Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

CUDA: noncont MMVQ + batched bs1 MUL_MAT_ID (#13014)

Browse Source 
* CUDA: noncont MMVQ + batched bs1 MUL_MAT_ID

* fix logic for RoPE support, CUDA graphs
This commit is contained in:
Johannes Gäßler
2025-04-22 21:27:40 +02:00
committed by GitHub
parent dc39a5e7a8
commit 658987cfc9
9 changed files with 548 additions and 426 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/mmv.cu
+109 -70
View File
@@ -4,18 +4,23 @@
template <typename T, typename type_acc, int block_size>
static __global__ void mul_mat_vec(
const T * __restrict__ x, const float * __restrict__ y, float * __restrict__ dst, const int64_t ncols2, const int64_t stride_row,
const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
const int64_t ncols2, const int64_t nchannels_y, const int64_t stride_row,
const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
const int64_t sample_ratio, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst) {
const int64_t row = blockIdx.x;
const int64_t channel = blockIdx.y;
const int64_t sample = blockIdx.z;
const int tid = threadIdx.x;
constexpr int warp_size = ggml_cuda_get_physical_warp_size();
const int64_t row = blockIdx.x;
const int64_t channel_dst = blockIdx.y;
const int64_t channel_x = ids ? ids[channel_dst] : channel_dst / channel_ratio;
const int64_t channel_y = ids ? channel_dst % nchannels_y : channel_dst;
const int64_t sample_dst = blockIdx.z;
const int64_t sample_x = sample_dst / sample_ratio;
const int64_t sample_y = sample_dst;
const int tid = threadIdx.x;
constexpr int warp_size = ggml_cuda_get_physical_warp_size();
x += (sample/sample_ratio)*stride_sample_x + (channel/channel_ratio)*stride_channel_x + row*stride_row;
y += sample *stride_sample_y + channel *stride_channel_y;
dst += sample *stride_sample_dst + channel *stride_channel_dst;
x += sample_x *stride_sample_x + channel_x *stride_channel_x + row*stride_row;
y += sample_y *stride_sample_y + channel_y *stride_channel_y;
dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst;
const float2 * y2 = (const float2 *) y;
@@ -31,12 +36,19 @@ static __global__ void mul_mat_vec(
float sumf = 0.0f;
if constexpr (std::is_same<T, half>::value) {
if constexpr (std::is_same<T, float>::value) {
const float2 * x2 = (const float2 *) x;
for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
const float2 tmpx = x2[col2];
const float2 tmpy = y2[col2];
sumf += tmpx.x*tmpy.x;
sumf += tmpx.y*tmpy.y;
}
} else if constexpr (std::is_same<T, half>::value) {
const half2 * x2 = (const half2 *) x;
if (std::is_same<type_acc, float>::value) {
sumf = 0.0f;
for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
const float2 tmpx = __half22float2(x2[col2]);
const float2 tmpy = y2[col2];
@@ -59,8 +71,6 @@ static __global__ void mul_mat_vec(
}
} else if constexpr (std::is_same<T, nv_bfloat16>::value) {
const int * x2 = (const int *) x;
sumf = 0.0f;
for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
const int tmpx = x2[col2];
const float2 tmpy = y2[col2];
@@ -92,17 +102,17 @@ static __global__ void mul_mat_vec(
template <typename T, typename type_acc>
static void launch_mul_mat_vec_cuda(
const T * x, const float * y, float * dst,
const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
const T * x, const float * y, const int32_t * ids, float * dst,
const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
const int64_t nsamples_y, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
cudaStream_t stream) {
GGML_ASSERT(ncols % 2 == 0);
GGML_ASSERT(stride_row % 2 == 0);
GGML_ASSERT(nchannels_y % nchannels_x == 0);
GGML_ASSERT(nsamples_y % nsamples_x == 0);
const int64_t channel_ratio = nchannels_y / nchannels_x;
const int64_t sample_ratio = nsamples_y / nsamples_x;
GGML_ASSERT(ids || nchannels_dst % nchannels_x == 0);
GGML_ASSERT( nsamples_dst % nsamples_x == 0);
const int64_t channel_ratio = nchannels_dst / nchannels_x;
const int64_t sample_ratio = nsamples_dst / nsamples_x;
int device;
int warp_size;
@@ -124,48 +134,48 @@ static void launch_mul_mat_vec_cuda(
}
const int smem = warp_size*sizeof(float);
const dim3 block_nums(nrows, nchannels_y, nsamples_y);
const dim3 block_nums(nrows, nchannels_dst, nsamples_dst);
const dim3 block_dims(block_size_best, 1, 1);
switch (block_size_best) {
case 32: {
mul_mat_vec<T, type_acc, 32><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
(x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
} break;
case 64: {
mul_mat_vec<T, type_acc, 64><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
(x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
} break;
case 96: {
mul_mat_vec<T, type_acc, 96><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
(x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
} break;
case 128: {
mul_mat_vec<T, type_acc, 128><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
(x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
} break;
case 160: {
mul_mat_vec<T, type_acc, 160><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
(x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
} break;
case 192: {
mul_mat_vec<T, type_acc, 192><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
(x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
} break;
case 224: {
mul_mat_vec<T, type_acc, 224><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
(x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
} break;
case 256: {
mul_mat_vec<T, type_acc, 256><<<block_nums, block_dims, smem, stream>>>
(x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
(x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
} break;
default: {
GGML_ABORT("fatal error");
@@ -175,28 +185,28 @@ static void launch_mul_mat_vec_cuda(
template<typename T>
static void mul_mat_vec_cuda(
const T * x, const float * y, float * dst,
const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
const T * x, const float * y, const int32_t * ids, float * dst,
const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
const int64_t nsamples_y, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
enum ggml_prec prec, cudaStream_t stream) {
switch (prec) {
case GGML_PREC_DEFAULT: {
if constexpr(std::is_same<T, half>::value) {
if (prec == GGML_PREC_DEFAULT) {
launch_mul_mat_vec_cuda<T, half>
(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
} break;
case GGML_PREC_F32: {
launch_mul_mat_vec_cuda<T, float>
(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
} break;
(x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
return;
}
}
launch_mul_mat_vec_cuda<T, float>
(x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
}
void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
GGML_ASSERT( src1->type == GGML_TYPE_F32);
GGML_ASSERT(!ids || ids->type == GGML_TYPE_I32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
GGML_TENSOR_BINARY_OP_LOCALS;
@@ -204,21 +214,24 @@ void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor *
const size_t ts_src1 = ggml_type_size(src1->type);
const size_t ts_dst = ggml_type_size(dst->type);
GGML_ASSERT(ne11 == 1);
GGML_ASSERT(ne12 == ne2);
GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1.
GGML_ASSERT(ne13 == ne3);
GGML_ASSERT(nb00 == ts_src0);
GGML_ASSERT(nb10 == ts_src1);
GGML_ASSERT(nb0 == ts_dst);
GGML_ASSERT( nb00 == ts_src0);
GGML_ASSERT( nb10 == ts_src1);
GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
GGML_ASSERT( nb0 == ts_dst);
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
const float * src1_d = (const float *) src1->data;
float * dst_d = (float *) dst->data;
const float * src1_d = (const float *) src1->data;
const int32_t * ids_d = ids ? (const int32_t *) ids->data : nullptr;
float * dst_d = (float *) dst->data;
const int64_t s01 = src0->nb[1] / ts_src0;
const int64_t s11 = src1->nb[1] / ts_src1;
const int64_t s1 = dst->nb[1] / ts_dst;
const int64_t s02 = src0->nb[2] / ts_src0;
const int64_t s12 = src1->nb[2] / ts_src1;
const int64_t s2 = dst->nb[2] / ts_dst;
@@ -226,14 +239,33 @@ void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor *
const int64_t s13 = src1->nb[3] / ts_src1;
const int64_t s3 = dst->nb[3] / ts_dst;
// For MUL_MAT_ID the memory layout is different than for MUL_MAT:
const int64_t ncols_dst = ids ? ne2 : ne1;
const int64_t nchannels_y = ids ? ne11 : ne12;
const int64_t nchannels_dst = ids ? ne1 : ne2;
const int64_t stride_channel_dst = ids ? s1 : s2;
const int64_t stride_channel_y = ids ? s11 : s12;
GGML_ASSERT(ncols_dst == 1);
switch (src0->type) {
case GGML_TYPE_F32: {
const float * src0_d = (const float *) src0->data;
mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
ne03, ne3, s03, s13, s3, prec, ctx.stream());
} break;
case GGML_TYPE_F16: {
const half * src0_d = (const half *) src0->data;
mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, s01, ne02, ne12, s02, s12, s2, ne03, ne13, s03, s13, s3, prec, ctx.stream());
mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
ne03, ne3, s03, s13, s3, prec, ctx.stream());
} break;
case GGML_TYPE_BF16: {
const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data;
mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, s01, ne02, ne12, s02, s12, s2, ne03, ne13, s03, s13, s3, prec, ctx.stream());
mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
ne03, ne3, s03, s13, s3, prec, ctx.stream());
} break;
default:
GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
@@ -262,27 +294,34 @@ void ggml_cuda_op_mul_mat_vec(
const int64_t stride_row = ne00;
const int64_t nchannels_x = 1;
const int64_t nchannels_y = 1;
const int64_t nchannels_dst = 1;
const int64_t stride_channel_x = 0;
const int64_t stride_channel_y = 0;
const int64_t stride_channel_dst = 0;
const int64_t nsamples_x = 1;
const int64_t nsamples_y = 1;
const int64_t nsamples_dst = 1;
const int64_t stride_sample_x = 0;
const int64_t stride_sample_y = 0;
const int64_t stride_sample_dst = 0;
switch (src0->type) {
case GGML_TYPE_F32: {
const float * src0_d = (const float *) src0_dd_i;
mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
} break;
case GGML_TYPE_F16: {
const half * src0_d = (const half *) src0_dd_i;
mul_mat_vec_cuda(src0_d, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
} break;
case GGML_TYPE_BF16: {
const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i;
mul_mat_vec_cuda(src0_d, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
} break;
default:
GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));