quantized-modules-functions
Direct use of bitsandbytes quantized PyTorch modules, autograd matmul helpers, and functional quantization primitives for custom modules and research code.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- quantized-modules-functions
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Direct use of bitsandbytes quantized PyTorch modules, autograd matmul helpers, and functional quantization primitives for custom modules and research code.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/api-reference.md8.4 KB
- 参考文档references/troubleshooting.md6.3 KB
- 参考文档references/workflows.md6.3 KB
- 脚本scripts/quantized-module-smoke.py5.2 KB
03 SKILL.md 正文
Quantized Modules and Functional Primitives
Use this sub-skill when the task needs direct bitsandbytes.nn, bitsandbytes.functional, bitsandbytes.matmul, bitsandbytes.matmul_4bit, or torch.ops.bitsandbytes APIs rather than Hugging Face model-loading wrappers.
Choose this sub-skill for
- Replacing
torch.nn.Linearortorch.nn.EmbeddingwithLinear8bitLt,Linear4bit,LinearNF4,LinearFP4,Embedding8bit,Embedding4bit,EmbeddingNF4, orEmbeddingFP4in custom PyTorch code. - Working directly with
Int8Params,Params4bit,QuantState,matmul,matmul_4bit, int8 vectorwise quantization, 4-bit quantization, or blockwise quantization primitives. - Debugging direct module state dicts, quantized checkpoints,
SCB/CBfields,.to(device)quantization timing,quant_statehandling, blocksize rules, or quantized tensor storage dtype. - Creating CPU-safe construction/signature checks for agents or CI without requiring quantized GPU kernels.
Route elsewhere
- Hugging Face
BitsAndBytesConfig,load_in_8bit,load_in_4bit, QLoRA model loading, PEFT, Accelerate, Diffusers, and FSDP-QLoRA workflows route totransformers-integrations. - Package installation, backend compatibility, missing native library, CUDA/ROCm/XPU/HPU/MPS setup, and
python -m bitsandbytesfailures route toinstallation-diagnostics. Adam8bit,AdamW8bit, paged optimizers,GlobalOptimManager,StableEmbedding, and training-loop optimizer behavior route tooptimizers-training.
Start here
- Read
references/api-reference.mdfor verified constructor/function signatures and parameter notes. - Use
references/workflows.mdfor direct layer replacement, quantize-on-.to(device), save/load caveats, and primitive recipes. - Use
references/troubleshooting.mdto map common symptoms to recovery steps. - Run
scripts/quantized-module-smoke.pyfor a deterministic CPU-safe import/construction/signature check; pass--device cudaor another backend only when that backend is available and native quantized kernels are expected to work.
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).