master

openFuyao-botfeat(tokenizer): expose tokenizerSource and wire Helm charts

93ab1c05创建于 27 天前历史提交

文件	最后提交记录	最后更新时间
inflight	refactor: resolve codecheck findings across EPP plugins Address static-analysis findings without changing runtime behavior: - fix exported-name stutter (NPUState->State, PredictionScorer->Scorer, PrefixCacheProducer->Producer, score.ScoreAggregate->Aggregate, etc.) - extract magic values to named constants (TCP port, percent scale, file perms, float tolerance) - collapse long parameter lists into context/config structs - deduplicate JSON param decoding, prefix-cache scoring, and picker candidate prep into shared helpers - add exported doc comments and package comments Also ignore generated lockfiles and packaged chart deps, and add the Mulan license header to the sidecar pyproject files. go build ./... and go test ./pkg/epp/... pass. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	1 个月前
pdgroup	fix(epp): address remaining codecheck findings	1 个月前
predictionfeature	fix(prediction): treat KVCacheUsagePercent as a 0-1 fraction, not a percent vllm:kv_cache_usage_perc is stored raw as a [0,1] fraction, but the prediction feature extractor and the badness scorers divided it by 100 again. That compressed real 10-90% usage into ~0.001-0.009 and effectively removed KV utilization as a routing signal — in the prediction fallback and the kvaware/autokvaware/pdkvaware baselines alike, since they share score.AggregateBadness / PrefillBadness / DecodeBadness. Replace the /100 (and normalizePercent) on KVCacheUsagePercent with Clamp01 in the feature extractor and the score badness paths. NPU utilization fields keep /100 since the exporter reports 0-100. Add a regression test pinning the fraction semantics. Note: models must be retrained after this change — serving now feeds the unscaled fraction, so an old model would see a 100x feature-scale shift. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>	1 个月前
prefixcache	refactor(epp): reorganize shared plugin utilities	1 个月前
tokenizer	feat(tokenizer): expose tokenizerSource and wire Helm charts - Rename EPP tokenizer plugin JSON key to tokenizerSource for source/identity split - Render tokenizerSource in all routing profile Helm branches - Document the option in values and example profiles - Replace prediction sidecar noqa ANN001 with explicit type annotations Co-authored-by: lileqi<lileqi@huawei.com>	27 天前