Model Pr History Knowledge

Model PR History Knowledge

This is a PR-driven knowledge base for model optimization history. It is not a set of per-model skills. Each model family keeps bilingual docs with inspected PR diffs, implementation file coverage, timelines, changed files, code excerpts, and validation/risk notes.

Use it before patching model-specific serving paths, choosing an SGLang SOTA optimization target, or explaining why a framework already has a faster path.

Query

Run commands from this directory:

python3 scripts/query.py --list
python3 scripts/query.py --framework sglang --model qwen3-core --paths-only
python3 scripts/query.py --framework sglang --model qwen3-core "fused qk norm rope"
python3 scripts/query.py --framework vllm "DeepSeek-V4 fused norm router" --limit 5
python3 scripts/query.py --framework tokenspeed qwen35 --paths-only

Useful options:

• --framework sglang|vllm|tensorrt_llm|tokenspeed: restrict to one serving framework.
• --model <slug>: restrict to one model family directory.
• --lang en|zh|both: select English, Chinese, or both docs.
• --paths-only: print the exact docs to read without snippets.
• --limit N: bound search results.

Workflow

1. Infer the model-family slug from the user's model id, checkpoint path, or SGLang source path. If unsure, run scripts/query.py "<model name>".
2. Read the matching SGLang history first for SGLang patch work. Read competitor history too when vLLM, TensorRT-LLM, or TokenSpeed is the leading competitor or its trace suggests a missing SGLang fast path. If the doc opens with a dated PR Backfill Audit section, read it first: it lists the most recent PR-numbered merges that are not yet folded into the older timeline / diff-audit cards.
3. Extract only actionable evidence:
   • model implementation files and symbols
   • PRs that changed the hot source path
   • prior fusions, overlap work, quantization, MoE, attention, cache, sampler, or loader changes
   • open/watch PRs that may explain a known gap or pending support issue
   • validation lanes and regression risks implied by the PR cards
4. Save a short note in the active run artifacts, for example history/model-pr-history-notes.md, with paths read, PR numbers, source files, and the decision each item influenced.
5. Do not copy long PR cards into the final answer. Cite paths and summarize the relevant implementation/risk.

Model Slugs

Current frameworks:

• sglang
• vllm
• tensorrt_llm
• tokenspeed

Current model-family slugs include:

deepseek-ocr, deepseek-ocr-2, deepseek-v3-r1, deepseek-v31, deepseek-v32,
deepseek-v4, ernie45, gemma4, glm-vlm-ocr, glm45, glm46-glm47, glm5-glm51,
gpt-oss, hunyuan3-preview, intern-s1, internvl35, jina-reranker-m0, kimi, ling25, llada21,
llama31, llama33-70b, llama4, mimo-v2-flash, minimax, mistral-small-4,
mixtral-quark-int4fp8-moe, moss-vl, nemotron-super, qwen-vlm-omni-asr,
qwen3-coder, qwen3-core, qwen3-next, qwen35, qwen36, qwen38, ring25, step35

Availability is framework-specific. In particular, the current SGLang index includes hunyuan3-preview, moss-vl, qwen36, and qwen38; the current vLLM index includes hunyuan3-preview and qwen36, but not moss-vl or a dedicated qwen38 surface. vLLM v0.27 lists Qwen/Qwen3.8-27B in the model registry, but it rides the existing qwen3_5 implementation (model_type=qwen3_5 in the public HF config) rather than a separate file tree. Query qwen35 on vLLM when comparing that checkpoint.

SOTA Loop Contract

For sglang-sota-humanize-loop, this knowledge base is an early context source:

• Read it after model identification and before patch planning.
• Include the history paths and key PR evidence in analysis/root-cause.md or history/model-pr-history-notes.md.
• If the profiler points at a known model path, check whether the history has prior changes on that file before writing a new patch.
• If a competitor is faster, search that competitor's model history for the same model family and stage before assuming the gap is kernel-local. Refresh live source/PRs for the exact target commit before patch planning when the comparison depends on latest upstream behavior.