Skip to content

Commit cdf7658

Browse files
JohannesGaesslerJohannesGaessler
authored
CUDA: fix non-cont. inputs for batched mat mul (ggml-org#13155)
1 parent 7d3af70 commit cdf7658

File tree

4 files changed

+94
-42
lines changed

4 files changed

+94
-42
lines changed

‎ggml/src/ggml-cuda/convert.cu‎

Lines changed: 41 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,8 @@
11
#include "convert.cuh"
22
#include "dequantize.cuh"
33

4+
#include <cstdint>
5+
46
#define CUDA_Q8_0_NE_ALIGN 2048
57

68
template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
@@ -570,30 +572,46 @@ static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t
570572
}
571573

572574
template <typename src_t, typename dst_t>
573-
static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
574-
const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
575+
static __global__ void convert_unary(
576+
const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01, const int64_t ne02,
577+
const int64_t s01, const int64_t s02, const int64_t s03) {
578+
const int64_t i00 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
575579

576-
if (i >= k) {
580+
if (i00 >= ne00) {
577581
return;
578582
}
579583

584+
const int64_t i01 = blockIdx.y;
585+
const int64_t i02 = blockIdx.z % ne02;
586+
const int64_t i03 = blockIdx.z / ne02;
587+
580588
const src_t * x = (const src_t *) vx;
581589

582-
y[i] = float(x[i]);
590+
const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
591+
const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
592+
y[iy] = float(x[ix]);
583593
}
584594

585595
template <typename src_t, typename dst_t>
586-
static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
587-
const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
588-
convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
596+
static void convert_unary_cuda(const void * vx, dst_t * y,
597+
const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
598+
const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
599+
const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, ne02*ne03);
600+
convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
601+
(vx, y, ne00, ne01, ne02, s01, s02, s03);
602+
}
603+
604+
template <typename src_t, typename dst_t>
605+
static void convert_unary_cont_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
606+
convert_unary_cuda<src_t>(vx, y, k, 1, 1, 1, k, k, k, stream);
589607
}
590608

591609
to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
592610
switch (type) {
593611
case GGML_TYPE_F32:
594-
return convert_unary_cuda<float>;
612+
return convert_unary_cont_cuda<float>;
595613
case GGML_TYPE_F16:
596-
return convert_unary_cuda<half>;
614+
return convert_unary_cont_cuda<half>;
597615
default:
598616
return nullptr;
599617
}
@@ -643,9 +661,9 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
643661
case GGML_TYPE_IQ3_S:
644662
return dequantize_row_iq3_s_cuda;
645663
case GGML_TYPE_F32:
646-
return convert_unary_cuda<float>;
664+
return convert_unary_cont_cuda<float>;
647665
case GGML_TYPE_BF16:
648-
return convert_unary_cuda<nv_bfloat16>;
666+
return convert_unary_cont_cuda<nv_bfloat16>;
649667
default:
650668
return nullptr;
651669
}
@@ -692,7 +710,18 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
692710
case GGML_TYPE_IQ3_S:
693711
return dequantize_row_iq3_s_cuda;
694712
case GGML_TYPE_F16:
695-
return convert_unary_cuda<half>;
713+
return convert_unary_cont_cuda<half>;
714+
case GGML_TYPE_BF16:
715+
return convert_unary_cont_cuda<nv_bfloat16>;
716+
default:
717+
return nullptr;
718+
}
719+
}
720+
721+
to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
722+
switch (type) {
723+
case GGML_TYPE_F32:
724+
return convert_unary_cuda<float>;
696725
case GGML_TYPE_BF16:
697726
return convert_unary_cuda<nv_bfloat16>;
698727
default:

‎ggml/src/ggml-cuda/convert.cuh‎

Lines changed: 11 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -3,7 +3,7 @@
33
#define CUDA_DEQUANTIZE_BLOCK_SIZE 256
44

55
template<typename T>
6-
using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int64_t k, cudaStream_t stream);
6+
using to_t_cuda_t = void (*)(const void * x, T * y, int64_t k, cudaStream_t stream);
77

88
typedef to_t_cuda_t<float> to_fp32_cuda_t;
99
typedef to_t_cuda_t<half> to_fp16_cuda_t;
@@ -14,3 +14,13 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type);
1414
to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type);
1515

1616
to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type);
17+
18+
// TODO more general support for non-contiguous inputs
19+
20+
template<typename T>
21+
using to_t_nc_cuda_t = void (*)(const void * x, T * y,
22+
int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03,
23+
int64_t s01, int64_t s02, int64_t s03, cudaStream_t stream);
24+
25+
typedef to_t_nc_cuda_t<half> to_fp16_nc_cuda_t;
26+
to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);

‎ggml/src/ggml-cuda/ggml-cuda.cu‎

Lines changed: 42 additions & 28 deletions
Original file line numberDiff line numberDiff line change
@@ -1720,15 +1720,15 @@ static __global__ void k_compute_batched_ptrs(
17201720
size_t nb12, size_t nb13,
17211721
size_t nbd2, size_t nbd3,
17221722
int64_t r2, int64_t r3) {
1723-
int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
1724-
int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
1723+
const int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
1724+
const int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
17251725

17261726
if (i13 >= ne13 || i12 >= ne12) {
17271727
return;
17281728
}
17291729

1730-
int64_t i03 = i13 / r3;
1731-
int64_t i02 = i12 / r2;
1730+
const int64_t i03 = i13 / r3;
1731+
const int64_t i02 = i12 / r2;
17321732

17331733
ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03;
17341734
ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13;
@@ -1742,6 +1742,10 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
17421742
GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
17431743
GGML_ASSERT(src0->type == GGML_TYPE_F16);
17441744

1745+
// Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.
1746+
// As long as dst is contiguous this does not matter though.
1747+
GGML_ASSERT(ggml_is_contiguous(dst));
1748+
17451749
GGML_TENSOR_BINARY_OP_LOCALS
17461750

17471751
const int64_t ne_dst = ggml_nelements(dst);
@@ -1750,21 +1754,31 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
17501754

17511755
CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));
17521756

1753-
void * src0_ddq = src0->data;
1754-
half * src0_f16 = (half *) src0_ddq;
1755-
float * src1_ddf = (float *) src1->data;
1756-
float * dst_ddf = (float *) dst->data;
1757+
const half * src0_f16 = (const half *) src0->data;
1758+
float * dst_ddf = (float *) dst->data;
17571759

1758-
// convert src1 to fp16
1760+
const half * src1_f16 = (const half *) src1->data;
1761+
const size_t ts_src1 = ggml_type_size(src1->type);
1762+
GGML_ASSERT(nb10 == ts_src1);
1763+
int64_t s11 = nb11 / ts_src1;
1764+
int64_t s12 = nb12 / ts_src1;
1765+
int64_t s13 = nb13 / ts_src1;
17591766
ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
1767+
1768+
// convert src1 to fp16
17601769
if (src1->type != GGML_TYPE_F16) {
1761-
const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
1770+
const to_fp16_nc_cuda_t to_fp16_cuda = ggml_get_to_fp16_nc_cuda(src1->type);
17621771
const int64_t ne_src1 = ggml_nelements(src1);
17631772
src1_f16_alloc.alloc(ne_src1);
17641773
GGML_ASSERT(to_fp16_cuda != nullptr);
1765-
to_fp16_cuda(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream);
1774+
1775+
to_fp16_cuda(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
1776+
1777+
src1_f16 = src1_f16_alloc.get();
1778+
s11 = ne10;
1779+
s12 = ne11*s11;
1780+
s13 = ne12*s12;
17661781
}
1767-
half * src1_f16 = src1->type == GGML_TYPE_F16 ? (half *) src1_ddf : src1_f16_alloc.get();
17681782

17691783
ggml_cuda_pool_alloc<half> dst_f16(ctx.pool());
17701784
char * dst_t;
@@ -1824,13 +1838,13 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
18241838
int i02 = i12 / r2;
18251839

18261840
CUBLAS_CHECK(
1827-
cublasGemmEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
1828-
ne01, ne11, ne10,
1829-
alpha, (const char *) src0_as_f16 + i02*src0->nb[2] + i03*src0->nb[3] , CUDA_R_16F, nb01/sizeof(half),
1830-
(const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F, nb11/sizeof(float),
1831-
beta, ( char *) dst_t + i12*nbd2 + i13*nbd3, cu_data_type, ne01,
1832-
cu_compute_type,
1833-
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
1841+
cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
1842+
ne01, ne11, ne10,
1843+
alpha, (const char *) src0_f16 + i03*nb03 + i02*nb02, CUDA_R_16F, nb01/sizeof(half),
1844+
src1_f16 + i13*s13 + i12*s12, CUDA_R_16F, s11,
1845+
beta, ( char *) dst_t + i13*nbd3 + i12*nbd2, cu_data_type, ne0,
1846+
cu_compute_type,
1847+
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
18341848
}
18351849
}
18361850
}
@@ -1841,15 +1855,15 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
18411855
CUBLAS_CHECK(
18421856
cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
18431857
ne01, ne11, ne10,
1844-
alpha, (const char *) src0_f16, CUDA_R_16F, nb01/nb00, nb02/nb00, // strideA
1845-
(const char *) src1_f16, CUDA_R_16F, nb11/nb10, nb12/nb10, // strideB
1846-
beta, ( char *) dst_t, cu_data_type, ne01, nb2/nb0, // strideC
1858+
alpha, src0_f16, CUDA_R_16F, nb01/nb00, nb02/nb00, // strideA
1859+
src1_f16, CUDA_R_16F, s11, s12, // strideB
1860+
beta, dst_t, cu_data_type, ne0, ne1*ne0, // strideC
18471861
ne12*ne13,
18481862
cu_compute_type,
18491863
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
18501864
} else {
18511865
// use cublasGemmBatchedEx
1852-
const int ne23 = ne12*ne13;
1866+
const int64_t ne23 = ne12*ne13;
18531867

18541868
ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
18551869
ggml_cuda_pool_alloc< void *> ptrs_dst(ctx.pool(), 1*ne23);
@@ -1861,8 +1875,8 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
18611875
ne12, ne13,
18621876
ne23,
18631877
nb02, nb03,
1864-
src1->type == GGML_TYPE_F16 ? nb12 : nb12/2,
1865-
src1->type == GGML_TYPE_F16 ? nb13 : nb13/2,
1878+
src1->type == GGML_TYPE_F16 ? nb12 : s12*sizeof(half),
1879+
src1->type == GGML_TYPE_F16 ? nb13 : s13*sizeof(half),
18661880
nbd2, nbd3,
18671881
r2, r3);
18681882
CUDA_CHECK(cudaGetLastError());
@@ -1871,8 +1885,8 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
18711885
cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
18721886
ne01, ne11, ne10,
18731887
alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F, nb01/nb00,
1874-
(const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F, nb11/nb10,
1875-
beta, ( void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne01,
1888+
(const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F, s11,
1889+
beta, ( void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
18761890
ne23,
18771891
cu_compute_type,
18781892
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
@@ -1936,7 +1950,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
19361950
} else if (!split && use_mul_mat_vec_q) {
19371951
ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
19381952
} else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
1939-
dst->op_params[0] == GGML_PREC_DEFAULT && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
1953+
!ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
19401954
// general KQ + KQV multi-batch without FlashAttention
19411955
ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
19421956
} else if (use_mul_mat_vec) {

‎src/llama-model.cpp‎

Lines changed: 0 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -10195,7 +10195,6 @@ struct llm_build_deepseek2 : public llm_graph_context {
1019510195

1019610196
// {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
1019710197
ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
10198-
ggml_mul_mat_set_prec(q_nope_absorbed, GGML_PREC_F32);
1019910198
cb(q_nope_absorbed, "q_nope_absorbed", il);
1020010199

1020110200
// {kv_lora_rank, n_head, n_tokens}

0 commit comments

Comments
 (0)