scalar-quantizers
Operate FSQ, FSP, ResidualFSQ, and GroupedResidualFSQ scalar quantizers in vector-quantize-pytorch.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- scalar-quantizers
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Operate FSQ, FSP, ResidualFSQ, and GroupedResidualFSQ scalar quantizers in vector-quantize-pytorch.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/api-reference.md10.9 KB
- 参考文档references/troubleshooting.md9.0 KB
- 参考文档references/workflows.md7.6 KB
- 脚本scripts/smoke_scalar_quantizers.py10.4 KB
03 SKILL.md 正文
Scalar Quantizers
Use this sub-skill when a task needs scalar discretization in vector-quantize-pytorch: finite scalar quantization, finite scalar perturbation, residual scalar quantizers, grouped residual scalar quantizers, index roundtrips, tensor-layout decisions, or scalar-quantizer troubleshooting.
Route here for
FSQsetup withlevels,dim,num_codebooks,return_indices,channel_first,preserve_symmetry,bound_hard_clamp,orthogonal_rotation, projections, and index conversion.FSPsetup withlevels,dim,channel_first,act_name,quantize_rate,need_inv_act,vector_norm, four-value return tuples, stochastic training perturbation, and eval-time roundtrips.ResidualFSQandGroupedResidualFSQfor residual stacks, grouped feature splits,num_quantizers, residual reconstruction from indices,return_all_codes, and quantize-dropout caveats.- Debugging shape, dtype, projection, activation, norm, and exact-vs-allclose reconstruction issues in scalar quantizers.
Route elsewhere
- Use the top-level skill to choose package installation and overall routing.
- Use
../vector-quantization/for learned codebookVectorQuantize,RandomProjectionQuantizer, codebook EMA, masks, and top-k/manual EMA updates. - Use
../residual-quantizers/forResidualVQandGroupedResidualVQlearned-codebook residual workflows. - Use
../lookup-free-and-latent/forLFQ,ResidualLFQ,GroupedResidualLFQ,LatentQuantize, binary entropy losses, and lookup-free/latent quantization. - Use
../sim-and-hierarchical/forSimVQ,ResidualSimVQ, andHierarchicalVQ.
Required references
Read these bundled references before implementing or debugging a scalar workflow:
references/api-reference.mdfor constructor signatures, return tuples, shape contracts, index helpers, projections, and dtype behavior.references/workflows.mdfor safe FSQ/FSP/residual recipes and code patterns.references/troubleshooting.mdfor common errors, unsupported settings, roundtrip caveats, and precision/layout fixes.
Safe smoke helper
Run the bundled helper when you need a quick CPU sanity check of the installed package and the scalar APIs:
python scripts/smoke_scalar_quantizers.py --help
python scripts/smoke_scalar_quantizers.py
The helper uses small random tensors only. It does not download datasets, train MNIST models, or depend on any source tree.
Operating cautions
- FSQ returns exactly two values:
(quantized, indices). Ifreturn_indices=False, the second value isNoneand index roundtrip helpers are unavailable for that instance. - FSP returns exactly four values:
(quantized, indices, norm_loss, other_info). - Prefer eval mode and
torch.allclosefor reconstruction checks that involve FSP, projections, autocast, or low precision. Exact equality is reliable only for the narrow deterministic cases called out in the API reference. - For image-like tensors, confirm whether the class expects channel-first handling. FSQ treats 4D+ inputs as channel-first feature maps; FSP requires
channel_first=Truefor NCHW-style inputs.
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).