docs: add Hunyuan 3 Preview cookbook#23532
Merged
Qiaolin-Yu merged 2 commits intoApr 23, 2026
Merged
Conversation
- Section 1: MoE architecture (~276B / ~20B active), hybrid thinking
(reasoning_effort high/medium/low/none), 256K context, MTP
- Section 2: Docker image table (lmsysorg/sglang:hy3-preview{,-cu130})
- Section 3: Interactive Hunyuan3PreviewDeployment jsx generator
(NVIDIA A100/H100/H200/B200/B300/GB300, FP8/BF16,
Blackwell auto-injects --attention-backend trtllm_mha,
MTP toggle prepends SGLANG_ENABLE_SPEC_V2=1 and
--speculative-algorithm EAGLE flags)
- Section 4: Real invocation outputs (simple completion, thinking
high-effort, instant mode, non-stream tool call, streaming tool call)
- Section 5: GSM8K (95.0%), MMLU (82.5%), tool-call accuracy via
MiniMax-Provider-Verifier, low/high-concurrency bench_serving
results on 4x H200
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
- Remove Hy3-preview-FP8 from Available Models list - Remove FP8 Hardware Requirements block (Section 3.2) - JSX generator: drop A100/H100 hardware (BF16 won't fit single-node on 80GB GPUs) and drop the FP8 quantization option - All deploy commands switch from '--tp 4' to '--tp 8' (H200 BF16 default) - Docker table: narrow to H200/B200 and B300/GB300 - Section 5 benchmarks: replace FP8-Testing numbers with TODO placeholders for BF16 re-measure
Qiaolin-Yu
approved these changes
Apr 23, 2026
zhangying098
pushed a commit
to zhangying098/sglang
that referenced
this pull request
Apr 23, 2026
LucQueen
pushed a commit
to LucQueen/sglang
that referenced
this pull request
May 12, 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
2 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
Adds a cookbook entry for Tencent Hunyuan 3 Preview (
Hy3-preview/Hy3-preview-FP8/Hy3-preview-Base) underdocs_new/cookbook/autoregressive/Tencent/.cookbook/autoregressive/Tencent/Hunyuan3-Preview.mdx): 5-section Mintlify recipe covering model intro, install/Docker tags, deployment tips, invocation examples with real output (basic, hybrid thinkinghigh/none, non-stream + streaming tool call), and benchmarks.src/snippets/autoregressive/hunyuan3-preview-deployment.jsx): NVIDIA A100/H100/H200/B200/B300/GB300 × FP8/BF16 + hunyuan parsers + MTP toggle (prependsSGLANG_ENABLE_SPEC_V2=1); Blackwell hardware unconditionally adds--attention-backend trtllm_mha.docs.json): adds aTencentgroup under Autoregressive Models.Model characteristics surfaced in the doc
reasoning_effort, built-in MTP draft module<tool_call>/<arg_key>/<arg_value>— uses--tool-call-parser hunyuanand--reasoning-parser hunyuanBenchmarks included
Notes
HYV3ForCausalLM,hunyuantool-call / reasoning parsers, and the MTP draft loader are not yet upstream. This PR only adds documentation; it assumes the corresponding model-code / parser PRs land separately before theHy3-previewweights are public.lmsysorg/sglang:hy3-preview{,-cu130}) are placeholders for the release-specific image naming.Test plan
python3 -c "import json; json.load(open('docs_new/docs.json'))"— nav JSON parsesmint devpreview ondocs_new/to visually verify the interactive generator and table rendering