
Target hardware: CPU
Target models: qwen
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache Licens... | ## Motivation
Optimize the performance image preprocessor on intel Xeon CPU with AMX support.
## Modifications
Add kernel implementation of
```
std::tuple<at::Tensor, at::Tensor> image_preprocess_cpu(
at::TensorList images,
bool do_convert_rgb,
bool do_resize,
int64_t shortest_edge,
int64_t ... | [] | null | null |
sglang | 62d74549762a2c3684addc71353357f1e219bbc7 | 2026-03-19T13:17:53+08:00 | 16,040 | https://github.com/sgl-project/sglang/pull/16040 | optimize conv3d used in patch embedding (#16040) | optimize conv3d used in patch embedding (#16040) | 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
@@ -294,6 +294,11 @@
int64_t pad_slot_id,
bool is_vnni);
+// conv3d fast path for patch embedding
+at::Tensor conv3d_embed_weight_pack(const at::Tensor& weight);
+
+at::Tensor conv3d_embed_cpu(const at::Tensor& i... | 11 | 0 | 1 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/torch_extension_cpu.cpp).
Goal: optimize conv3d used in patch embedding
Target hardware: CPU
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.io to discuss further. -->
## Motivation
Optimize `nn.Conv3d` used in patch embedding ... | [] | null | null |
sglang | 3d58cd16d940e2437e58e146d780f54e1112e2b9 | 2026-03-16T18:44:42+08:00 | 20,406 | https://github.com/sgl-project/sglang/pull/20406 | [DP Attention] Optimize dp_padding_mode selection for dp_size=1 in extend mode (#20406) | [DP Attention] Optimize dp_padding_mode selection for dp_size=1 in extend mode (#20406)
Signed-off-by: wangfakang <fakangwang@gmail.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 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_at... | 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_at... | --- python/sglang/srt/layers/dp_attention.py
+++ python/sglang/srt/layers/dp_attention.py
@@ -64,13 +64,20 @@
def get_dp_padding_mode(
cls, is_extend_in_batch, global_num_tokens: List[int]
) -> DpPaddingMode:
- if is_extend_in_batch:
+ dp_size = get_attention_dp_size()
+
+ # When... | 9 | 2 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/dp_attention.py).
Goal: [DP Attention] Optimize dp_padding_mode selection for dp_size=1 in extend mode
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from __future__ import annotations
import functools
import logg... | CC @yizhang2077 @ShangmingCai @nvcastet @ch-wan @merrymercy @Fridge003 PTAL, thx.
## Motivation
When dp_size=1, the `MAX_LEN` and `SUM_LEN` modes have identical communication overhead since max_len equals sum_len. Previously, extend mode(`get_dp_padding_mode`) unconditionally used `SUM_LEN`, which prevented symmetri... | [] | null | null |
sglang | 80a6b32703db7f0fe1ef69fa9b5e2154f3e51258 | 2026-03-01T04:59:26-05:00 | 19,536 | https://github.com/sgl-project/sglang/pull/19536 | [Perf] Optimize NSA backend metadata under MTP (#19536) | [Perf] Optimize NSA backend metadata under MTP (#19536)
Co-authored-by: Baidu-AIAK <Baidu_AIAK@163.com>
Co-authored-by: zengpai <zengpai@baidu.com> | python/sglang/srt/layers/attention/utils.py | triton | Apache-2.0 | import torch
import triton
import triton.language as tl
from sglang.srt.utils import is_cuda
_FLASHMLA_CREATE_KV_BLOCK_SIZE = 4096
FLASHMLA_CREATE_KV_BLOCK_SIZE_TRITON = tl.constexpr(_FLASHMLA_CREATE_KV_BLOCK_SIZE)
_is_cuda = is_cuda()
if _is_cuda:
from sgl_kernel import concat_mla_absorb_q
@triton.jit
def cr... | import torch
import triton
import triton.language as tl
from sglang.srt.utils import is_cuda
_FLASHMLA_CREATE_KV_BLOCK_SIZE = 4096
FLASHMLA_CREATE_KV_BLOCK_SIZE_TRITON = tl.constexpr(_FLASHMLA_CREATE_KV_BLOCK_SIZE)
_is_cuda = is_cuda()
if _is_cuda:
from sgl_kernel import concat_mla_absorb_q
@triton.jit
def cr... | --- python/sglang/srt/layers/attention/utils.py
+++ python/sglang/srt/layers/attention/utils.py
@@ -284,6 +284,67 @@
)
return B, output, attn_mask
+
+
+@triton.jit
+def seqlens_expand_kernel(
+ extend_seq_lens_ptr, # [N]
+ seq_lens_ptr, # [N]
+ offsets_ptr, # [N+1]
+ output_ptr, # [sum(exten... | 61 | 0 | 2 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/utils.py).
Goal: [Perf] Optimize NSA backend metadata under MTP
Target hardware: unspecified
Target models: unspecified
Current file:
```python
import torch
import triton
import triton.language as tl
from sglang.srt.utils import 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.io to discuss further. -->
## Motivation
This PR is authored by @Baidu-AIAK in https:/... | [{"scope": "e2e", "hardware": null, "model": "nvidia/DeepSeek-V3.2-NVFP4", "shape": null, "batch_size": 200, "metric": "latency", "unit": "s", "higher_is_better": false, "baseline": 139.013, "optimized": 153.288, "stated_speedup": null, "speedup": 0.9069}, {"scope": "e2e", "hardware": null, "model": "nvidia/DeepSeek-V3... | null | 0.9912 |
sglang | a1b39c1c26d41203bcce61be35c2676c5c646929 | 2026-02-25T15:40:55+08:00 | 18,088 | https://github.com/sgl-project/sglang/pull/18088 | Perf/fuse mamba state scatter mtp verify (#18088) | Perf/fuse mamba state scatter mtp verify (#18088) | python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py | triton | Apache-2.0 | import logging
import math
from typing import Optional, Union
import torch
import triton
import triton.language as tl
from einops import rearrange
from sglang.srt.environ import Envs
from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
from sglang.srt.layers.attention.fla.fused_gdn_gating import... | import logging
import math
from typing import Optional, Union
import torch
import triton
import triton.language as tl
from einops import rearrange
from sglang.srt.environ import Envs
from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
from sglang.srt.layers.attention.fla.fused_gdn_gating import... | --- python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
+++ python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
@@ -31,6 +31,9 @@
from sglang.srt.layers.attention.mamba.mamba2_metadata import (
ForwardMetadata,
Mamba2Metadata,
+)
+from sglang.srt.layers.attention.mamba.mamba_state_s... | 39 | 37 | 1 | [
"mamba"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py).
Goal: Perf/fuse mamba state scatter mtp verify
Target hardware: unspecified
Target models: mamba
Current file:
```python
import logging
import math
from typing import Optional, Union
import torch
import t... | ## Summary
This PR introduces a fused Triton kernel for Mamba state scatter operations in MTP (Multi-Token Prediction) verify stage, reducing kernel launch overhead and improving throughput.
## Changes
- **New fused kernel**: `fused_mamba_state_scatter_with_mask` in `mamba_state_scatter_triton.py`
- Fuses multiple `i... | [{"scope": "e2e", "hardware": "8xH20", "model": "Qwen3-Next-80B-A3B", "shape": null, "batch_size": 48, "metric": "throughput", "unit": "tok/s", "higher_is_better": true, "baseline": 1830.6, "optimized": 1901.81, "stated_speedup": 1.039, "speedup": 1.0389}] | null | 1.0389 |
sglang | fa0ef6e4f7f7c8937e498d8db6fa8599405d2a18 | 2026-02-14T14:35:26+08:00 | 18,782 | https://github.com/sgl-project/sglang/pull/18782 | [VLM][LLM] Optimize fused_moe triton kernel tma (#18782) | [VLM][LLM] Optimize fused_moe triton kernel tma (#18782)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com> | python/sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py | triton | Apache-2.0 | from __future__ import annotations
import functools
import os
from typing import Any, Dict, List, Optional
import torch
import triton
import triton.language as tl
from sglang.srt.batch_invariant_ops import is_batch_invariant_mode_enabled
from sglang.srt.layers.quantization.fp8_kernel import (
per_token_group_qua... | from __future__ import annotations
import functools
import os
from collections import OrderedDict
from typing import Any, Dict, List, Optional
import torch
import triton
import triton.language as tl
from sglang.srt.batch_invariant_ops import is_batch_invariant_mode_enabled
from sglang.srt.layers.quantization.fp8_ker... | --- 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
@@ -2,6 +2,7 @@
import functools
import os
+from collections import OrderedDict
from typing import Any, Dict, List, Optional
import torch
@@ -610,6 +611,67 ... | 68 | 9 | 1 | [
"llava/vlm"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py).
Goal: [VLM][LLM] Optimize fused_moe triton kernel tma
Target hardware: unspecified
Target models: llava/vlm
Current file:
```python
from __future__ import annotations
import functools
import os
f... | <!-- 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.io to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [{"scope": "e2e", "hardware": null, "model": "Qwen3-VL-30B-A3B-Instruct", "shape": null, "batch_size": 32, "metric": "request throughput", "unit": "req/s", "higher_is_better": true, "baseline": 13.09, "optimized": 14.34, "stated_speedup": null, "speedup": 1.0955}, {"scope": "e2e", "hardware": null, "model": "Qwen3-VL-3... | null | 1.0952 |
sglang | 8be18c655d0ff1c2e673395102c97c9d5cbd7d1c | 2026-02-13T17:30:50-08:00 | 17,613 | https://github.com/sgl-project/sglang/pull/17613 | [Perf] refactor piecewise cuda graph support of Qwen3-Next (#17613) | [Perf] refactor piecewise cuda graph support of Qwen3-Next (#17613) | 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
@@ -14,6 +14,7 @@
import triton.language as tl
from einops import rearrange
+from sglang.srt.server_args import get_global_server_args
from sglang.srt.utils import (
cdiv,
cpu_has_a... | 13 | 1 | 2 | [
"qwen"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/fla/layernorm_gated.py).
Goal: [Perf] refactor piecewise cuda graph support of Qwen3-Next
Target hardware: unspecified
Target models: qwen
Current file:
```python
# Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/f... | <!-- 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.io to discuss further. -->
## Motivation
Currently in Qwen3Next, we hide all ops inside... | [{"scope": "e2e", "hardware": null, "model": "Qwen3-Next-80B-A3B-Instruct", "shape": null, "batch_size": null, "metric": "latency", "unit": "s", "higher_is_better": false, "baseline": 12.379, "optimized": 11.041, "stated_speedup": null, "speedup": 1.1212}, {"scope": "e2e", "hardware": null, "model": "Qwen3-Next-80B-A3B... | null | 1.143 |
sglang | 9e9e94926162d0b2bf189c9722e97dbdc59b7375 | 2026-02-12T23:43:22+08:00 | 18,586 | https://github.com/sgl-project/sglang/pull/18586 | speed up sgl-kernel build (#18586) | speed up sgl-kernel build (#18586) | 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"
#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... | #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
@@ -215,3 +215,4 @@
cudaError_t err = cudaGetLastError();
TORCH_CHECK(err == cudaSuccess, "CUDA kernel launch failed: ", cudaGetErrorString(err));
}
+// test-1 | 1 | 0 | 1 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/elementwise/concat_mla.cu).
Goal: speed up sgl-kernel build
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
#include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/CUDADataType.h>
#include <cuda_runtime.h>
#include "pytorch_exte... | <!-- 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.io to discuss further. -->
## Motivation
<img width="2280" height="1032" alt="e2b78baf... | [] | null | null |
sglang | 4637970dfb369d900936136c98cb5a6663385884 | 2026-02-06T19:59:42-08:00 | 18,271 | https://github.com/sgl-project/sglang/pull/18271 | [Qwen3Next] Optimize fused_sigmoid_gating_delta_rule_update_kernel (#18271) | [Qwen3Next] Optimize fused_sigmoid_gating_delta_rule_update_kernel (#18271) | python/sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py | triton | Apache-2.0 | from typing import Optional
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.fla.utils import input_guard
@triton.jit(do_not_specialize=["T"])
def fused_sigmoid_gating_delta_rule_update_kernel(
A_log,
a,
dt_bias,
softplus_beta,
softplus_threshold,
q,
... | from typing import Optional
import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.fla.utils import input_guard
@triton.jit(do_not_specialize=["T"])
def fused_sigmoid_gating_delta_rule_update_kernel(
A_log,
a,
dt_bias,
softplus_beta,
softplus_threshold,
q,
... | --- python/sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py
+++ python/sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py
@@ -190,7 +190,7 @@
B, T, H, K, V = *k.shape, v.shape[-1]
HV = v.shape[2]
N = B if cu_seqlens is None else len(cu_seqlens) - 1
- BK, BV = triton.nex... | 1 | 1 | 1 | [
"qwen"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py).
Goal: [Qwen3Next] Optimize fused_sigmoid_gating_delta_rule_update_kernel
Target hardware: unspecified
Target models: qwen
Current file:
```python
from typing import Optional
import torch
import tr... | <!-- 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.io to discuss further. -->
## Motivation
Apply the same optimization in https://github... | [{"scope": "e2e", "hardware": "B200", "model": null, "shape": null, "batch_size": 512, "metric": "TPOT", "unit": "ms", "higher_is_better": false, "baseline": 35.79, "optimized": 34.64, "stated_speedup": null, "speedup": 1.0332}, {"scope": "kernel", "hardware": "B200", "model": null, "shape": null, "batch_size": null, "... | 1.5921 | 1.0332 |
sglang | 760ae933bb3878a6897e7e552a746929c29e9d90 | 2026-02-04T19:59:41+08:00 | 16,043 | https://github.com/sgl-project/sglang/pull/16043 | optimize get_topk_ragged by fusing get k and k_scale triton kernel (#16043) | optimize get_topk_ragged by fusing get k and k_scale triton kernel (#16043)
Co-authored-by: abing <wangbingjia.wbj@alibaba-inc.com> | 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
from sglang.srt.layers.quantization.fp8_kernel import is_fp8_fnuz
from sglang.srt.utils import is_hip
_is_hip = is_hip()
_is_fp8_fnuz = is_fp8_fnuz()
if TYPE_CHECKING:
from sglang.srt.mem_cache.memory_pool import NSATokenTo... | from typing import TYPE_CHECKING
import torch
import triton
import triton.language as tl
from sglang.srt.layers.quantization.fp8_kernel import is_fp8_fnuz
from sglang.srt.utils import is_hip
_is_hip = is_hip()
_is_fp8_fnuz = is_fp8_fnuz()
if TYPE_CHECKING:
from sglang.srt.mem_cache.memory_pool import NSATokenTo... | --- python/sglang/srt/layers/attention/nsa/index_buf_accessor.py
+++ python/sglang/srt/layers/attention/nsa/index_buf_accessor.py
@@ -167,11 +167,20 @@
@classmethod
def triton(
- cls, pool: "NSATokenToKVPool", buf, seq_len: int, page_indices: torch.Tensor
+ cls,
+ pool: "NSATokenToKVPoo... | 70 | 26 | 2 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/nsa/index_buf_accessor.py).
Goal: optimize get_topk_ragged by fusing get k and k_scale triton kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```python
from typing import TYPE_CHECKING
import torch
import tr... | - **Summary**
Optimize the get_index_k_scale_buffer function to reduce the number of concatenation operations of k_fp8 and k_scale.
<img width="707" height="328" alt="image" src="https://github.com/user-attachments/assets/a78f6c66-1857-4faa-baea-db31c8a87b05" />
| [] | null | null |
sglang | afebb7ab7893869ec8cf7ed6dd6f06981bc8ccef | 2026-02-01T18:59:31+08:00 | 17,674 | https://github.com/sgl-project/sglang/pull/17674 | Optimize custom-all-reduce (#17674) | Optimize custom-all-reduce (#17674)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com> | sgl-kernel/csrc/allreduce/custom_all_reduce.cuh | cuda | Apache-2.0 | // Adapted from https://github.com/vllm-project/vllm/blob/v0.8.2/csrc/custom_all_reduce.cuh
#pragma once
#include <cuda.h>
#include <cuda_bf16.h>
#include <cuda_fp16.h>
#include <cuda_runtime.h>
#include <array>
#include <iostream>
#include <limits>
#include <map>
#include <unordered_map>
#include <vector>
#include ... | // Adapted from https://github.com/vllm-project/vllm/blob/v0.8.2/csrc/custom_all_reduce.cuh
#pragma once
#include <cuda.h>
#include <cuda_bf16.h>
#include <cuda_fp16.h>
#include <cuda_runtime.h>
#include <array>
#include <iostream>
#include <limits>
#include <map>
#include <unordered_map>
#include <vector>
#include ... | --- sgl-kernel/csrc/allreduce/custom_all_reduce.cuh
+++ sgl-kernel/csrc/allreduce/custom_all_reduce.cuh
@@ -442,21 +442,44 @@
size /= d;
auto bytes = size * sizeof(typename packed_t<T>::P);
int blocks = std::min(block_limit, (size + threads - 1) / threads);
+
+ // Check environment variable once
+ ... | 35 | 12 | 1 | [] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/allreduce/custom_all_reduce.cuh).
Goal: Optimize custom-all-reduce
Target hardware: unspecified
Target models: unspecified
Current file:
```cuda
// Adapted from https://github.com/vllm-project/vllm/blob/v0.8.2/csrc/custom_all_reduce.cuh
#pragma once
#includ... | <!-- 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.io to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [{"scope": "kernel", "hardware": null, "model": null, "shape": null, "batch_size": null, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 386, "optimized": 166, "stated_speedup": null, "speedup": 2.3253}] | 2.3253 | null |
sglang | 81449b4beebf761eab1ad67ad19e3cab4c0109d4 | 2026-01-31T01:02:12+08:00 | 17,094 | https://github.com/sgl-project/sglang/pull/17094 | Optimize GDN decode for Qwen3 Next (#17094) | Optimize GDN decode for Qwen3 Next (#17094) | 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
@@ -136,7 +136,7 @@
B, T, H, K, V = *k.shape, v.shape[-1]
HV = v.shape[2]
N = B if cu_seqlens is None else len(cu_seqlens) - 1
- BK, BV = triton.next_power_of_2(K), min(triton.ne... | 1 | 1 | 1 | [
"qwen"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/fla/fused_recurrent.py).
Goal: Optimize GDN decode for Qwen3 Next
Target hardware: unspecified
Target models: qwen
Current file:
```python
# Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/gated_delta_rule/... | <!-- 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.io to discuss further. -->
## Motivation
As MoE/FullAttn/GEMM has been optimized, the G... | [{"scope": "e2e", "hardware": "B200", "model": "Qwen3", "shape": null, "batch_size": null, "metric": "Latency", "unit": "s", "higher_is_better": false, "baseline": 74.87, "optimized": 67.961, "stated_speedup": null, "speedup": 1.1017}, {"scope": "e2e", "hardware": "B200", "model": "Qwen3", "shape": null, "batch_size": ... | null | 1.1013 |
sglang | 88f77594023d6a0c9833c450eddc73b4a146541c | 2026-01-30T14:07:05+08:00 | 15,708 | https://github.com/sgl-project/sglang/pull/15708 | [CPU] optimize flash_attn_varlen_func (#15708) | [CPU] optimize flash_attn_varlen_func (#15708) | 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
@@ -1049,7 +1049,7 @@
int64_t k_strideH,
int64_t v_strideN,
int64_t v_strideH,
- float scaling,
+ float sm_scale,
float logit_cap,
int64_t max_num_reqs,
int64_t max_context_len,
@@ -1103,7 +1103,7 @@
/... | 7 | 7 | 5 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/decode.cpp).
Goal: [CPU] optimize flash_attn_varlen_func
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 value ... |
## Motivation
provide apple to apple function of `flash_attn_varlen_func` on CPU, optimized with intel AMX. Newly enabled models such as omni and sglang diffusion used `flash_attn_varlen_func` directly on CUDA device, so need to provide a counterpart on other devices.
## Modifications
Add `flash_attn.cpp` in sgl-ke... | [{"scope": "kernel", "hardware": "Intel(R) Xeon(R) 6979P", "model": "qwen3 omni", "shape": "T=8160,H=6,Hkv=6,K=72", "batch_size": 1, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 23.993, "optimized": 19.787, "stated_speedup": null, "speedup": 1.2126}] | 1.2126 | null |
sglang | 88f77594023d6a0c9833c450eddc73b4a146541c | 2026-01-30T14:07:05+08:00 | 15,708 | https://github.com/sgl-project/sglang/pull/15708 | [CPU] optimize flash_attn_varlen_func (#15708) | [CPU] optimize flash_attn_varlen_func (#15708) | sgl-kernel/csrc/cpu/extend.cpp | cpp | Apache-2.0 | #include "common.h"
#include "gemm.h"
#include "vec.h"
#include "vec_pack.h"
namespace {
// [NOTE]: extend attention for CPU
// 1. tune BLOCK_M and BLOCK_N
// 2. can handle non-contiguous k_exttend and v_extend
// 3. computes attention for prefix and extend separately
// 4. TODO: vectorize `pack_vnni` and `pa... | #include "common.h"
#include "flash_attn.h"
#include "gemm.h"
namespace {
// [NOTE]: extend attention for CPU
// 1. BLOCK_M and BLOCK_N tuned for various seq lengths
// 2. can handle non-contiguous k_extend and v_extend
// 3. computes attention for prefix and extend separately
// 4. TODO: apply head dimension... | --- sgl-kernel/csrc/cpu/extend.cpp
+++ sgl-kernel/csrc/cpu/extend.cpp
@@ -1,70 +1,15 @@
#include "common.h"
+#include "flash_attn.h"
#include "gemm.h"
-#include "vec.h"
-#include "vec_pack.h"
namespace {
// [NOTE]: extend attention for CPU
-// 1. tune BLOCK_M and BLOCK_N
-// 2. can handle non-contiguous k_e... | 81 | 217 | 5 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/extend.cpp).
Goal: [CPU] optimize flash_attn_varlen_func
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#include "common.h"
#include "gemm.h"
#include "vec.h"
#include "vec_pack.h"
namespace {
// [NOTE]: extend attention for CPU
//... |
## Motivation
provide apple to apple function of `flash_attn_varlen_func` on CPU, optimized with intel AMX. Newly enabled models such as omni and sglang diffusion used `flash_attn_varlen_func` directly on CUDA device, so need to provide a counterpart on other devices.
## Modifications
Add `flash_attn.cpp` in sgl-ke... | [{"scope": "kernel", "hardware": "Intel(R) Xeon(R) 6979P", "model": "qwen3 omni", "shape": "T=8160,H=6,Hkv=6,K=72", "batch_size": 1, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 23.993, "optimized": 19.787, "stated_speedup": null, "speedup": 1.2126}] | 1.2126 | null |
sglang | 88f77594023d6a0c9833c450eddc73b4a146541c | 2026-01-30T14:07:05+08:00 | 15,708 | https://github.com/sgl-project/sglang/pull/15708 | [CPU] optimize flash_attn_varlen_func (#15708) | [CPU] optimize flash_attn_varlen_func (#15708) | 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
@@ -108,6 +108,17 @@
int64_t max_len_extend,
double sm_scale,
double logit_cap);
+
+// flash attention
+at::Tensor flash_attn_varlen_func(
+ const at::Tensor& q,
+ const at::Tensor& k,
+ const at::Ten... | 17 | 0 | 5 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/torch_extension_cpu.cpp).
Goal: [CPU] optimize flash_attn_varlen_func
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
/* Copyright 2025 SGLang Team. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");... |
## Motivation
provide apple to apple function of `flash_attn_varlen_func` on CPU, optimized with intel AMX. Newly enabled models such as omni and sglang diffusion used `flash_attn_varlen_func` directly on CUDA device, so need to provide a counterpart on other devices.
## Modifications
Add `flash_attn.cpp` in sgl-ke... | [{"scope": "kernel", "hardware": "Intel(R) Xeon(R) 6979P", "model": "qwen3 omni", "shape": "T=8160,H=6,Hkv=6,K=72", "batch_size": 1, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 23.993, "optimized": 19.787, "stated_speedup": null, "speedup": 1.2126}] | 1.2126 | null |
sglang | 88f77594023d6a0c9833c450eddc73b4a146541c | 2026-01-30T14:07:05+08:00 | 15,708 | https://github.com/sgl-project/sglang/pull/15708 | [CPU] optimize flash_attn_varlen_func (#15708) | [CPU] optimize flash_attn_varlen_func (#15708) | 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
@@ -148,7 +148,7 @@
#endif
// vector to scalar reduction
-#if defined(CPU_CAPABILITY_AVX512) && 0
+#if defined(CPU_CAPABILITY_AVX512)
inline float vec_reduce_sum(const Vectorized<float>& a) {
return _mm512_reduce_add_ps(__m512(a));
}
@@ -323,6 +323,56... | 51 | 1 | 5 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/vec.h).
Goal: [CPU] optimize flash_attn_varlen_func
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
#pragma once
#if defined(__AVX512F__) && defined(__AVX512BF16__) && defined(__AMX_BF16__)
#define CPU_CAPABILITY_AVX512
#endif
#incl... |
## Motivation
provide apple to apple function of `flash_attn_varlen_func` on CPU, optimized with intel AMX. Newly enabled models such as omni and sglang diffusion used `flash_attn_varlen_func` directly on CUDA device, so need to provide a counterpart on other devices.
## Modifications
Add `flash_attn.cpp` in sgl-ke... | [{"scope": "kernel", "hardware": "Intel(R) Xeon(R) 6979P", "model": "qwen3 omni", "shape": "T=8160,H=6,Hkv=6,K=72", "batch_size": 1, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 23.993, "optimized": 19.787, "stated_speedup": null, "speedup": 1.2126}] | 1.2126 | null |
sglang | 88f77594023d6a0c9833c450eddc73b4a146541c | 2026-01-30T14:07:05+08:00 | 15,708 | https://github.com/sgl-project/sglang/pull/15708 | [CPU] optimize flash_attn_varlen_func (#15708) | [CPU] optimize flash_attn_varlen_func (#15708) | sgl-kernel/csrc/cpu/vec_pack.h | cpp | Apache-2.0 | // To use the transpose functions
#include <ATen/native/cpu/utils.h>
#include "vec.h"
namespace {
using namespace at::vec;
template <typename index_t>
inline index_t get_index(index_t* ind, int i) {
return (ind == nullptr) ? (index_t)i : ind[i];
}
#if defined(CPU_CAPABILITY_AVX512)
// key: from [N, 32] to [32/2,... | // To use the transpose functions
#include <ATen/native/cpu/utils.h>
#include "vec.h"
namespace {
using namespace at::vec;
template <typename index_t>
inline index_t get_index(index_t* ind, int i) {
return (ind == nullptr) ? (index_t)i : ind[i];
}
#if defined(CPU_CAPABILITY_AVX512)
// key: from [N, 32] to [32/2,... | --- sgl-kernel/csrc/cpu/vec_pack.h
+++ sgl-kernel/csrc/cpu/vec_pack.h
@@ -13,30 +13,6 @@
}
#if defined(CPU_CAPABILITY_AVX512)
-// key: from [N, 32] to [32/2, N, 2]
-template <typename scalar_t>
-inline void
-pack_vnni_Nx32(scalar_t* __restrict__ dst, const scalar_t* __restrict__ src, int N, int ld_src, int ld_dst) ... | 144 | 131 | 5 | [] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/vec_pack.h).
Goal: [CPU] optimize flash_attn_varlen_func
Target hardware: CPU
Target models: unspecified
Current file:
```cpp
// To use the transpose functions
#include <ATen/native/cpu/utils.h>
#include "vec.h"
namespace {
using namespace at::vec;
te... |
## Motivation
provide apple to apple function of `flash_attn_varlen_func` on CPU, optimized with intel AMX. Newly enabled models such as omni and sglang diffusion used `flash_attn_varlen_func` directly on CUDA device, so need to provide a counterpart on other devices.
## Modifications
Add `flash_attn.cpp` in sgl-ke... | [{"scope": "kernel", "hardware": "Intel(R) Xeon(R) 6979P", "model": "qwen3 omni", "shape": "T=8160,H=6,Hkv=6,K=72", "batch_size": 1, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 23.993, "optimized": 19.787, "stated_speedup": null, "speedup": 1.2126}] | 1.2126 | null |
sglang | 336dc4579e944728d3efe902a49ad75f195da5e4 | 2026-01-30T14:03:58+08:00 | 12,525 | https://github.com/sgl-project/sglang/pull/12525 | [CPU] Optimize Qwen3-next model on CPU (#12525) | [CPU] Optimize Qwen3-next model on CPU (#12525)
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
Co-authored-by: Fan Yin <1106310035@qq.com> | 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
@@ -14,9 +14,17 @@
import triton.language as tl
from einops import rearrange
-from sglang.srt.utils import cdiv, device_context, is_npu, next_power_of_2
+from sglang.srt.utils import (
+ cd... | 27 | 11 | 8 | [
"qwen"
] | [
"CPU"
] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/fla/layernorm_gated.py).
Goal: [CPU] Optimize Qwen3-next model on CPU
Target hardware: CPU
Target models: qwen
Current file:
```python
# Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/modules/layernorm_gated.p... | This PR adds unified CPU optimizations for Qwen3-next models, including:
1. Add CPU paths to call optimized kernels, which is depending on below sgl-kernels:
a. chunk_gated_delta_rule https://github.com/sgl-project/sglang/pull/12441
b. fused_sigmoid_gating_delta_rule_update and fused_gdn_gating https://git... | [] | null | null |
sglang | 336dc4579e944728d3efe902a49ad75f195da5e4 | 2026-01-30T14:03:58+08:00 | 12,525 | https://github.com/sgl-project/sglang/pull/12525 | [CPU] Optimize Qwen3-next model on CPU (#12525) | [CPU] Optimize Qwen3-next model on CPU (#12525)
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
Co-authored-by: Fan Yin <1106310035@qq.com> | python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py | triton | Apache-2.0 | from typing import Optional, Tuple, Union
import torch
import triton
import triton.language as tl
from einops import rearrange
from sglang.jit_kernel.cutedsl_gdn import cutedsl_fused_sigmoid_gating_delta_rule_update
from sglang.srt.environ import Envs
from sglang.srt.layers.attention.base_attn_backend import Attentio... | from typing import Optional, Tuple, Union
import torch
import triton
import triton.language as tl
from einops import rearrange
from sglang.srt.environ import Envs
from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
from sglang.srt.layers.attention.fla.fused_gdn_gating import fused_gdn_gating
fr... | --- python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
+++ python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py
@@ -5,11 +5,8 @@
import triton.language as tl
from einops import rearrange
-from sglang.jit_kernel.cutedsl_gdn import cutedsl_fused_sigmoid_gating_delta_rule_update
from sglang.... | 35 | 10 | 8 | [
"qwen"
] | [
"CPU"
] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/hybrid_linear_attn_backend.py).
Goal: [CPU] Optimize Qwen3-next model on CPU
Target hardware: CPU
Target models: qwen
Current file:
```python
from typing import Optional, Tuple, Union
import torch
import triton
import triton.language as... | This PR adds unified CPU optimizations for Qwen3-next models, including:
1. Add CPU paths to call optimized kernels, which is depending on below sgl-kernels:
a. chunk_gated_delta_rule https://github.com/sgl-project/sglang/pull/12441
b. fused_sigmoid_gating_delta_rule_update and fused_gdn_gating https://git... | [] | null | null |
sglang | 336dc4579e944728d3efe902a49ad75f195da5e4 | 2026-01-30T14:03:58+08:00 | 12,525 | https://github.com/sgl-project/sglang/pull/12525 | [CPU] Optimize Qwen3-next model on CPU (#12525) | [CPU] Optimize Qwen3-next model on CPU (#12525)
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
Co-authored-by: Fan Yin <1106310035@qq.com> | 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
@@ -524,6 +524,12 @@
break;
case 256:
LAUNCH_TOPK_SOFTMAX_KERNEL(256);
+ break;
+ case 384:
+ LAUNCH_TOPK_SOFTMAX_KERNEL(384);
+ break;
+ case 512:
+ LAUNCH_TOPK_SOFTMAX_KERNEL(512);
... | 6 | 0 | 8 | [
"qwen"
] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/topk.cpp).
Goal: [CPU] Optimize Qwen3-next model on CPU
Target hardware: CPU
Target models: qwen
Current file:
```cpp
#include "common.h"
#include "vec.h"
namespace {
template <typename scalar_t, int SIZE>
inline void softmax(float* __restrict__ out, co... | This PR adds unified CPU optimizations for Qwen3-next models, including:
1. Add CPU paths to call optimized kernels, which is depending on below sgl-kernels:
a. chunk_gated_delta_rule https://github.com/sgl-project/sglang/pull/12441
b. fused_sigmoid_gating_delta_rule_update and fused_gdn_gating https://git... | [] | null | null |
sglang | 82a1b645bad31c3071220c1acd4171209c524013 | 2026-01-17T00:10:52+08:00 | 17,133 | https://github.com/sgl-project/sglang/pull/17133 | [DeepSeek V3.1/V3.2] Optimize fused moe configs for H20 & H20-3E based on swapab (#17133) | [DeepSeek V3.1/V3.2] Optimize fused moe configs for H20 & H20-3E based on swapab (#17133) | python/sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py | triton | Apache-2.0 | from __future__ import annotations
import functools
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 ... | from __future__ import annotations
import functools
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 ... | --- 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
@@ -744,8 +744,8 @@
B.stride(0),
B.stride(2),
B.stride(1),
- C.stride(1),
- C.stride(2),
+ C... | 2 | 2 | 1 | [
"deepseek"
] | [
"H20"
] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py).
Goal: [DeepSeek V3.1/V3.2] Optimize fused moe configs for H20 & H20-3E based on swapab
Target hardware: H20
Target models: deepseek
Current file:
```python
from __future__ import annotations
impo... | <!-- 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.io to discuss further. -->
## Motivation
1. Performance tuning based on the code after ... | [] | null | null |
sglang | d1ec93e3acc483d37e33b1378f2b2ff7efbb3cfc | 2026-01-10T20:55:31+08:00 | 16,397 | https://github.com/sgl-project/sglang/pull/16397 | Optimize layernorm_gated for Qwen3-Next (#16397) | Optimize layernorm_gated for Qwen3-Next (#16397)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com> | 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
@@ -6,13 +6,15 @@
# The models we train have hidden dim up to 8k anyway (e.g. Llama 70B), so this is fine.
+from functools import lru_cache
+
import torch
import torch.nn.functional as F
i... | 90 | 36 | 1 | [
"qwen"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/fla/layernorm_gated.py).
Goal: Optimize layernorm_gated for Qwen3-Next
Target hardware: unspecified
Target models: qwen
Current file:
```python
# Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/modules/layernor... | <!-- 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.io to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [{"scope": "kernel", "hardware": null, "model": null, "shape": "N=128", "batch_size": 65536, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 41.43, "optimized": 18.13, "stated_speedup": null, "speedup": 2.2852}, {"scope": "kernel", "hardware": null, "model": null, "shape": "N=128", "batch_size... | 2.3053 | 1.027 |
sglang | 67b61a4e8d0dba9c8c1d52a42769f658ad20bc0b | 2026-01-10T13:57:44+08:00 | 16,723 | https://github.com/sgl-project/sglang/pull/16723 | [Rework] Add SwapAB Optimization for triton fused_moe_kernel on SM90. (#16723) | [Rework] Add SwapAB Optimization for triton fused_moe_kernel on SM90. (#16723) | 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 functools
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 ... | --- 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
@@ -1,5 +1,6 @@
from __future__ import annotations
+import functools
import os
from typing import Any, Dict, List, Optional
@@ -20,9 +21,11 @@
from sglang.s... | 46 | 1 | 1 | [] | [
"SM90"
] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py).
Goal: [Rework] Add SwapAB Optimization for triton fused_moe_kernel on SM90.
Target hardware: SM90
Target models: unspecified
Current file:
```python
from __future__ import annotations
import os
f... | <!-- 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.io to discuss further. -->
## Motivation
Rework of reverted pr https://github.com/sgl-p... | [] | null | null |
sglang | ee4d2287ab64a196adb316255eb768cdf826962a | 2026-01-07T23:45:35+08:00 | 15,712 | https://github.com/sgl-project/sglang/pull/15712 | Add SwapAB Optimization for triton fused_moe_kernel on SM90. (#15712) | Add SwapAB Optimization for triton fused_moe_kernel on SM90. (#15712) | 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 functools
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 ... | --- 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
@@ -1,5 +1,6 @@
from __future__ import annotations
+import functools
import os
from typing import Any, Dict, List, Optional
@@ -20,9 +21,11 @@
from sglang.s... | 40 | 1 | 1 | [] | [
"SM90"
] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py).
Goal: Add SwapAB Optimization for triton fused_moe_kernel on SM90.
Target hardware: SM90
Target models: unspecified
Current file:
```python
from __future__ import annotations
import os
from typin... | <!-- 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.io to discuss further. -->
## Motivation
<!-- Describe the purpose and goals of this p... | [{"scope": "kernel", "hardware": "H20-3e", "model": "GLM-4.6V-FP8", "shape": null, "batch_size": 1, "metric": "latency", "unit": "ms", "higher_is_better": false, "baseline": 0.058496, "optimized": 0.051936, "stated_speedup": 1.126309304, "speedup": 1.1263}, {"scope": "kernel", "hardware": "H20-3e", "model": "GLM-4.6V-F... | 1.1297 | 1.16 |
sglang | 9a414b164c5eb538a6b3677e5f3d9dcd27ee9f3d | 2026-01-03T15:15:36+08:00 | 15,640 | https://github.com/sgl-project/sglang/pull/15640 | [Performance] Optimze the performance of Qwen25VL (#15640) | [Performance] Optimze the performance of Qwen25VL (#15640)
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.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."""
from __future__ import annotations
import itertools
import math
from typing import Any, Dict, List, Optional, Tuple, Union
import torch
import to... | # 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."""
from __future__ import annotations
import itertools
import math
from typing import Any, Dict, List, Optional, Tuple, Union
import torch
import to... | --- python/sglang/srt/layers/rotary_embedding.py
+++ python/sglang/srt/layers/rotary_embedding.py
@@ -16,6 +16,7 @@
from sglang.srt.utils import (
cpu_has_amx_support,
get_bool_env_var,
+ get_compiler_backend,
is_cpu,
is_cuda,
is_hip,
@@ -2748,6 +2749,7 @@
return torch.cat((-x2, x1), d... | 2 | 0 | 1 | [
"qwen"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/rotary_embedding.py).
Goal: [Performance] Optimze the performance of Qwen25VL
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/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.io to discuss further. -->
## Motivation
To improve the performance of Qwen25VL model.... | [{"scope": "kernel", "hardware": null, "model": "Qwen25VL", "shape": null, "batch_size": null, "metric": "latency", "unit": "us", "higher_is_better": false, "baseline": 96, "optimized": 24, "stated_speedup": null, "speedup": 4.0}, {"scope": "kernel", "hardware": null, "model": "Qwen25VL", "shape": null, "batch_size": n... | 9.2778 | 1.2747 |
sglang | 9d878c1f3e1e5bd1ac7be9417e5b3cd6308cd23a | 2025-12-25T12:35:39-08:00 | 15,522 | https://github.com/sgl-project/sglang/pull/15522 | Optimize FP8 MLA KV cache writes with Triton kernel (#15522) | Optimize FP8 MLA KV cache writes with Triton kernel (#15522) | python/sglang/srt/layers/attention/nsa/quant_k_cache.py | triton | Apache-2.0 | import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.nsa.utils import NSA_QUANT_K_CACHE_FAST
def quantize_k_cache(cache_k):
# TODO upstream can skip concat([k_nope, k_pe]) since we split them here
if NSA_QUANT_K_CACHE_FAST:
return _quantize_k_cache_fast_wrapped(cac... | import torch
import triton
import triton.language as tl
from sglang.srt.layers.attention.nsa.utils import NSA_QUANT_K_CACHE_FAST
def quantize_k_cache(cache_k):
# TODO upstream can skip concat([k_nope, k_pe]) since we split them here
if NSA_QUANT_K_CACHE_FAST:
return _quantize_k_cache_fast_wrapped(cac... | --- python/sglang/srt/layers/attention/nsa/quant_k_cache.py
+++ python/sglang/srt/layers/attention/nsa/quant_k_cache.py
@@ -11,6 +11,84 @@
return _quantize_k_cache_fast_wrapped(cache_k)
else:
return _quantize_k_cache_slow(cache_k)
+
+
+def quantize_k_cache_separate(
+ k_nope: torch.Tensor,
+ ... | 199 | 1 | 1 | [] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/attention/nsa/quant_k_cache.py).
Goal: Optimize FP8 MLA KV cache writes with Triton kernel
Target hardware: unspecified
Target models: unspecified
Current file:
```python
import torch
import triton
import triton.language as tl
from sglang.srt.lay... | # Updated Summary
Optimize FP8 MLA KV cache writes in NSA decode by:
- skipping `torch.cat([k_nope, k_rope])` in the FP8 path, and
- reusing the existing `set_mla_kv_buffer_triton` two-tensor write kernel.
**No new Triton write kernel is added in the final version.**
## Motivation
In DeepSeek-V3.2 dec... | [{"scope": "e2e", "hardware": null, "model": "DeepSeek-V3.2", "shape": null, "batch_size": null, "metric": "Benchmark duration", "unit": "s", "higher_is_better": false, "baseline": 153.09, "optimized": 150.03, "stated_speedup": null, "speedup": 1.0204}, {"scope": "e2e", "hardware": null, "model": "DeepSeek-V3.2", "shap... | null | 1.0205 |
sglang | 793c96c3d269086249eb279672fd0475ccab1b82 | 2025-12-18T18:13:22+08:00 | 12,921 | https://github.com/sgl-project/sglang/pull/12921 | [perf]optimize w4afp8 kernel on deepseek-v3-0324 (#12921) | [perf]optimize w4afp8 kernel on deepseek-v3-0324 (#12921)
Signed-off-by: bruce.xu <bruce.x@gmicloud.ai> | sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_grouped_mm_c3x.cu | cuda | Apache-2.0 | #include <c10/cuda/CUDAGuard.h>
#include <cudaTypedefs.h>
#include <torch/all.h>
#include <type_traits>
#include "cutlass/cutlass.h"
#include "w4a8_grouped_mm_c3x.cuh"
using namespace cute;
namespace {
enum class Sched { PP, CO };
template <int M, int N, int K, int A, int B, int C, Sched S>
struct SM90W4A8Config ... | #include <c10/cuda/CUDAGuard.h>
#include <cudaTypedefs.h>
#include <torch/all.h>
#include <type_traits>
#include "cutlass/cutlass.h"
#include "w4a8_grouped_mm_c3x.cuh"
using namespace cute;
namespace {
enum class Sched { PP, CO };
template <int M, int N, int K, int A, int B, int C, Sched S>
struct SM90W4A8Config ... | --- sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_grouped_mm_c3x.cu
+++ sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_grouped_mm_c3x.cu
@@ -66,6 +66,25 @@
chunk_size);
}
+// Helper macro to reduce code duplication
+// Note: Config must be wrapped in parentheses when it contains commas (e.g., template parameters)
+// ... | 69 | 241 | 3 | [
"deepseek"
] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_grouped_mm_c3x.cu).
Goal: [perf]optimize w4afp8 kernel on deepseek-v3-0324
Target hardware: unspecified
Target models: deepseek
Current file:
```cuda
#include <c10/cuda/CUDAGuard.h>
#include <cudaTypedefs.h>
#include <torch/all.h>
... | ## Motivation
we use w4afp8 deepseekv3-0324 online, and we find its performance is not good enough when decode batch size < 32
## Modifications
fine-grained tiling config
and based on https://github.com/sgl-project/sglang/pull/10027/files
I use cuda-int4 memory access to decrease memory-access pressure
## Accuracy ... | [{"scope": "e2e", "hardware": null, "model": "DeepSeek-V3-0324", "shape": null, "batch_size": 20, "metric": "request throughput", "unit": "req/s", "higher_is_better": true, "baseline": 0.69, "optimized": 0.71, "stated_speedup": null, "speedup": 1.029}, {"scope": "e2e", "hardware": null, "model": "DeepSeek-V3-0324", "sh... | 1.1165 | 1.072 |
sglang | 793c96c3d269086249eb279672fd0475ccab1b82 | 2025-12-18T18:13:22+08:00 | 12,921 | https://github.com/sgl-project/sglang/pull/12921 | [perf]optimize w4afp8 kernel on deepseek-v3-0324 (#12921) | [perf]optimize w4afp8 kernel on deepseek-v3-0324 (#12921)
Signed-off-by: bruce.xu <bruce.x@gmicloud.ai> | sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_moe_data.cu | cuda | Apache-2.0 | #include <c10/cuda/CUDAGuard.h>
#include <cudaTypedefs.h>
#include <torch/all.h>
#include <iostream>
constexpr uint64_t THREADS_PER_EXPERT = 512;
__global__ void compute_problem_sizes_w4a8(
const int32_t* __restrict__ topk_ids,
int32_t* problem_sizes1,
int32_t* problem_sizes2,
int32_t* atomic_buffer,... | #include <c10/cuda/CUDAGuard.h>
#include <cudaTypedefs.h>
#include <torch/all.h>
#include <cub/block/block_reduce.cuh>
#include <cub/block/block_scan.cuh>
template <int BLOCK_SIZE>
__global__ void compute_problem_sizes_w4a8(
const int32_t* __restrict__ topk_ids,
int32_t* problem_sizes1,
int32_t* problem_s... | --- sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_moe_data.cu
+++ sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_moe_data.cu
@@ -2,29 +2,48 @@
#include <cudaTypedefs.h>
#include <torch/all.h>
-#include <iostream>
+#include <cub/block/block_reduce.cuh>
+#include <cub/block/block_scan.cuh>
-constexpr uint64_t THREADS_PER_EX... | 96 | 27 | 3 | [
"deepseek"
] | [] | [] | Optimize this cuda kernel file from sglang (sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_moe_data.cu).
Goal: [perf]optimize w4afp8 kernel on deepseek-v3-0324
Target hardware: unspecified
Target models: deepseek
Current file:
```cuda
#include <c10/cuda/CUDAGuard.h>
#include <cudaTypedefs.h>
#include <torch/all.h>
#inclu... | ## Motivation
we use w4afp8 deepseekv3-0324 online, and we find its performance is not good enough when decode batch size < 32
## Modifications
fine-grained tiling config
and based on https://github.com/sgl-project/sglang/pull/10027/files
I use cuda-int4 memory access to decrease memory-access pressure
## Accuracy ... | [{"scope": "e2e", "hardware": null, "model": "DeepSeek-V3-0324", "shape": null, "batch_size": 20, "metric": "request throughput", "unit": "req/s", "higher_is_better": true, "baseline": 0.69, "optimized": 0.71, "stated_speedup": null, "speedup": 1.029}, {"scope": "e2e", "hardware": null, "model": "DeepSeek-V3-0324", "sh... | 1.1165 | 1.072 |
sglang | d36299ad774fe139b2b9afc85a97d9115e6a3067 | 2025-12-13T17:39:55+08:00 | 14,423 | https://github.com/sgl-project/sglang/pull/14423 | [NPU] perf update with kvcache nz & w4a8 quant (#14423) | [NPU] perf update with kvcache nz & w4a8 quant (#14423) | 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
@@ -806,6 +806,10 @@
/ yarn_get_mscale(self.scaling_factor, float(mscale_all_dim))
* attn_factor
)
+ self.cos_cached_total = None
+ self.sin_cached_total = None
+ sel... | 38 | 8 | 1 | [] | [
"NPU"
] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/rotary_embedding.py).
Goal: [NPU] perf update with kvcache nz & w4a8 quant
Target hardware: NPU
Target models: unspecified
Current file:
```python
# Adapted from https://raw.githubusercontent.com/vllm-project/vllm/refs/tags/v0.6.6.post1/vllm/model... | <!-- 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、Use the nz format for kv cache, thie method... | [{"scope": "e2e", "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}] | null | 1.1 |
sglang | 03b835e7d13e4564e2beb72845486b8dc37606b9 | 2025-12-08T09:24:58+08:00 | 14,141 | https://github.com/sgl-project/sglang/pull/14141 | Refactor tuning block wise kernel and opt Qwen/Qwen3-VL-32B-Instruct-FP8 (#14141) | Refactor tuning block wise kernel and opt Qwen/Qwen3-VL-32B-Instruct-FP8 (#14141) | 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
@@ -719,6 +719,7 @@
BLOCK_SIZE_N: tl.constexpr,
BLOCK_SIZE_K: tl.constexpr,
GROUP_SIZE_M: tl.constexpr,
+ needs_masking: tl.constexpr,
):
"""Triton-accelerated function used to perform... | 94 | 67 | 1 | [
"qwen"
] | [] | [] | Optimize this triton kernel file from sglang (python/sglang/srt/layers/quantization/fp8_kernel.py).
Goal: Refactor tuning block wise kernel and opt Qwen/Qwen3-VL-32B-Instruct-FP8
Target hardware: unspecified
Target models: qwen
Current file:
```python
# Copyright 2024 SGLang Team
# 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
Add a needs_masking flag to optimize `w8a8_bl... | [{"scope": "e2e", "hardware": null, "model": "Qwen/Qwen3-VL-32B-Instruct-FP8", "shape": null, "batch_size": 1, "metric": "output_throughput", "unit": "token/s", "higher_is_better": true, "baseline": 52.033, "optimized": 56.681, "stated_speedup": null, "speedup": 1.0893}, {"scope": "e2e", "hardware": null, "model": "Qwe... | null | 1.0105 |
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/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
@@ -85,6 +85,26 @@
}
template <typename scalar_t>
+inline void copy_stub(float* __restrict__ out, const scalar_t* __restrict__ input, int64_t size) {
+ using bVec = at::vec::Vectorized<scalar_t>;
+ using fVec = at::vec::Vectorized<float>;
+ conste... | 262 | 11 | 3 | [
"qwen"
] | [
"CPU"
] | [] | Optimize this cpp kernel file from sglang (sgl-kernel/csrc/cpu/gemm.cpp).
Goal: [CPU] Optimize small oc GEMM for Qwen3-next on CPU
Target hardware: CPU
Target models: qwen
Current file:
```cpp
#include "gemm.h"
#include "common.h"
#include "vec.h"
namespace {
// packed layout:
// quants {N, K} int8_t
// co... | 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 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.