Nandakishor 008deb86a4 Merge pull request #615 from RagingSilence/perf/reuse-question-tokenization ...

perf(agent): reuse question tokenization within prediction batches

# Conflicts:
#	.github/workflows/ci.yml
#	laya/agent.py

2026-09-28 00:54:22 +05:30

..

bench_apps.py

fix(research): point the benchmark scripts at paths that exist (#133)

2026-09-23 12:34:55 +05:30

bench_latency.py

fix(research): let the bench harnesses take their model root from LAYA_MODELS

2026-09-26 12:28:33 +02:00

bench_length_batching.py

feat(agent): group batch inputs by encoded length (#294)

2026-09-24 08:24:19 +05:30

bench_local.py

fix(research): let the bench harnesses take their model root from LAYA_MODELS

2026-09-26 12:28:33 +02:00

bench_long_context.py

release: 0.3.19, measured long-document table for laya-multilingual

2026-09-24 11:04:11 +05:30

bench_question_token_reuse.py

perf(agent): reuse question tokenization within prediction batches

2026-09-27 18:23:48 +08:00

bench_router_grouping.py

feat(router): group heterogeneous requests by checkpoint (#138)

2026-09-23 23:28:15 +05:30

build_benchmark_nb.py

fix: include zero confidence in ECE (#39)

2026-09-23 12:33:32 +05:30

confidence_definitions.py

research: recompute ECE under both definitions of confidence (#327)

2026-09-24 09:11:12 +05:30

laya_benchmark_colab.ipynb

research: benchmark harnesses and raw results (#14)

2026-09-20 23:06:53 +05:30

make_plots.py

fix(research): point the benchmark scripts at paths that exist (#133)

2026-09-23 12:34:55 +05:30

plot_long_context.py

release: 0.3.20, recommend max_len=8192 for long documents

2026-09-24 11:09:14 +05:30