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

ggml : group all experts in a single ggml_mul_mat_id (#6505)

Browse source 
* ggml : group all experts in a single ggml_mul_mat_id
cuda : improve mmid row copy

* cuda : fix bin bcast with non-cont src0

* test-backend-ops : only run all mul mat tests for base types

* llama : disable moe offloading with SYCL

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
slaren 2024-04-18 15:18:48 +02:00 • committed by GitHub
parent 03c0946d73
commit 0d56246f4b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
12 changed files with 971 additions and 821 deletions
Download patch file Download diff file Expand all files Collapse all files

125
ggml-metal.m
View file

@ -1732,15 +1732,10 @@ static enum ggml_status ggml_metal_graph_compute(
} break;
case GGML_OP_MUL_MAT_ID:
{
//GGML_ASSERT(ne00 == ne10);
//GGML_ASSERT(ne03 == ne13);
const int n_as = src0->ne[2];
// max size of the src1ids array in the kernel shared buffer
GGML_ASSERT(ne11 <= 4096);
// src2 = ids
const int64_t ne20 = src2->ne[0]; GGML_UNUSED(ne20);
const int64_t ne20 = src2->ne[0];
const int64_t ne21 = src2->ne[1];
const int64_t ne22 = src2->ne[2]; GGML_UNUSED(ne22);
const int64_t ne23 = src2->ne[3]; GGML_UNUSED(ne23);
@ -1761,15 +1756,13 @@ static enum ggml_status ggml_metal_graph_compute(
// find the break-even point where the matrix-matrix kernel becomes more efficient compared
// to the matrix-vector kernel
int ne11_mm_min = n_as;
// ne20 = n_used_experts
// ne21 = n_rows
const int dst_rows = ne20*ne21;
const int dst_rows_min = n_as;
const int idx = ((int32_t *) dst->op_params)[0];
// batch size
GGML_ASSERT(ne21 == ne11); // ?
GGML_ASSERT(ne12 == 1 && ne13 == 1); // no broadcasting
const uint r2 = 1;
const uint r3 = 1;
// max size of the rowids array in the kernel shared buffer
GGML_ASSERT(dst_rows <= 2048);
// for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
// AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
@ -1779,7 +1772,7 @@ static enum ggml_status ggml_metal_graph_compute(
// !!!
if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
ne00 % 32 == 0 && ne00 >= 64 &&
ne11 > ne11_mm_min) {
dst_rows > dst_rows_min) {
// some Metal matrix data types require aligned pointers
// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
@ -1821,26 +1814,26 @@ static enum ggml_status ggml_metal_graph_compute(
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
[encoder setBytes:&nb21 length:sizeof(nb21) atIndex:4];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:5];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:9];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:10];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:11];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:12];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:13];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:14];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:15];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:16];
[encoder setBytes:&r2 length:sizeof(r2) atIndex:17];
[encoder setBytes:&r3 length:sizeof(r3) atIndex:18];
[encoder setBytes:&idx length:sizeof(idx) atIndex:19];
[encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
[encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
[encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:8];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:9];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:10];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:17];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:18];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:19];
[encoder setThreadgroupMemoryLength:GGML_PAD(8192 + 2*ne11, 16) atIndex:0];
[encoder setThreadgroupMemoryLength:GGML_PAD(8192 + dst_rows*4/*sizeof(ushort2)*/, 16) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne11 + 31)/32, (ne01 + 63)/64, n_as*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, n_as) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
} else {
int nth0 = 32;
int nth1 = 1;
@ -1993,72 +1986,72 @@ static enum ggml_status ggml_metal_graph_compute(
GGML_ASSERT(ne00 >= nth0*nth1);
}
const int64_t _ne1 = 1; // kernels needs a reference in constant memory
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
[encoder setBytes:&nb21 length:sizeof(nb21) atIndex:4];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:5];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:6];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:7];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:8];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:9];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:10];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
[encoder setBytes:&_ne1 length:sizeof(_ne1) atIndex:12];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:18];
[encoder setBytes:&_ne1 length:sizeof(_ne1) atIndex:19];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:20];
[encoder setBytes:&r2 length:sizeof(r2) atIndex:21];
[encoder setBytes:&r3 length:sizeof(r3) atIndex:22];
[encoder setBytes:&idx length:sizeof(idx) atIndex:23];
[encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
[encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
[encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:8];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:9];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:10];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:11];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:12];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:13];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:14];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:15];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:16];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:17];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:18];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:19];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:20];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:21];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:22];
const int64_t _ne1 = 1;
const int tgz = dst_rows;
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q5_0 ||
src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K ||
src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
const int mem_size = 32*sizeof(float);
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q3_K) {
#ifdef GGML_QKK_64
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#else
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
#endif
}
else if (src0t == GGML_TYPE_Q5_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q6_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else {
const int64_t ny = (_ne1 + nrows - 1)/nrows;
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne21*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
const int64_t ny = (_ne1 + nrows - 1)/nrows; // = _ne1
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
}
} break;