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

Merge branch 'master' into gg/flash-attn

Browse source
This commit is contained in:
Georgi Gerganov 2024-01-31 18:49:43 +02:00
commit 2ddc9bbef1
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
30 changed files with 1758 additions and 1597 deletions
Download patch file Download diff file Expand all files Collapse all files

13
ggml-metal.m
View file

@ -135,6 +135,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_ROPE_F16,
GGML_METAL_KERNEL_TYPE_ALIBI_F32,
GGML_METAL_KERNEL_TYPE_IM2COL_F16,
GGML_METAL_KERNEL_TYPE_IM2COL_F32,
GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
GGML_METAL_KERNEL_TYPE_PAD_F32,
GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
@ -515,6 +516,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16, rope_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32, alibi_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16, im2col_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32, im2col_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32, upscale_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32, pad_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true);
@ -645,6 +647,10 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
case GGML_OP_ALIBI:
case GGML_OP_ROPE:
case GGML_OP_IM2COL:
return true;
case GGML_OP_POOL_1D:
case GGML_OP_POOL_2D:
return false;
case GGML_OP_UPSCALE:
case GGML_OP_PAD:
case GGML_OP_ARGSORT:
@ -2031,7 +2037,7 @@ static bool ggml_metal_graph_compute(
{
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F16);
GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
const int32_t s1 = ((const int32_t *)(dst->op_params))[1];
@ -2039,6 +2045,7 @@ static bool ggml_metal_graph_compute(
const int32_t p1 = ((const int32_t *)(dst->op_params))[3];
const int32_t d0 = ((const int32_t *)(dst->op_params))[4];
const int32_t d1 = ((const int32_t *)(dst->op_params))[5];
const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1;
const int32_t N = src1->ne[is_2D ? 3 : 2];
@ -2059,8 +2066,8 @@ static bool ggml_metal_graph_compute(
id<MTLComputePipelineState> pipeline = nil;
switch (src0->type) {
case GGML_TYPE_F32: GGML_ASSERT(false && "not implemented"); break;
switch (dst->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F16].pipeline; break;
default: GGML_ASSERT(false);
};