inference-model-loading
Route Qwen model loading, local or cloud inference, batch generation, backend choice, precision, quantized checkpoints, and long-context planning.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- inference-model-loading
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Route Qwen model loading, local or cloud inference, batch generation, backend choice, precision, quantized checkpoints, and long-context planning.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/model-and-backend-matrix.md3.4 KB
- 参考文档references/model-loading-and-inference.md4.9 KB
- 参考文档references/troubleshooting.md2.0 KB
- 参考文档references/vllm-and-batch-inference.md2.4 KB
- 脚本scripts/qwen_inference_checklist.py3.4 KB
03 SKILL.md 正文
Qwen Inference and Model Loading
Use this sub-skill when the user wants to load a Qwen checkpoint, run chat or text generation, choose a model/backend, plan batch inference, diagnose local checkpoint loading, or decide between Transformers, ModelScope, DashScope, CPU, CUDA, quantized, and long-context paths.
First decisions
- Identify the model family: base
Qwen-*for continuation/generation, orQwen-*-Chatforchat/chat_streamand multi-turn assistant behavior. - Identify where the checkpoint comes from: Hugging Face model id, ModelScope id, DashScope hosted API, or an already-downloaded local directory.
- Identify runtime limits: CPU-only, single CUDA GPU, multiple GPUs, vLLM serving, no network, no credentials, or no checkpoint yet.
- Run the safe checklist before model loading when the user is debugging setup:
python scripts/qwen_inference_checklist.py --model Qwen/Qwen-7B-Chat --backend auto --precision auto
Routes
| User request | Read |
|---|---|
Basic Transformers or ModelScope loading, local checkpoint fallback, model.chat, model.generate, history, or no-network setup |
references/model-loading-and-inference.md |
| Model sizes, base vs chat choice, CPU/CUDA/DashScope/vLLM backend, precision, quantization, KV cache, or long-context tradeoffs | references/model-and-backend-matrix.md |
| Batch inference, ChatML stop words, vLLM transformer-like wrapper behavior, or when to route service launch elsewhere | references/vllm-and-batch-inference.md |
| Errors about missing tokenizer files, remote code, slow CPU, bad generations, optional dependencies, quantization, or memory | references/troubleshooting.md |
Boundaries
- For CLI, web UI, OpenAI-compatible API, Docker, vLLM/FastChat service launch, TensorRT, Ascend, or DCU deployment, use
../serving-deployment/SKILL.md. - For fine-tuning, LoRA, Q-LoRA, adapter merge, or creating GPTQ checkpoints, use
../finetuning-quantization/SKILL.md. - For ReAct prompts, function-calling schemas, ChatML internals, tokenizer special tokens, or BPE merge extension, use
../prompting-tool-use-tokenization/SKILL.md. - For benchmark reproduction, use
../evaluation-reproduction/SKILL.md.
Operating rules
- Keep
trust_remote_code=Truevisible for historical Qwen checkpoints. Remote code means code from the checkpoint is executed locally; use a trusted checkpoint source. - Do not load a remote model as a smoke test unless the user accepts network, storage, and model-code execution. A dependency import check is safer.
- Do not present CPU as a performance equivalent to CUDA; CPU-only loading is mainly for compatibility or very small/debug runs.
- Do not infer that quantized, vLLM, FlashAttention, or KV-cache behavior is verified from a base Transformers import.
- When the user has a local checkpoint, validate
config.json, tokenizer assets, and model shards before changing Python code.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).