module-computation
Compute evaluate metrics, comparisons, and measurements with direct or streaming inputs; combine modules; configure cache/distributed runs; save results; and debug input validation.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- module-computation
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Compute evaluate metrics, comparisons, and measurements with direct or streaming inputs; combine modules; configure cache/distributed runs; save results; and debug input validation.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/api-reference.md5.8 KB
- 参考文档references/computation-workflows.md5.8 KB
- 参考文档references/troubleshooting.md4.8 KB
- 脚本scripts/compute_smoke.py3.4 KB
03 SKILL.md 正文
Module Computation
Use this sub-skill when a user has already selected or loaded an evaluate module and needs to run it correctly. Route module discovery, Hub/local loading, revisions, and evaluate.load(...) decisions to ../module-loading/. Route evaluator pipeline orchestration around datasets/models/pipelines to ../evaluator-pipelines/. Route authoring or sharing new modules to ../hub-and-cli/.
Start Here
- Inspect the loaded module before computing: check
module.features,module.inputs_description,module.module_type, andmodule.infoso input names and shapes match the module contract. - Choose a computation style:
- Direct:
module.compute(predictions=[...], references=[...], **module_kwargs)for all inputs at once. - Streaming examples: call
module.add(prediction=..., reference=...)repeatedly, thenmodule.compute(). - Streaming batches: call
module.add_batch(predictions=[...], references=[...])for each batch, thenmodule.compute().
- Direct:
- For non-standard modules, use the input names shown in
features, such asinputs=andtargets=, instead of assumingpredictions/references. - For multiple modules, use
evaluate.combine(...); passforce_prefix=Trueor a name-to-module dict when output keys may collide. - For distributed runs, set the same
cache_dir,num_process, andexperiment_idon every worker, a uniqueprocess_idper worker, and read the result only fromprocess_id == 0. - Save finished result dictionaries with
evaluate.save(path_or_file, **result, **metadata)when the user needs a JSON record.
Core References
references/computation-workflows.md: direct compute, streaming, combined evaluations, distributed/cache patterns, result saving, and integration examples.references/api-reference.md: signatures, accepted inputs, result shapes, and constructor/settings table for computation APIs.references/troubleshooting.md: validation, shape, cache, distributed, casting, and combined-key failure modes.scripts/compute_smoke.py: safe smoke helper for tiny computations against a named module or local module path.
Fast Patterns
import evaluate
metric = evaluate.load("accuracy")
result = metric.compute(predictions=[0, 1, 1, 0], references=[0, 1, 0, 1])
metric = evaluate.load("accuracy")
for predictions, references in [([0, 1], [0, 1]), ([1, 0], [0, 1])]:
metric.add_batch(predictions=predictions, references=references)
result = metric.compute()
combined = evaluate.combine({"acc": "accuracy", "f1_binary": "f1"}, force_prefix=True)
result = combined.compute(predictions=[0, 1, 0], references=[0, 1, 1])
Guardrails
compute,add, andadd_batchare keyword-oriented forEvaluationModule; do not rely on positional calls for normal modules.featuresdescribes one example;add_batchandcomputeusually receive lists, arrays, or tensors of those examples.addtakes singular aliases (prediction,reference) for standard modules; use custom feature names directly for non-standard modules.compute()returnsNoneon non-zero distributed processes; only process0returns the final result dict.keep_in_memory=Trueis single-process only; do not combine it withnum_process > 1.evaluate.saverecords metadata including interpreter information in the JSON output; avoid publishing sensitive local paths from saved files without review.
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).