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

llama : add qwen2moe (#6074)

Browse source 
* support qwen2moe

* fix-review

* metal : support unary ops for nelements % 4 != 0

* metal : require contiguousness for float4 unary kernels

* metal : require contiguousness for float4 unary kernels (cont)

* fix-review

* names : for brevity "SHARED_EXP" -> "SHEXP"

* llama : reuse build_moe_ffn()

* llama : add model type name

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Shijie 2024-04-16 23:40:48 +08:00 • committed by GitHub
parent 8a56075b07
commit f4dea7da18
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 537 additions and 101 deletions
Download patch file Download diff file Expand all files Collapse all files

57
ggml-metal.m
View file

@ -41,8 +41,11 @@ enum ggml_metal_kernel_type {
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,
@ -473,8 +476,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
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);
@ -1178,6 +1184,9 @@ static enum ggml_status ggml_metal_graph_compute(
} 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;
@ -1204,42 +1213,60 @@ 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:
{