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

ggml : full ALiBi support

Browse source
This commit is contained in:
Georgi Gerganov 2024-05-10 10:40:33 +03:00
parent d11afd6652
commit 7fdca3348c
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
10 changed files with 82 additions and 680 deletions
Download patch file Download diff file Expand all files Collapse all files

8
tests/test-backend-ops.cpp
View file

@ -1111,11 +1111,7 @@ struct test_soft_max : public test_case {
if (this->mask) {
mask = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, ne[0], ne[1]);
}
ggml_tensor * pos = nullptr;
if (max_bias > 0.0f) {
pos = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ne[0]);
}
ggml_tensor * out = ggml_soft_max_ext(ctx, a, mask, pos, scale, max_bias);
ggml_tensor * out = ggml_soft_max_ext(ctx, a, mask, scale, max_bias);
return out;
}
};
@ -1611,7 +1607,7 @@ public:
struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
kq = ggml_soft_max_ext(ctx, kq, kq_mask, nullptr, kq_scale, 0.0f);
kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale, 0.0f);
// split cached v into n_head heads
struct ggml_tensor * v =