Thanks to visit codestin.com
Credit goes to github.com

Commits

Commits on Oct 23, 2025

Major cleanup of profiler zones, add Caller annotation for all pool.Run

jan-wassenberg
authored and
copybara-github
committed
3ed403e

Commits on Oct 22, 2025

Use correct offsets in BlobWriter.

niting
authored and
copybara-github
committed
9e8ac7e
Merge pull request #735 from Hitesh-ed:gemma.cpp-windows-build-fix
copybara-github
committed
64a82ed
Merge branch 'dev' into gemma.cpp-windows-build-fix
Hitesh-ed
authored
027288b

Commits on Oct 21, 2025

Warning fix (unused var), Windows build fix (missing member variable)

jan-wassenberg
authored and
copybara-github
committed
acede9d

Commits on Oct 16, 2025

Merge branch 'dev' into gemma.cpp-windows-build-fix
Hitesh-ed
authored
c55120f
Remove multi-package support from topology

jan-wassenberg
authored and
copybara-github
committed
f59eb2e
Update CMakePresets.json
Hitesh-ed
authored
cc1d256

Commits on Oct 15, 2025

gemma_batch_bench: generate more unique prompts

jan-wassenberg
authored and
copybara-github
committed
9b6ed1a
Add 8-bit integer quantization (I8Stream) to Gemma.cpp.

pculliton
authored and
copybara-github
committed
503aadd
Removed the PROFILER_ZONE from the most highly called functions to reduce the overhead.

theraysmith
authored and
copybara-github
committed
ee18916
Initialization of profiler zones.

theraysmith
authored and
copybara-github
committed
e3e8511

Commits on Oct 14, 2025

Added a global (to gemma) zones list to enable most call sites to PROFILER_ZONE3 to avoid the sychronization required for the static const initialization of the zone handle.

theraysmith
authored and
copybara-github
committed
fb6fa79

Commits on Oct 7, 2025

tune pool kSpin mode in threading_context

jan-wassenberg
authored and
copybara-github
committed
035273c

Commits on Oct 6, 2025

Add logging to io.cc on failed write and read.

niting
authored and
copybara-github
committed
9dc802c

Commits on Oct 2, 2025

Reduced parallelism for TransposeQ, making each thread read and write within its own cache lines

theraysmith
authored and
copybara-github
committed
684a044
Avoid transposing Q when it isn't needed

theraysmith
authored and
copybara-github
committed
1424466
Improve FlashAttention threading:

jan-wassenberg
authored and
copybara-github
committed
fe5a399

Commits on Oct 1, 2025

Increased parallelism for RMSNormAndPositionalEncoding

theraysmith
authored and
copybara-github
committed
6098a02

Commits on Sep 30, 2025

Added a smaller tile size to flash attention for smaller batch sizes

theraysmith
authored and
copybara-github
committed
2f6cbde
Fixed bug with softcap in single flash attention

theraysmith
authored and
copybara-github
committed
4974f24

Commits on Sep 29, 2025

Remove less useful spammy log lines.

niting
authored and
copybara-github
committed
1653699

Commits on Sep 26, 2025

Utilize multiple cores to read weight batches.

niting
authored and
copybara-github
committed
667a3f1

Commits on Sep 25, 2025

Used hn::BroadcastLane instead of Set(..., x.raw)

theraysmith
authored and
copybara-github
committed
d15731d

Commits on Sep 24, 2025

(1) Added QueryResultAndMetrics and BatchQueryModelWithMetrics to also return TimingInfo besides query results.

Charles Zhao
authored and
copybara-github
committed
4f0c633

Commits on Sep 22, 2025

Internal change

jan-wassenberg
authored and
copybara-github
committed
fac8aac

Commits on Sep 19, 2025

Remove no longer used MatVec

jan-wassenberg
authored and
copybara-github
committed
501fdf0

Commits on Sep 16, 2025

Fix batch inference: dangling reference

jan-wassenberg
authored and
copybara-github
committed
b603425

Commits on Sep 15, 2025

1.03x speedup: fused FFN

jan-wassenberg
authored and
copybara-github
committed
f3bc1c1

Commits on Sep 14, 2025

add const restriction for benchmark_helper.cc, and paligemma_helper.cc to remove a few uncessary copies.

Charles Zhao
authored and
copybara-github
committed
59db30e

Commits on Sep 12, 2025

Added zero-initialization to att_out.

theraysmith
authored and
copybara-github
committed
c9b8479

Commits on Sep 10, 2025

Temporarily disable flash pending msan fix

jan-wassenberg
authored and
copybara-github
committed
2695aab
Fix gemma_batch_bench for flash attention

jan-wassenberg
authored and
copybara-github
committed
ba61313
Refactor MatMul to accept views in the kernel functions

jan-wassenberg
authored and
copybara-github
committed
9457258

Commits on Sep 9, 2025

Added flash attention, with both a single-q function, and a register-tiled function.

theraysmith
authored and
copybara-github
committed
f10ac41