chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

update usage of trtllm_fp8_per_tensor_scale_moe (#12569)

Browse Source
This commit is contained in:
b8zhong
2025-11-03 14:25:32 -08:00
committed by GitHub
parent 8834260739
commit d31d48b341
2 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/advanced_features/attention_backend.md
View File

@@ -21,7 +21,7 @@ The support matrix is split into two parts: MHA (standard attention) and MLA (mu
| **Triton** | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ |
| **Torch Native (SDPA)** | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| **FlexAttention (PyTorch)** | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
| **TRTLLM MHA** | 16, 32 or 64 | ✅ | ✅ | ❌ | ❌ | ❌ |
| **TRTLLM MHA** | 16, 32 or 64 | ✅ | ✅ | ❌ | ✅ | ❌ |
| **Dual Chunk FlashAttention** | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
| **AITER (ROCm)** | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ |
| **Wave (ROCm)** | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |

2
python/sglang/srt/layers/quantization/modelopt_quant.py
View File

@@ -689,7 +689,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
else 1.0
),
use_routing_scales_on_input=use_routing_scales_on_input,
tile_tokens_dim=8, # TODO(brayden): use the FI tile calculation
tile_tokens_dim=None,
routing_method_type=routing_method_type,
)
sm.tag(output)