haraldwolff/llama.cpp
1
0
Fork 0
Code Issues Pull requests Packages Projects Releases Wiki Activity

llama : pad KV cache size (#4280)

Browse source 
* llama : pad KV cache size to 32

* metal : try to improve batched decoding
This commit is contained in:
Georgi Gerganov 2023-12-03 10:58:16 +02:00 committed by GitHub
parent 5a7d3125e7
commit d7b800b8bc
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 2 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
ggml-metal.m
View file

@ -1083,7 +1083,7 @@ void ggml_metal_graph_compute(
// find the break-even point where the matrix-matrix kernel becomes more efficient compared
// to the matrix-vector kernel
int ne11_mm_min = 1;
int ne11_mm_min = src0t == GGML_TYPE_F16 ? 1 : 16;
#if 0
// the numbers below are measured on M2 Ultra for 7B and 13B models

3
llama.cpp
View file

@ -5744,8 +5744,7 @@ static int llama_decode_internal(
// a heuristic, to avoid attending the full cache if it is not yet utilized
// after enough generations, the benefit from this heuristic disappears
// if we start defragmenting the cache, the benefit from this will be more important
//kv_self.n = std::max(32, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32)); // TODO: this might be better for CUDA?
kv_self.n = std::min((int32_t) cparams.n_ctx, std::max(32, llama_kv_cache_cell_max(kv_self)));
kv_self.n = std::min((int32_t) cparams.n_ctx, std::max(32, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32)));
//printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head);