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

Verify the Ling-3.0-flash-VL FP4 lane on H200 and disable shared-expert fusion in quant recipes (#39419)

Browse Source
This commit is contained in:
Xinyuan Tong
2026-09-17 23:20:18 -07:00
committed by GitHub
parent 3ce3b4969f
commit 4dbba37965
3 changed files with 24 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/cookbook/autoregressive/InclusionAI/Ling-3.0-flash-VL.mdx
+2 -2
View File
@@ -95,8 +95,8 @@ It is a thinking model: the chat template turns chain-of-thought on by default a
- Native context is 128K, and the generated recipes default to the 256K YaRN variant (`--context-length 262144` plus the `rope_scaling` override and `SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1`), matching the model card. For the native 128K window, drop all three from the command.
- Audio content parts are rejected before preprocessing; send images as `image_url` and clips as `video_url` content parts (§3).
- The FP8 variant (`inclusionAI/Ling-3.0-flash-VL-FP8`) uses 128×128 block quantization, so plain TP must keep `768 ÷ tp` divisible by 128 (TP 1, 2, 3, or 6 — TP=4 fails to load). FP8 recipes default to a single GPU where the ~126 GB weights fit (`--tp 1` on 288 GB-class and B200, verified on GB300) or `--tp 2` on tighter cards. As a manual alternative on four GPUs, `--tp 4 --ep 4` keeps experts whole per rank (measured on 4×GB300). The BF16 checkpoint can also be quantized online with `--quantization fp8`; measured accuracy and speed for both paths are in the benchmark cards below.
- The INT4 variant (`inclusionAI/Ling-3.0-flash-VL-int4`) is a GPTQ 4-bit export (compressed-tensors, group size 32). It runs on all listed hardware with single-GPU (`--tp 1`) recipes; GB300, H200, and DGX Spark cells are verified, the rest carry the unverified badge.
- The FP4 variant (`inclusionAI/Ling-3.0-flash-VL-fp4`) is an MXFP4 export (e2m1 routed-expert weights with e8m0 scales, group 32, plus FP8 e4m3 dynamic activations) and runs on Blackwell only (GB300, B300, B200, DGX Spark). The `flashinfer_mxfp4` MoE backend is selected automatically on the `lmsysorg/sglang:dev-Ling-3.0-flash-VL` image and on sglang with sgl-project/sglang#38526; on builds that already include the VL integration but predate the auto-selection, add `--moe-runner-backend flashinfer_mxfp4` explicitly. GB300 and DGX Spark cells are verified, the rest carry the unverified badge.
- The INT4 variant (`inclusionAI/Ling-3.0-flash-VL-int4`) is a GPTQ 4-bit export (compressed-tensors, group size 32). It runs on all listed hardware with single-GPU (`--tp 1`) recipes; GB300, H200, and DGX Spark cells are verified, the rest carry the unverified badge. Shared-expert fusion is auto-disabled for the INT4 and FP4 checkpoints by the current dev image and by sglang main; older images need `--disable-shared-experts-fusion` explicitly.
- The FP4 variant (`inclusionAI/Ling-3.0-flash-VL-fp4`) is an MXFP4 export (e2m1 routed-expert weights with e8m0 scales, group 32, plus FP8 e4m3 dynamic activations) and runs on Hopper (SM90 CUTLASS W4A16) and Blackwell (SM100/SM120): H200, GB300, B300, B200, and DGX Spark. The `flashinfer_mxfp4` MoE backend is selected automatically on the `lmsysorg/sglang:dev-Ling-3.0-flash-VL` image and on sglang main; on builds that already include the VL integration but predate the auto-selection, add `--moe-runner-backend flashinfer_mxfp4` explicitly. GB300, H200, and DGX Spark cells are verified, the rest carry the unverified badge.
## 3. Advanced Usage