Skip to content

Commit 36ca8b3

Browse files
CISCggerganov
authored andcommitted
CUDA: don't convert BF16 weights to FP32 (ggml/1174)
* add bf16 support * use convert_from_bf16_cuda instead of convert_unary_cuda for f32 * revert 7ec5085 * move functionality into convert_unary with constexpr
1 parent 995083e commit 36ca8b3

File tree

3 files changed

+52
-2
lines changed

3 files changed

+52
-2
lines changed

ggml/src/ggml-cuda/convert.cu

Lines changed: 20 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -579,7 +579,13 @@ static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __res
579579

580580
const src_t * x = (const src_t *) vx;
581581

582-
y[i] = x[i];
582+
if constexpr (std::is_same_v<src_t, nv_bfloat16>) {
583+
y[i] = __bfloat162float(x[i]);
584+
} else if constexpr (std::is_same_v<dst_t, nv_bfloat16> && std::is_same_v<src_t, half>) {
585+
y[i] = (float)x[i];
586+
} else {
587+
y[i] = x[i];
588+
}
583589
}
584590

585591
template <typename src_t, typename dst_t>
@@ -588,6 +594,17 @@ static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict_
588594
convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
589595
}
590596

597+
to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
598+
switch (type) {
599+
case GGML_TYPE_F32:
600+
return convert_unary_cuda<float>;
601+
case GGML_TYPE_F16:
602+
return convert_unary_cuda<half>;
603+
default:
604+
return nullptr;
605+
}
606+
}
607+
591608
to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
592609
switch (type) {
593610
case GGML_TYPE_Q4_0:
@@ -633,6 +650,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
633650
return dequantize_row_iq3_s_cuda;
634651
case GGML_TYPE_F32:
635652
return convert_unary_cuda<float>;
653+
case GGML_TYPE_BF16:
654+
return convert_unary_cuda<nv_bfloat16>;
636655
default:
637656
return nullptr;
638657
}

ggml/src/ggml-cuda/convert.cuh

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -7,7 +7,10 @@ using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, in
77

88
typedef to_t_cuda_t<float> to_fp32_cuda_t;
99
typedef to_t_cuda_t<half> to_fp16_cuda_t;
10+
typedef to_t_cuda_t<nv_bfloat16> to_bf16_cuda_t;
1011

1112
to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type);
1213

14+
to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type);
15+
1316
to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type);

ggml/src/ggml-cuda/ggml-cuda.cu

Lines changed: 29 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1194,7 +1194,35 @@ static void ggml_cuda_op_mul_mat_cublas(
11941194

11951195
const bool use_fp16 = (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT;
11961196

1197-
if (((GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) || GGML_CUDA_CC_IS_AMD(cc)) && use_fp16) {
1197+
if (src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) {
1198+
ggml_cuda_pool_alloc<nv_bfloat16> src1_as_bf16(ctx.pool(id));
1199+
if (src1->type != GGML_TYPE_BF16) {
1200+
const to_bf16_cuda_t to_bf16_cuda = ggml_get_to_bf16_cuda(src1->type);
1201+
GGML_ASSERT(to_bf16_cuda != nullptr);
1202+
size_t ne = src1_ncols*ne10;
1203+
src1_as_bf16.alloc(ne);
1204+
to_bf16_cuda(src1_ddf_i, src1_as_bf16.get(), ne, stream);
1205+
}
1206+
const nv_bfloat16 * src1_ptr = src1->type == GGML_TYPE_BF16 ? (const nv_bfloat16 *) src1_ddf_i : src1_as_bf16.get();
1207+
const nv_bfloat16 * src0_ptr = (const nv_bfloat16 *)src0_dd_i;
1208+
ggml_cuda_pool_alloc<nv_bfloat16> dst_bf16(ctx.pool(id), row_diff*src1_ncols);
1209+
1210+
const float alpha_f32 = 1.0f;
1211+
const float beta_f32 = 0.0f;
1212+
1213+
CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
1214+
CUBLAS_CHECK(
1215+
cublasGemmEx(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N,
1216+
row_diff, src1_ncols, ne10,
1217+
&alpha_f32, src0_ptr, CUDA_R_16BF, ne00,
1218+
src1_ptr, CUDA_R_16BF, ne10,
1219+
&beta_f32, dst_bf16.get(), CUDA_R_16BF, ldc,
1220+
CUBLAS_COMPUTE_32F,
1221+
CUBLAS_GEMM_DEFAULT_TENSOR_OP));
1222+
1223+
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_BF16);
1224+
to_fp32_cuda(dst_bf16.get(), dst_dd_i, row_diff*src1_ncols, stream);
1225+
} else if (((GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) || GGML_CUDA_CC_IS_AMD(cc)) && use_fp16) {
11981226
// convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
11991227
ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
12001228
if (src0->type != GGML_TYPE_F16) {

0 commit comments

Comments
 (0)