CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-14 11:07:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: batched+noncont MMQ, refactor bs>1 MoE code (#13199)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-04-30 23:12:59 +02:00
committed by GitHub
parent 6f67cf1f48
commit e1e8e0991f
9 changed files with 869 additions and 440 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
ggml/src/ggml-cuda/quantize.cu
View File

@@ -49,29 +49,38 @@ static __global__ void quantize_q8_1(
template <mmq_q8_1_ds_layout ds_layout>
static __global__ void quantize_mmq_q8_1(
const float * __restrict__ x, void * __restrict__ vy, const int64_t kx0, const int64_t kx1, const int64_t kx0_padded) {
const float * __restrict__ x, const int32_t * __restrict__ ids, void * __restrict__ vy,
const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
const int64_t ne0, const int ne1, const int ne2) {
constexpr int vals_per_scale = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 64 : 32;
constexpr int vals_per_sum = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 16 : 32;
const int64_t ix0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
const int64_t i0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
if (ix0 >= kx0_padded) {
if (i0 >= ne0) {
return;
}
const float4 * x4 = (const float4 *) x;
const int64_t i1 = blockIdx.y;
const int64_t i2 = blockIdx.z % ne2;
const int64_t i3 = blockIdx.z / ne2;
const int64_t ix1 = kx1*blockIdx.z + blockIdx.y;
const int64_t i00 = i0;
const int64_t i01 = ids ? ids[i1] : i1;
const int64_t i02 = i2;
const int64_t i03 = i3;
const float4 * x4 = (const float4 *) x;
block_q8_1_mmq * y = (block_q8_1_mmq *) vy;
const int64_t ib0 = blockIdx.z*((int64_t)gridDim.y*gridDim.x*blockDim.x/QK8_1); // first block of channel
const int64_t ib = ib0 + (ix0 / (4*QK8_1))*kx1 + blockIdx.y; // block index in channel
const int64_t iqs = ix0 % (4*QK8_1); // quant index in block
const int64_t ib = ib0 + (i0 / (4*QK8_1))*ne1 + blockIdx.y; // block index in channel
const int64_t iqs = i0 % (4*QK8_1); // quant index in block
// Load 4 floats per thread and calculate max. abs. value between them:
const float4 xi = ix0 < kx0 ? x4[(ix1*kx0 + ix0)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
const float4 xi = i0 < ne00 ? x4[(i03*s03 + i02*s02 + i01*s01 + i00)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
float amax = fabsf(xi.x);
amax = fmaxf(amax, fabsf(xi.y));
amax = fmaxf(amax, fabsf(xi.z));
@@ -87,7 +96,7 @@ static __global__ void quantize_mmq_q8_1(
if (ds_layout != MMQ_Q8_1_DS_LAYOUT_D4) {
sum = xi.x + xi.y + xi.z + xi.w;
// Exchange calculate sum across vals_per_sum/4 threads.
// Calculate sums across vals_per_sum/4 threads.
#pragma unroll
for (int offset = vals_per_sum/8; offset > 0; offset >>= 1) {
sum += __shfl_xor_sync(0xFFFFFFFF, sum, offset, WARP_SIZE);
@@ -137,9 +146,10 @@ static __global__ void quantize_mmq_q8_1(
}
void quantize_row_q8_1_cuda(
const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
const float * x, const int32_t * ids, void * vy, const ggml_type type_src0,
const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
GGML_ASSERT(!ids);
GGML_ASSERT(ne0 % QK8_1 == 0);
const int64_t block_num_x = (ne0 + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
@@ -150,9 +160,9 @@ void quantize_row_q8_1_cuda(
}
void quantize_mmq_q8_1_cuda(
const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
const float * x, const int32_t * ids, void * vy, const ggml_type type_src0,
const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
GGML_ASSERT(ne0 % (4*QK8_1) == 0);
const int64_t block_num_x = (ne0 + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
@@ -161,21 +171,18 @@ void quantize_mmq_q8_1_cuda(
switch (mmq_get_q8_1_ds_layout(type_src0)) {
case MMQ_Q8_1_DS_LAYOUT_D4:
quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D4>
<<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, ne1, ne0);
<<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
break;
case MMQ_Q8_1_DS_LAYOUT_DS4:
quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_DS4>
<<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, ne1, ne0);
<<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
break;
case MMQ_Q8_1_DS_LAYOUT_D2S6:
quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D2S6>
<<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, ne1, ne0);
<<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
break;
default:
GGML_ABORT("fatal error");
break;
}
GGML_UNUSED(s01);
GGML_UNUSED(s02);
GGML_UNUSED(s03);
}