gemma.cpp/gemma
Jan Wassenberg 2ebbe4076f 1.03-1.08x decode speedup: precompute Rope theta, fuse
Split attention into functions, move into class.
Fuse Rope and MulBy, allow non-in-place version to avoid copy from q to KV.
Sink if() into MaybeLogitsSoftCap.

PiperOrigin-RevId: 661168418
2024-08-09 01:23:24 -07:00
..
evals Add MMLU eval to github 2024-05-20 10:20:53 -07:00
instantiations Internal change 2024-07-30 20:24:54 -07:00
activations.h 1.03-1.08x decode speedup: precompute Rope theta, fuse 2024-08-09 01:23:24 -07:00
common.cc Internal change 2024-07-30 20:24:54 -07:00
common.h Internal change 2024-07-30 20:24:54 -07:00
configs.h Internal change 2024-07-30 20:24:54 -07:00
gemma-inl.h 1.03-1.08x decode speedup: precompute Rope theta, fuse 2024-08-09 01:23:24 -07:00
gemma.cc 1.1x prefill speedup, revamp threading in preparation for hierarchical parallelism. 2024-08-05 18:50:09 -07:00
gemma.h Improve performance logging 2024-08-07 14:15:43 -07:00
kv_cache.cc Major revamp #2 of Prefill: fix token order, parallel for multi-query 2024-07-25 03:28:55 -07:00
kv_cache.h Major revamp #2 of Prefill: fix token order, parallel for multi-query 2024-07-25 03:28:55 -07:00
run.cc Improve performance logging 2024-08-07 14:15:43 -07:00
tokenizer.cc 7x compile time speedup: shard gemma.cc 2024-07-03 06:35:04 -07:00
tokenizer.h 7x compile time speedup: shard gemma.cc 2024-07-03 06:35:04 -07:00
weights.cc Add scale parameter to MatMul. 2024-07-18 06:58:56 -07:00
weights.h Fix msan uninitialized scale in optimize_test 2024-07-22 10:50:25 -07:00