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

ggml : add ggml_soft_max_ext (#4256)

Browse source 
* metal : implement soft_max_ext

* cuda : implement soft_max_ext

* ggml : implement soft_max_ext (CPU)

* batched-bench : print threads

ggml-ci

* metal : simplify soft_max encoding

ggml-ci

* cuda : use 512 threads for soft_max instead of 32

* ggml : update soft max cpu

* cuda : do warp-based block reduce

* cuda : increase max block size to 1024

* cuda : fix warp reduction initialization of shared mem

* metal : warp-based reduction for soft max kernel

* metal : warp-based reduce for rms_norm

* metal : simplify soft max kernel

ggml-ci

* alloc : fix build with debug
This commit is contained in:
Georgi Gerganov 2023-12-01 10:51:24 +02:00 • committed by GitHub
parent 1d144112c0
commit ef47ec18da
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
8 changed files with 311 additions and 196 deletions
Download patch file Download diff file Expand all files Collapse all files

35
llama.cpp
View file

@ -3704,23 +3704,29 @@ static struct ggml_tensor * llm_build_kqv(
struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
cb(kq, "kq", il);
kq = ggml_scale(ctx, kq, kq_scale);
cb(kq, "kq_scaled", il);
if (max_alibi_bias > 0.0f) {
// TODO: n_head or n_head_kv
// TODO: K-shift is likely not working
// TODO: change to ggml_add
kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, max_alibi_bias);
cb(kq, "kq_scaled_alibi", il);
// temporary branch until we figure out how to handle ggml_alibi through ggml_add
kq = ggml_scale(ctx, kq, kq_scale);
cb(kq, "kq_scaled", il);
if (max_alibi_bias > 0.0f) {
// TODO: n_head or n_head_kv
// TODO: K-shift is likely not working
// TODO: change to ggml_add
kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, max_alibi_bias);
cb(kq, "kq_scaled_alibi", il);
}
kq = ggml_add(ctx, kq, kq_mask);
cb(kq, "kq_masked", il);
kq = ggml_soft_max(ctx, kq);
cb(kq, "kq_soft_max", il);
} else {
kq = ggml_soft_max_ext(ctx, kq, kq_mask, 1.0f/sqrtf(float(n_embd_head)));
cb(kq, "kq_soft_max_ext", il);
}
kq = ggml_add(ctx, kq, kq_mask);
cb(kq, "kq_masked", il);
kq = ggml_soft_max(ctx, kq);
cb(kq, "kq_soft_max", il);
// split cached v into n_head heads
struct ggml_tensor * v =
ggml_view_3d(ctx, kv.v,
@ -5041,6 +5047,7 @@ static const std::unordered_map<const char *, llm_offload_func_e> k_offload_map
{ "kq_scaled_alibi", OFFLOAD_FUNC_KQ },
{ "kq_masked", OFFLOAD_FUNC_KQ },
{ "kq_soft_max", OFFLOAD_FUNC_V },
{ "kq_soft_max_ext", OFFLOAD_FUNC_V },
{ "v", OFFLOAD_FUNC_V },
{ "kqv", OFFLOAD_FUNC_V },
{ "kqv_merged", OFFLOAD_FUNC_V },