minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[FlashInfer v0.6.18] add FlashInfer CuTe DSL NVFP4 W4A16 mode (#35120)

Browse Source
This commit is contained in:
Ziang Li
2026-08-31 18:47:30 -07:00
committed by GitHub
parent 175973d834
commit 9a85473a89
12 changed files with 287 additions and 68 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/docs/references/environment_variables.mdx
+5
View File
@@ -759,6 +759,11 @@ SGLang supports various environment variables that can be used to configure its
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Enable FlashInfer TRT-LLM or CuTe DSL v2 (no A2A or FlashInfer A2A) per-token FP32 activation scaling for serialized <code>modelopt_fp4</code> checkpoints; checkpoint activation scales are treated as <code>1</code></td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>false</code></td>
</tr>
<tr>
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>SGLANG_FLASHINFER_CUTEDSL_NVFP4_W4A16</code></td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Use BF16 activations and outputs with FlashInfer CuTe DSL NVFP4 weights on NVIDIA SM100-family GPUs (SM100/SM103). Serialized dense linear layers use this mode with <code>--fp4-gemm-backend flashinfer_cutedsl</code>. MoE layers use it with <code>--moe-runner-backend flashinfer_cutedsl</code>; the MoE path supports online weight quantization and serialized ModelOpt NVFP4 weights with either no A2A or FlashInfer A2A, and honors <code>SGLANG_FLASHINFER_MOE_FUSED_FINALIZE</code>.</td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.02)"}}><code>false</code></td>
</tr>
<tr>
<td style={{padding: "9px 12px", fontWeight: 500, backgroundColor: "rgba(255,255,255,0.02)"}}><code>SGLANG_FLASHINFER_MOE_FUSED_FINALIZE</code></td>
<td style={{padding: "9px 12px", backgroundColor: "rgba(255,255,255,0.05)"}}>Use FlashInfer's fused atomic CUTLASS and CuTe DSL MoE finalize for best performance. Deterministic inference overrides this to <code>false</code>.</td>