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

falcon : minor changes (still chasing the Metal problem)

Browse source
This commit is contained in:
Georgi Gerganov 2023-08-23 12:25:49 +03:00
parent a0dc47a501
commit e2d23bed1b
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
3 changed files with 51 additions and 52 deletions
Download patch file Download diff file Expand all files Collapse all files

2
ggml-metal.m
View file

@ -992,7 +992,9 @@ void ggml_metal_graph_compute(
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16]; [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17]; [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
[encoder setBytes:&m0 length:sizeof( float) atIndex:18]; [encoder setBytes:&m0 length:sizeof( float) atIndex:18];
const int nth = 32; const int nth = 32;
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break; } break;
case GGML_OP_ROPE: case GGML_OP_ROPE:

15
llama.cpp
View file

@ -2545,26 +2545,23 @@ static struct ggml_cgraph * llm_build_falcon(
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
ggml_set_name(KQV_merged, "KQV_merged"); ggml_set_name(KQV_merged, "KQV_merged");
cur = ggml_cpy(ctx0, cur = ggml_cpy(ctx0, KQV_merged, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
KQV_merged,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
ggml_set_name(cur, "KQV_merged_contiguous"); ggml_set_name(cur, "KQV_merged_contiguous");
cur = ggml_cpy(ctx0,
KQV_merged,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur); cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur);
ggml_set_name(cur, "result_wo"); ggml_set_name(cur, "result_wo");
} }
struct ggml_tensor * attn_out = cur;
// feed forward
{
struct ggml_tensor * inpFF = attn_norm; struct ggml_tensor * inpFF = attn_norm;
struct ggml_tensor * attn_out = ggml_cpy(
ctx0, cur, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
cur = ggml_mul_mat(ctx0, model.layers[il].w3, inpFF); cur = ggml_mul_mat(ctx0, model.layers[il].w3, inpFF);
cur = ggml_gelu(ctx0, cur); cur = ggml_gelu(ctx0, cur);
cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur); cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur);
}
cur = ggml_add(ctx0, cur, attn_out); cur = ggml_add(ctx0, cur, attn_out);
cur = ggml_add(ctx0, cur, inpL); cur = ggml_add(ctx0, cur, inpL);