minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

Feat/add w4a16 moe support to nemotron (#25655)

Browse Source
This commit is contained in:
Shaun Kotek
2026-06-02 22:42:26 -07:00
committed by GitHub
parent 512bfbb1e1
commit b8d7351a74
19 changed files with 999 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs_new/docs/advanced_features/quantization.mdx
+15 -5
View File
@@ -142,10 +142,10 @@ The following table summarizes quantization method support across NVIDIA and AMD
</tr>
<tr>
<td><code>modelopt_fp4</code></td>
<td>Yes (Blackwell/SM100+)</td>
<td>Yes (SM80-SM90 via Marlin; SM100+ native FP4)</td>
<td>No</td>
<td>No</td>
<td><a href="https://github.com/NVIDIA/Model-Optimizer">NVIDIA ModelOpt</a>; native FP4 on Blackwell (B200, GB200)</td>
<td><a href="https://github.com/NVIDIA/Model-Optimizer">NVIDIA ModelOpt</a>; use Marlin W4A16 fallback on Ampere/Hopper and native FP4 backends on Blackwell</td>
</tr>
<tr>
<td><code>petit_nvfp4</code></td>
@@ -257,8 +257,8 @@ Backend selection is supported only for **blockwise FP8** and **NVFP4** GEMM. Wh
<tbody>
<tr>
<td><code>auto</code></td>
<td>SM100/120</td>
<td>Auto-selects: <code>flashinfer_cudnn</code> on SM120; <code>flashinfer_cutlass</code> on SM100</td>
<td>SM80+</td>
<td>Auto-selects: <code>flashinfer_cudnn</code> on SM120; <code>flashinfer_cutedsl</code> on SM100; <code>marlin</code> on SM80-SM90; <code>flashinfer_cutlass</code> otherwise</td>
</tr>
<tr>
<td><code>cutlass</code></td>
@@ -275,15 +275,25 @@ Backend selection is supported only for **blockwise FP8** and **NVFP4** GEMM. Wh
<td>SM100/120 (CUDA 13+, cuDNN 9.15+)</td>
<td>FlashInfer cuDNN backend; used on SM120 for performance</td>
</tr>
<tr>
<td><code>flashinfer_cutedsl</code></td>
<td>SM100</td>
<td>FlashInfer CuTe DSL backend</td>
</tr>
<tr>
<td><code>flashinfer_trtllm</code></td>
<td>SM100</td>
<td>FlashInfer TensorRT-LLM backend</td>
</tr>
<tr>
<td><code>marlin</code></td>
<td>SM80-SM90</td>
<td>Weight-only W4A16 fallback for NVFP4 checkpoints</td>
</tr>
</tbody>
</table>
When FlashInfer is unavailable for NVFP4, the SGLang CUTLASS kernel is used as an automatic fallback.
On Blackwell, when FlashInfer is unavailable for NVFP4, the SGLang CUTLASS kernel is used as an automatic fallback. On SM80-SM90, `auto` selects Marlin for NVFP4.
## Offline Quantization
docs_new/docs/advanced_features/server_arguments.mdx
+3 -3
View File
@@ -1220,9 +1220,9 @@ Please consult the documentation below and [server_args.py](https://github.com/s
</tr>
<tr>
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>`--fp4-gemm-backend`</td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Choose the runner backend for NVFP4 GEMM operations. Options: 'auto' (default; selects <code>flashinfer_cudnn</code> on SM120, <code>flashinfer_cutedsl</code> on SM100, <code>flashinfer_cutlass</code> otherwise), 'cutlass' (SGLang CUTLASS kernel), 'flashinfer_cutlass' (FlashInfer CUTLASS backend), 'flashinfer_cudnn' (FlashInfer cuDNN backend, optimal on CUDA 13+ with cuDNN 9.15+), 'flashinfer_cutedsl' (FlashInfer CuTe DSL backend), 'flashinfer_trtllm' (FlashInfer TensorRT-LLM backend, requires different weight preparation with shuffling). All FlashInfer backends fall back to sgl-kernel CUTLASS when FlashInfer is unavailable.</td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Choose the runner backend for NVFP4 GEMM operations. Options: 'auto' (default; selects <code>flashinfer_cudnn</code> on SM120, <code>flashinfer_cutedsl</code> on SM100, <code>marlin</code> on SM80-SM90, <code>flashinfer_cutlass</code> otherwise), 'cutlass' (SGLang CUTLASS kernel), 'flashinfer_cutlass' (FlashInfer CUTLASS backend), 'flashinfer_cudnn' (FlashInfer cuDNN backend, optimal on CUDA 13+ with cuDNN 9.15+), 'flashinfer_cutedsl' (FlashInfer CuTe DSL backend), 'flashinfer_trtllm' (FlashInfer TensorRT-LLM backend, requires different weight preparation with shuffling), 'marlin' (weight-only W4A16 fallback for SM80-SM90). All FlashInfer backends fall back to sgl-kernel CUTLASS when FlashInfer is unavailable.</td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>auto</code></td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}><code>auto</code>, <code>cutlass</code>, <code>flashinfer_cudnn</code>, <code>flashinfer_cutedsl</code>, <code>flashinfer_cutlass</code>, <code>flashinfer_trtllm</code></td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}><code>auto</code>, <code>cutlass</code>, <code>flashinfer_cudnn</code>, <code>flashinfer_cutedsl</code>, <code>flashinfer_cutlass</code>, <code>flashinfer_trtllm</code>, <code>marlin</code></td>
</tr>
<tr>
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>`--disable-flashinfer-autotune`</td>
@@ -1452,7 +1452,7 @@ Please consult the documentation below and [server_args.py](https://github.com/s
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>` --moe-runner-backend`</td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Choose the runner backend for MoE.</td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}>` auto`</td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}><code>auto</code>, <code>deep_gemm</code>, <code>triton</code>, <code>triton_kernel</code>, <code>flashinfer_trtllm</code>, <code>flashinfer_trtllm_routed</code>, <code>flashinfer_cutlass</code>, <code>flashinfer_mxfp4</code>, <code>flashinfer_cutedsl</code>, <code>cutlass</code></td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}><code>auto</code>, <code>deep_gemm</code>, <code>triton</code>, <code>triton_kernel</code>, <code>flashinfer_trtllm</code>, <code>flashinfer_trtllm_routed</code>, <code>flashinfer_cutlass</code>, <code>flashinfer_mxfp4</code>, <code>flashinfer_cutedsl</code>, <code>cutlass</code>, <code>marlin</code></td>
</tr>
<tr>
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}>` --flashinfer-mxfp4-moe-precision`</td>