inference-workflow
Adapt mixtral-offloading's notebook evidence into safe local offloaded Mixtral inference workflows.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- inference-workflow
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Adapt mixtral-offloading's notebook evidence into safe local offloaded Mixtral inference workflows.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/configuration.md2.6 KB
- 参考文档references/troubleshooting.md2.6 KB
- 参考文档references/workflow.md4.5 KB
- 脚本scripts/create_offload_config.py2.3 KB
- 脚本scripts/render_generation_skeleton.py4.0 KB
03 SKILL.md 正文
Inference Workflow
Use this sub-skill when the task is to run, script, adapt, or troubleshoot the
Mixtral offloading demo workflow: building an offloaded MixtralForCausalLM
from quantized safetensors, choosing offload sizes, and generating chat text.
Read this when
- The user asks to turn the demo notebook into a Python script.
- The task mentions
OffloadConfig,QuantConfig,build_model,offload_per_layer, or a quantized Mixtral state directory. - The user needs to balance GPU VRAM, system RAM, and generation speed.
- A local run fails before or during model construction or generation.
Route map
- Read references/workflow.md for the full notebook-to-script recipe, validation gates, and generation-loop outline.
- Read references/configuration.md when the task is about offload sizing, HQQ quantization settings, or memory trade-offs.
- Use scripts/create_offload_config.py to
compute
main_size,offload_size, andbuffer_sizefrom layer/expert counts without importing the repository. - Use scripts/render_generation_skeleton.py
to render a safe starter script that shows where to call
build_modelwithout downloading weights or executing the model. - Read references/troubleshooting.md for state-path, CUDA, memory, HQQ config, tokenizer, and generation-cache errors.
What this sub-skill owns
- Runtime prerequisites specific to offloaded inference after the repository dependencies are installed.
- State directory checks for
model.safetensors.index.jsonand safetensors weight-map keys used bybuild_model. - The notebook's model setup pattern: tokenizer, quant configs,
OffloadConfig,build_model, streamer, sampling parameters, and cached generation. - Memory and speed trade-offs for
offload_per_layer.
What to route elsewhere
- Use the root ../../references/installation-and-runtime.md
for source-only installation,
PYTHONPATH, and package import checks. - Use ../quantization-kernels/SKILL.md for HQQ layer internals, packing functions, and Triton kernel errors.
- Use ../expert-cache/SKILL.md for LRU expert-cache, storage, and SparseMoeWrapper internals.
Minimal decision flow
- Confirm the user has a CUDA-capable PyTorch runtime for actual inference. CPU-only checks can validate imports and config math, but not the advertised offloaded generation path.
- Confirm the quantized state path is local and contains a Hugging Face safetensors index. Do not start a large network download unless the user explicitly asks.
- Compute
offload_per_layerand derived cache sizes from the model config. - Build the quantization configs exactly before calling
build_model. - Keep generation code explicit about
attention_maskandpast_key_valuesif continuing multi-turn chat.
Verification expectations
Safe checks are --help on the bundled scripts and dry-run config rendering.
Full Mixtral generation requires large external model artifacts and a CUDA GPU;
classify it as an approved native run only when the user accepts downloads,
memory use, and runtime cost.
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).