From d31d48b341dfa254814c75a6d6df2d68e4bc895b Mon Sep 17 00:00:00 2001 From: b8zhong Date: Mon, 3 Nov 2025 14:25:32 -0800 Subject: [PATCH] update usage of `trtllm_fp8_per_tensor_scale_moe` (#12569) --- docs/advanced_features/attention_backend.md | 2 +- python/sglang/srt/layers/quantization/modelopt_quant.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/docs/advanced_features/attention_backend.md b/docs/advanced_features/attention_backend.md index bbe39e0a1..3a7645305 100644 --- a/docs/advanced_features/attention_backend.md +++ b/docs/advanced_features/attention_backend.md @@ -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)** | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | diff --git a/python/sglang/srt/layers/quantization/modelopt_quant.py b/python/sglang/srt/layers/quantization/modelopt_quant.py index ab75206e9..841c3cb44 100755 --- a/python/sglang/srt/layers/quantization/modelopt_quant.py +++ b/python/sglang/srt/layers/quantization/modelopt_quant.py @@ -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)