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 2024-04-17 10:13:09 +03:00
commit 2c41180e88
No known key found for this signature in database
GPG key ID: BF970631944C16B7
110 changed files with 11660 additions and 6357 deletions
Download patch file Download diff file Expand all files Collapse all files

84
ggml-metal.m
View file

@ -37,11 +37,15 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_DIV_ROW,
GGML_METAL_KERNEL_TYPE_SCALE,
GGML_METAL_KERNEL_TYPE_SCALE_4,
GGML_METAL_KERNEL_TYPE_CLAMP,
GGML_METAL_KERNEL_TYPE_TANH,
GGML_METAL_KERNEL_TYPE_RELU,
GGML_METAL_KERNEL_TYPE_GELU,
GGML_METAL_KERNEL_TYPE_GELU_4,
GGML_METAL_KERNEL_TYPE_GELU_QUICK,
GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,
GGML_METAL_KERNEL_TYPE_SILU,
GGML_METAL_KERNEL_TYPE_SILU_4,
GGML_METAL_KERNEL_TYPE_SOFT_MAX,
GGML_METAL_KERNEL_TYPE_SOFT_MAX_4,
GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,
@ -477,11 +481,15 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW, div_row, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE, scale, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4, scale_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP, clamp, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH, tanh, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU, relu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU, gelu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4, gelu_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK, gelu_quick, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, gelu_quick_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU, silu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4, silu_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX, soft_max, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_4, soft_max_4, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, diag_mask_inf, true);
@ -728,6 +736,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
case GGML_OP_MUL:
case GGML_OP_DIV:
case GGML_OP_SCALE:
case GGML_OP_CLAMP:
case GGML_OP_SQR:
case GGML_OP_SUM_ROWS:
return true;
@ -1170,8 +1179,30 @@ static enum ggml_status ggml_metal_graph_compute(
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_CLAMP:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CLAMP].pipeline;
float min;
float max;
memcpy(&min, ((int32_t *) dst->op_params) + 0, sizeof(float));
memcpy(&max, ((int32_t *) dst->op_params) + 1, sizeof(float));
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&min length:sizeof(min) atIndex:2];
[encoder setBytes:&max length:sizeof(max) atIndex:3];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_UNARY:
switch (ggml_get_unary_op(gf->nodes[i])) {
// we are not taking into account the strides, so for now require contiguous tensors
GGML_ASSERT(ggml_is_contiguous(src0));
case GGML_UNARY_OP_TANH:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TANH].pipeline;
@ -1198,42 +1229,60 @@ static enum ggml_status ggml_metal_graph_compute(
} break;
case GGML_UNARY_OP_GELU:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU].pipeline;
int64_t n = ggml_nelements(dst);
id<MTLComputePipelineState> pipeline = nil;
if (n % 4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_4].pipeline;
n /= 4;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU].pipeline;
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
GGML_ASSERT(n % 4 == 0);
[encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_GELU_QUICK:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK].pipeline;
int64_t n = ggml_nelements(dst);
id<MTLComputePipelineState> pipeline = nil;
if (n % 4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK_4].pipeline;
n /= 4;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK].pipeline;
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
GGML_ASSERT(n % 4 == 0);
[encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_SILU:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU].pipeline;
int64_t n = ggml_nelements(dst);
id<MTLComputePipelineState> pipeline = nil;
if (n % 4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU_4].pipeline;
n /= 4;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU].pipeline;
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
GGML_ASSERT(n % 4 == 0);
[encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
default:
{
@ -1944,7 +1993,12 @@ static enum ggml_status ggml_metal_graph_compute(
{
nth0 = 4;
nth1 = 16;
#if QK_K == 64
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
#else
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline;
#endif
} break;
default:
{