docs(cookbook): add Qwen3.6-27B dense variant#23486
Merged
mickqian merged 1 commit intosgl-project:mainfrom Apr 22, 2026
Merged
docs(cookbook): add Qwen3.6-27B dense variant#23486mickqian merged 1 commit intosgl-project:mainfrom
mickqian merged 1 commit intosgl-project:mainfrom
Conversation
Qwen3.6 ships a 27B dense variant alongside the existing 35B-A3B MoE. Update the cookbook page and deployment snippet: - Rewrite intro / available-models / hardware-requirements sections to cover both variants. - Add a Model Size radio (35B-A3B MoE / 27B Dense) to the deployment snippet; model path is derived from the selected baseName. - Drop the stale 'sglang[all]' install hint in favor of plain 'uv pip install sglang' (matches the install doc; '[all]' pulls diffusion/tracing/http2, irrelevant for an autoregressive VLM). - Note in Section 4 that 27B swaps only '--model-path'; other flags carry over. Verified with the sanity + MMMU runs against the FP8 fix from sgl-project#23467 on H200 TP=2: BF16 MMMU 900/val: 55.1% FP8 MMMU 900/val: 53.0% (within Wilson 95% CI, i.e. parity)
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
Contributor
|
The MMMU result is very different from the official result (82.9). Is this test performed on the sglang main branch? |
zhangying098
pushed a commit
to zhangying098/sglang
that referenced
this pull request
Apr 23, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Qwen3.6 ships a 27B dense variant (
Qwen3.6-27B/Qwen3.6-27B-FP8) alongside the existing 35B-A3B MoE. Update the cookbook page and deployment snippet to cover both.Model Sizeradio (35B-A3B MoE / 27B Dense) to the deployment snippet;--model-pathis derived from the selectedbaseName, same flags carry oversglang[all]install hint in favor of plainuv pip install sglang(matches the install doc;[all]only pulls diffusion/tracing/http2, irrelevant for an autoregressive VLM)--model-pathTest plan
Verified on H200 TP=2 on top of #23467 (FP8 MLP shard-matching fix) via
benchmark/mmmu/bench_sglang.pywith--reasoning-parser qwen3-thinking --tool-call-parser qwen25, concurrency 48:The 2.1% gap is within the Wilson 95% CI on n=900 (±3.2%) — FP8 at BF16 parity, i.e. no accuracy regression from the FP8 fix that landed in #23467.
mint devpreview renders the updated Qwen3.6 page (both variants in Available Models / Hardware tables; Model Size radio toggles the command)@babel/parserparses the updatedqwen36-deployment.jsx