EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Minor style and doc fix (#7228)

Browse Source
This commit is contained in:
Lianmin Zheng
2025-06-16 01:03:13 -07:00
committed by GitHub
parent 0ae1e9a755
commit 21615cc3fe
4 changed files with 10 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
docs/backend/attention_backend.md
View File

@@ -3,13 +3,16 @@
## Supporting matrix for different attention backends ## Supporting matrix for different attention backends
| **Backend** | **Page Size > 1** | **Spec Decoding** | **MLA** | **Sliding Window** | **MultiModal** | | **Backend** | **Page Size > 1** | **Spec Decoding** | **MLA** | **Sliding Window** | **MultiModal** |
|--------------------------|-------------------|-------------------|--------|--------------------|------------| |--------------------------|-------------------|-------------------|---------|--------------------|----------------|
| **FlashInfer** | ✅ | ✅ | ✅ | ✅ | ✅ | | **FlashInfer** | ❌ | ✅ | ✅ | ✅ | ✅ |
| **FA3** | ✅ | ✅ | ✅ | ✅ | ✅ | | **FA3** | ✅ | ✅ | ✅ | ✅ | ✅ |
| **Triton** | ❌ | ✅ | ✅ | ❌ | ❌ | | **Triton** | ❌ | ✅ | ✅ | ✅ | ❌ |
| **Torch Native** | ❌ | ❌ | ❌ | ❌ | ❌ | | **Torch Native** | ❌ | ❌ | ❌ | ❌ | ❌ |
| **FlashMLA** | ✅ | ✅ | ✅ | ❌ | ❌ | | **FlashMLA** | ✅ | ✅ | ✅ | ❌ | ❌ |
Note: Every kernel backend is compatible with a page size > 1 by specifying an argument such as `--page-size 16`.
This is because a page size of 16 can be converted to a page size of 1 in the kernel backend.
The "❌" and "✅" symbols in the table above under "Page Size > 1" indicate whether the kernel actually operates with a page size greater than 1, rather than treating a page size of 16 as a page size of 1.
## User guide ## User guide

3
python/sglang/srt/layers/attention/cutlass_mla_backend.py
View File

@@ -11,8 +11,6 @@ from typing import TYPE_CHECKING, Optional, Union
import torch import torch
import triton import triton
from sglang.global_config import global_config
from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
from sglang.srt.layers.attention.flashinfer_mla_backend import FlashInferMLAAttnBackend from sglang.srt.layers.attention.flashinfer_mla_backend import FlashInferMLAAttnBackend
from sglang.srt.layers.attention.utils import create_flashmla_kv_indices_triton from sglang.srt.layers.attention.utils import create_flashmla_kv_indices_triton
from sglang.srt.layers.dp_attention import get_attention_tp_size from sglang.srt.layers.dp_attention import get_attention_tp_size
@@ -22,7 +20,6 @@ from sglang.srt.utils import is_cuda
if TYPE_CHECKING: if TYPE_CHECKING:
from sglang.srt.layers.radix_attention import RadixAttention from sglang.srt.layers.radix_attention import RadixAttention
from sglang.srt.model_executor.model_runner import ModelRunner from sglang.srt.model_executor.model_runner import ModelRunner
from sglang.srt.speculative.eagle_utils import EagleDraftInput, EagleVerifyInput
from sglang.srt.speculative.spec_info import SpecInfo from sglang.srt.speculative.spec_info import SpecInfo
_is_cuda = is_cuda() _is_cuda = is_cuda()

1
python/sglang/srt/layers/attention/flashattention_backend.py
View File

@@ -11,7 +11,6 @@ from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
from sglang.srt.managers.schedule_batch import global_server_args_dict from sglang.srt.managers.schedule_batch import global_server_args_dict
from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
from sglang.srt.speculative.eagle_utils import EagleDraftInput, EagleVerifyInput from sglang.srt.speculative.eagle_utils import EagleDraftInput, EagleVerifyInput
from sglang.srt.utils import get_compiler_backend
if TYPE_CHECKING: if TYPE_CHECKING:
from sglang.srt.layers.radix_attention import RadixAttention from sglang.srt.layers.radix_attention import RadixAttention

5
python/sglang/srt/layers/attention/flashmla_backend.py
View File

@@ -14,8 +14,6 @@ import torch
import triton import triton
from flash_mla import flash_mla_with_kvcache, get_mla_metadata from flash_mla import flash_mla_with_kvcache, get_mla_metadata
from sglang.global_config import global_config
from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
from sglang.srt.layers.attention.flashinfer_mla_backend import FlashInferMLAAttnBackend from sglang.srt.layers.attention.flashinfer_mla_backend import FlashInferMLAAttnBackend
from sglang.srt.layers.attention.utils import create_flashmla_kv_indices_triton from sglang.srt.layers.attention.utils import create_flashmla_kv_indices_triton
from sglang.srt.layers.dp_attention import get_attention_tp_size from sglang.srt.layers.dp_attention import get_attention_tp_size
@@ -24,7 +22,6 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMo
if TYPE_CHECKING: if TYPE_CHECKING:
from sglang.srt.layers.radix_attention import RadixAttention from sglang.srt.layers.radix_attention import RadixAttention
from sglang.srt.model_executor.model_runner import ModelRunner from sglang.srt.model_executor.model_runner import ModelRunner
from sglang.srt.speculative.eagle_utils import EagleDraftInput, EagleVerifyInput
from sglang.srt.speculative.spec_info import SpecInfo from sglang.srt.speculative.spec_info import SpecInfo
@@ -330,7 +327,7 @@ class FlashMLABackend(FlashInferMLAAttnBackend):
) )
def get_cuda_graph_seq_len_fill_value(self): def get_cuda_graph_seq_len_fill_value(self):
return 1024 return 1
def forward_decode( def forward_decode(
self, self,