Skip to content

Commit d383c03

Browse files
LostRuinsLostRuins
committed
Merge branch 'upstream' into concedo_experimental
# Conflicts: # tests/test-backend-ops.cpp
2 parents 2594be7 + 27208bf commit d383c03

File tree

4 files changed

+182
-88
lines changed

4 files changed

+182
-88
lines changed

‎ggml/src/ggml-cuda/convert.cu‎

Lines changed: 22 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -728,3 +728,25 @@ to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
728728
return nullptr;
729729
}
730730
}
731+
732+
to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type) {
733+
switch (type) {
734+
case GGML_TYPE_F32:
735+
return convert_unary_cuda<float, nv_bfloat16>;
736+
case GGML_TYPE_F16:
737+
return convert_unary_cuda<half, nv_bfloat16>;
738+
default:
739+
return nullptr;
740+
}
741+
}
742+
743+
to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type) {
744+
switch (type) {
745+
case GGML_TYPE_F16:
746+
return convert_unary_cuda<half, float>;
747+
case GGML_TYPE_BF16:
748+
return convert_unary_cuda<nv_bfloat16, float>;
749+
default:
750+
return nullptr;
751+
}
752+
}

‎ggml/src/ggml-cuda/convert.cuh‎

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -22,5 +22,10 @@ using to_t_nc_cuda_t = void (*)(const void * x, T * y,
2222
int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03,
2323
int64_t s01, int64_t s02, int64_t s03, cudaStream_t stream);
2424

25+
typedef to_t_nc_cuda_t<float> to_fp32_nc_cuda_t;
2526
typedef to_t_nc_cuda_t<half> to_fp16_nc_cuda_t;
27+
typedef to_t_nc_cuda_t<nv_bfloat16> to_bf16_nc_cuda_t;
28+
29+
to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type);
2630
to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);
31+
to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type);

‎ggml/src/ggml-cuda/ggml-cuda.cu‎

Lines changed: 131 additions & 76 deletions
Original file line numberDiff line numberDiff line change
@@ -1750,7 +1750,7 @@ static void ggml_cuda_op_mul_mat(
17501750
}
17511751

17521752
static __global__ void k_compute_batched_ptrs(
1753-
const half * src0_as_f16, const half * src1_as_f16, char * dst,
1753+
const void * src0_as_f16, const void * src1_as_f16, char * dst,
17541754
const void ** ptrs_src, void ** ptrs_dst,
17551755
int64_t ne12, int64_t ne13,
17561756
int64_t ne23,
@@ -1773,91 +1773,139 @@ static __global__ void k_compute_batched_ptrs(
17731773
ptrs_dst[0*ne23 + i12 + i13*ne12] = ( char *) dst + i12*nbd2 + i13*nbd3;
17741774
}
17751775

1776-
static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
1776+
// Type traits for mapping ggml types to CUDA/cuBLAS types
1777+
template<ggml_type T>
1778+
struct batched_mul_mat_traits;
1779+
1780+
template<>
1781+
struct batched_mul_mat_traits<GGML_TYPE_F32> {
1782+
using cuda_type = float;
1783+
static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_32F;
1784+
static inline const cudaDataType_t data_type = CUDA_R_32F;
1785+
static inline const ggml_type ggml_type_val = GGML_TYPE_F32;
1786+
static inline const float alpha = 1.0f;
1787+
static inline const float beta = 0.0f;
1788+
static inline const void* get_alpha() { static const float val = alpha; return &val; }
1789+
static inline const void* get_beta() { static const float val = beta; return &val; }
1790+
static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_fp32_nc_cuda(src_type); }
1791+
};
1792+
1793+
template<>
1794+
struct batched_mul_mat_traits<GGML_TYPE_BF16> {
1795+
using cuda_type = nv_bfloat16;
1796+
static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_32F;
1797+
static inline const cudaDataType_t data_type = CUDA_R_16BF;
1798+
static inline const ggml_type ggml_type_val = GGML_TYPE_BF16;
1799+
static inline const float alpha = 1.0f;
1800+
static inline const float beta = 0.0f;
1801+
static inline const void* get_alpha() { static const float val = alpha; return &val; }
1802+
static inline const void* get_beta() { static const float val = beta; return &val; }
1803+
static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_bf16_nc_cuda(src_type); }
1804+
};
1805+
1806+
template<>
1807+
struct batched_mul_mat_traits<GGML_TYPE_F16> {
1808+
using cuda_type = half;
1809+
static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_16F;
1810+
static inline const cudaDataType_t data_type = CUDA_R_16F;
1811+
static inline const ggml_type ggml_type_val = GGML_TYPE_F16;
1812+
static inline const half alpha = 1.0;
1813+
static inline const half beta = 0.0;
1814+
static inline const void* get_alpha() { static const half val = alpha; return &val; }
1815+
static inline const void* get_beta() { static const half val = beta; return &val; }
1816+
static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_fp16_nc_cuda(src_type); }
1817+
};
1818+
1819+
template<ggml_type src0_type>
1820+
static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
1821+
using traits = batched_mul_mat_traits<src0_type>;
1822+
using cuda_t = typename traits::cuda_type;
1823+
17771824
GGML_ASSERT(!ggml_is_transposed(src0));
17781825
GGML_ASSERT(!ggml_is_transposed(src1));
1779-
17801826
GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft));
1781-
GGML_ASSERT(src0->type == GGML_TYPE_F16);
1827+
GGML_ASSERT(src0->type == src0_type);
1828+
GGML_ASSERT(ggml_is_contiguous(dst));
17821829

17831830
// Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.
17841831
// As long as dst is contiguous this does not matter though.
1785-
GGML_ASSERT(ggml_is_contiguous(dst));
17861832

17871833
GGML_TENSOR_BINARY_OP_LOCALS
17881834

17891835
const int64_t ne_dst = ggml_nelements(dst);
1790-
17911836
cudaStream_t main_stream = ctx.stream();
1792-
17931837
CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));
17941838

1795-
const half * src0_f16 = (const half *) src0->data;
17961839
float * dst_ddf = (float *) dst->data;
1797-
1798-
const half * src1_f16 = (const half *) src1->data;
17991840
const size_t ts_src1 = ggml_type_size(src1->type);
18001841
GGML_ASSERT(nb10 == ts_src1);
18011842
int64_t s11 = nb11 / ts_src1;
18021843
int64_t s12 = nb12 / ts_src1;
18031844
int64_t s13 = nb13 / ts_src1;
1804-
ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
18051845

1806-
// convert src1 to fp16
1807-
if (src1->type != GGML_TYPE_F16) {
1808-
const to_fp16_nc_cuda_t to_fp16_cuda = ggml_get_to_fp16_nc_cuda(src1->type);
1809-
const int64_t ne_src1 = ggml_nelements(src1);
1810-
src1_f16_alloc.alloc(ne_src1);
1811-
GGML_ASSERT(to_fp16_cuda != nullptr);
1846+
const cuda_t * src0_ptr = nullptr;
1847+
const cuda_t * src1_ptr = nullptr;
18121848

1813-
to_fp16_cuda(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
1849+
ggml_cuda_pool_alloc<cuda_t> src0_alloc(ctx.pool());
1850+
ggml_cuda_pool_alloc<cuda_t> src1_alloc(ctx.pool());
1851+
1852+
// Handle src0
1853+
src0_ptr = (const cuda_t *) src0->data;
1854+
1855+
// Handle src1 - convert if necessary
1856+
if (src1->type == src0_type) {
1857+
src1_ptr = (const cuda_t *) src1->data;
1858+
} else {
1859+
// Convert src1 to target type using traits conversion functions
1860+
const int64_t ne_src1 = ggml_nelements(src1);
1861+
src1_alloc.alloc(ne_src1);
18141862

1815-
src1_f16 = src1_f16_alloc.get();
1863+
const auto convert_func = traits::get_nc_converter(src1->type);
1864+
GGML_ASSERT(convert_func != nullptr);
1865+
convert_func(src1->data, src1_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
1866+
src1_ptr = src1_alloc.get();
18161867
s11 = ne10;
18171868
s12 = ne11*s11;
18181869
s13 = ne12*s12;
18191870
}
18201871

1821-
ggml_cuda_pool_alloc<half> dst_f16(ctx.pool());
1872+
// Setup destination buffer
1873+
ggml_cuda_pool_alloc<cuda_t> dst_temp(ctx.pool());
18221874
char * dst_t;
1823-
1824-
cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
1825-
cudaDataType_t cu_data_type = CUDA_R_16F;
1826-
1827-
// dst strides
18281875
size_t nbd2 = dst->nb[2];
18291876
size_t nbd3 = dst->nb[3];
18301877

1831-
const half alpha_f16 = 1.0f;
1832-
const half beta_f16 = 0.0f;
1833-
1878+
cublasComputeType_t cu_compute_type = traits::compute_type;
1879+
cudaDataType_t cu_data_type = traits::data_type;
1880+
cudaDataType_t cu_data_type_a = traits::data_type;
1881+
cudaDataType_t cu_data_type_b = traits::data_type;
1882+
const void * alpha = traits::get_alpha();
1883+
const void * beta = traits::get_beta();
18341884
const float alpha_f32 = 1.0f;
1835-
const float beta_f32 = 0.0f;
1836-
1837-
const void * alpha = &alpha_f16;
1838-
const void * beta = &beta_f16;
1885+
const float beta_f32 = 0.0f;
18391886

18401887
if (dst->op_params[0] == GGML_PREC_DEFAULT) {
1841-
dst_t = (char *) dst_f16.alloc(ne_dst);
1842-
1843-
nbd2 /= sizeof(float) / sizeof(half);
1844-
nbd3 /= sizeof(float) / sizeof(half);
1888+
if constexpr (src0_type == GGML_TYPE_F32) {
1889+
dst_t = (char *) dst_ddf; // Direct F32 output
1890+
} else {
1891+
dst_t = (char *) dst_temp.alloc(ne_dst);
1892+
nbd2 /= sizeof(float) / sizeof(cuda_t);
1893+
nbd3 /= sizeof(float) / sizeof(cuda_t);
1894+
}
18451895
} else {
18461896
dst_t = (char *) dst_ddf;
1847-
18481897
cu_compute_type = CUBLAS_COMPUTE_32F;
1849-
cu_data_type = CUDA_R_32F;
1850-
1898+
cu_data_type = CUDA_R_32F;
18511899
alpha = &alpha_f32;
1852-
beta = &beta_f32;
1900+
beta = &beta_f32;
18531901
}
18541902

18551903
int id = ggml_cuda_get_device();
18561904
const int cc = ggml_cuda_info().devices[id].cc;
18571905
if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
18581906
cu_compute_type = CUBLAS_COMPUTE_32F;
18591907
alpha = &alpha_f32;
1860-
beta = &beta_f32;
1908+
beta = &beta_f32;
18611909
}
18621910

18631911
GGML_ASSERT(ne12 % ne02 == 0);
@@ -1867,35 +1915,15 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
18671915
const int64_t r2 = ne12/ne02;
18681916
const int64_t r3 = ne13/ne03;
18691917

1870-
#if 0
1871-
// use cublasGemmEx
1872-
{
1873-
for (int i13 = 0; i13 < ne13; ++i13) {
1874-
for (int i12 = 0; i12 < ne12; ++i12) {
1875-
int i03 = i13 / r3;
1876-
int i02 = i12 / r2;
1877-
1878-
CUBLAS_CHECK(
1879-
cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
1880-
ne01, ne11, ne10,
1881-
alpha, (const char *) src0_f16 + i03*nb03 + i02*nb02, CUDA_R_16F, nb01/sizeof(half),
1882-
src1_f16 + i13*s13 + i12*s12, CUDA_R_16F, s11,
1883-
beta, ( char *) dst_t + i13*nbd3 + i12*nbd2, cu_data_type, ne0,
1884-
cu_compute_type,
1885-
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
1886-
}
1887-
}
1888-
}
1889-
#else
18901918
if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
18911919
// there is no broadcast and src0, src1 are contiguous across dims 2, 3
18921920
// use cublasGemmStridedBatchedEx
18931921
CUBLAS_CHECK(
18941922
cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
18951923
ne01, ne11, ne10,
1896-
alpha, src0_f16, CUDA_R_16F, nb01/nb00, nb02/nb00, // strideA
1897-
src1_f16, CUDA_R_16F, s11, s12, // strideB
1898-
beta, dst_t, cu_data_type, ne0, ne1*ne0, // strideC
1924+
alpha, src0_ptr, cu_data_type_a, nb01/nb00, nb02/nb00, // strideA
1925+
src1_ptr, cu_data_type_b, s11, s12, // strideB
1926+
beta, dst_t, cu_data_type, ne0, ne1*ne0, // strideC
18991927
ne12*ne13,
19001928
cu_compute_type,
19011929
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
@@ -1906,34 +1934,55 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
19061934
ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
19071935
ggml_cuda_pool_alloc< void *> ptrs_dst(ctx.pool(), 1*ne23);
19081936

1937+
size_t src1_stride_size = sizeof(cuda_t);
1938+
19091939
dim3 block_dims(ne13, ne12);
19101940
k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
1911-
src0_f16, src1_f16, dst_t,
1941+
src0_ptr, src1_ptr, dst_t,
19121942
ptrs_src.get(), ptrs_dst.get(),
19131943
ne12, ne13,
19141944
ne23,
19151945
nb02, nb03,
1916-
src1->type == GGML_TYPE_F16 ? nb12 : s12*sizeof(half),
1917-
src1->type == GGML_TYPE_F16 ? nb13 : s13*sizeof(half),
1946+
(src1->type == src0_type) ? nb12 : s12*src1_stride_size,
1947+
(src1->type == src0_type) ? nb13 : s13*src1_stride_size,
19181948
nbd2, nbd3,
19191949
r2, r3);
1950+
19201951
CUDA_CHECK(cudaGetLastError());
19211952

19221953
CUBLAS_CHECK(
19231954
cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
19241955
ne01, ne11, ne10,
1925-
alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F, nb01/nb00,
1926-
(const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F, s11,
1927-
beta, ( void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
1956+
alpha, (const void **) (ptrs_src.get() + 0*ne23), cu_data_type_a, nb01/nb00,
1957+
(const void **) (ptrs_src.get() + 1*ne23), cu_data_type_b, s11,
1958+
beta, ( void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
19281959
ne23,
19291960
cu_compute_type,
19301961
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
19311962
}
1932-
#endif
19331963

1934-
if (dst->op_params[0] == GGML_PREC_DEFAULT && cu_data_type == CUDA_R_16F) {
1935-
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
1936-
to_fp32_cuda(dst_f16.get(), dst_ddf, ne_dst, main_stream);
1964+
// Convert output back to F32 if needed
1965+
if (dst->op_params[0] == GGML_PREC_DEFAULT && cu_data_type != CUDA_R_32F) {
1966+
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(traits::ggml_type_val);
1967+
to_fp32_cuda(dst_temp.get(), dst_ddf, ne_dst, main_stream);
1968+
}
1969+
}
1970+
1971+
static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
1972+
GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || src0->type == GGML_TYPE_F32);
1973+
1974+
switch (src0->type) {
1975+
case GGML_TYPE_F32:
1976+
ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_F32>(ctx, src0, src1, dst);
1977+
break;
1978+
case GGML_TYPE_BF16:
1979+
ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_BF16>(ctx, src0, src1, dst);
1980+
break;
1981+
case GGML_TYPE_F16:
1982+
ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_F16>(ctx, src0, src1, dst);
1983+
break;
1984+
default:
1985+
GGML_ABORT("Unsupported type");
19371986
}
19381987
}
19391988

@@ -1985,6 +2034,12 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
19852034
//printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
19862035
//printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
19872036

2037+
//TODO update for generic tensor parallelism
2038+
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
2039+
bool use_batched_cublas_f16 = src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16);
2040+
bool use_batched_cublas_bf16 = src0->type == GGML_TYPE_BF16 && bf16_mma_hardware_available(cc);
2041+
bool use_batched_cublas_f32 = src0->type == GGML_TYPE_F32;
2042+
19882043
if (!split && use_mul_mat_vec) {
19892044
// the custom F16 vector kernel can be used over batched cuBLAS GEMM
19902045
// but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
@@ -1993,8 +2048,8 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
19932048
ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
19942049
} else if (!split && use_mul_mat_q) {
19952050
ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, dst);
1996-
} else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
1997-
!ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
2051+
} else if (!split && (use_batched_cublas_f16 || use_batched_cublas_bf16 || use_batched_cublas_f32)
2052+
&& !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
19982053
// general KQ + KQV multi-batch without FlashAttention
19992054
ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
20002055
} else if (use_mul_mat_vec) {

0 commit comments

Comments
 (0)