Repository navigation

Laya on Apple Silicon: worker, benchmarks and recipe - #30

Merged

hsliuustc0106 merged 22 commits into

ThinkFlowLab:mainfrom

cacheline999:laya-apple-silicon

Oct 3, 2026

Commits on Oct 3, 2026

Serve Laya on Apple Silicon: worker, benchmarks and recipe
cacheline999
committed
c92919b
Add measured Laya reports and a paired frontend-overhead probe
cacheline999
committed
a2e3609
Warm every loaded Laya model and harden the HTTP benchmark
cacheline999
committed
630d13b
Move the Laya benchmark scripts under recipe/laya/bench
cacheline999
committed
3bf7c3b
Add LAYA_WORKER_WEIGHTS=fp16 for the compiled worker
cacheline999
committed
bc42882
Compile the encoder for multi-question batches; make compile on/off
cacheline999
committed
836f5ac
Recipe: M5 results, test dependencies, first-request note
cacheline999
committed
2af396f
Say that fp16 weights are for MPS and warn on CPU
cacheline999
committed
37f22e0
Read /health from the live model; correct the Apple Silicon recipe
cacheline999
committed
a07dcfd
Run plain fp32 laya after a fallback to the CPU
cacheline999
committed
4785a39
Split the GPU optimizations out of worker.py; correct the Laya README
cacheline999
committed
a5be4f8
Lay the Laya worker out like the Cua-S1 worker (#19)
cacheline999
committed
5f43fc4
Laya worker: keep measurements out of comments, report compile.active, honour --log-level
cacheline999
committed
f883827
Bench scripts: mark the imports that follow the sys.path setup (E402)
cacheline999
committed
351b8ec
Laya worker: prepare checkpoints loaded while serving; fix report parity on paired files and bundled revisions
cacheline999
committed
235774a
Laya worker: restore checkpoints evicted for a failed late load; say which laya-serve variables apply
cacheline999
committed
02c68c5
Laya recipe: say what a late checkpoint load costs through the frontend; bench spawn fixes
cacheline999
committed
9a8ee98
Laya worker: drop comments that repeat the code or the test name
cacheline999
committed
64c965f
Laya: third review fixes, requirement-derived tests on CPU and GPU, and what the warm numbers leave out
cacheline999
committed
8d15710
Laya: keep profile runs out of the parity section; header test off macOS; memory growth is from compile
cacheline999
committed
fc0a526
Laya bench and tests: two-sided answer check in paired summaries; readiness test checks the warmup
cacheline999
committed
ec48a04
Laya: move the benchmark scripts from recipe/laya/bench to benchmarks/laya_mps
cacheline999
committed
539f06d

Navigation Menu

Repository navigation

Laya on Apple Silicon: worker, benchmarks and recipe - #30

Laya on Apple Silicon: worker, benchmarks and recipe#30
hsliuustc0106 merged 22 commits into
ThinkFlowLab:mainfrom
cacheline999:laya-apple-silicon

Commits on Oct 3, 2026

Serve Laya on Apple Silicon: worker, benchmarks and recipe

Add measured Laya reports and a paired frontend-overhead probe

Warm every loaded Laya model and harden the HTTP benchmark

Move the Laya benchmark scripts under recipe/laya/bench

Add LAYA_WORKER_WEIGHTS=fp16 for the compiled worker

Compile the encoder for multi-question batches; make compile on/off

Recipe: M5 results, test dependencies, first-request note

Say that fp16 weights are for MPS and warn on CPU

Read /health from the live model; correct the Apple Silicon recipe

Run plain fp32 laya after a fallback to the CPU

Split the GPU optimizations out of worker.py; correct the Laya README

Lay the Laya worker out like the Cua-S1 worker (#19)

Laya worker: keep measurements out of comments, report compile.active, honour --log-level

Bench scripts: mark the imports that follow the sys.path setup (E402)

Laya worker: prepare checkpoints loaded while serving; fix report parity on paired files and bundled revisions

Laya worker: restore checkpoints evicted for a failed late load; say which laya-serve variables apply

Laya recipe: say what a late checkpoint load costs through the frontend; bench spawn fixes

Laya worker: drop comments that repeat the code or the test name

Laya: third review fixes, requirement-derived tests on CPU and GPU, and what the warm numbers leave out

Laya: keep profile runs out of the parity section; header test off macOS; memory growth is from compile

Laya bench and tests: two-sided answer check in paired summaries; readiness test checks the warmup

Laya: move the benchmark scripts from recipe/laya/bench to benchmarks/laya_mps