Skip to content

Commit

Permalink
cuda : fix dequantize kernel names (ggerganov#4938)
Browse files Browse the repository at this point in the history
  • Loading branch information
ggerganov committed Jan 15, 2024
1 parent 2faaef3 commit ddb008d
Showing 1 changed file with 6 additions and 6 deletions.
12 changes: 6 additions & 6 deletions ggml-cuda.cu
Original file line number Diff line number Diff line change
Expand Up @@ -6309,14 +6309,14 @@ static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int k, cu
}

template<typename dst_t>
static void dequantize_q4_0_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
const int nb32 = k / 32;
const int nb = (k + 255) / 256;
dequantize_block_q4_0<<<nb, 32, 0, stream>>>(vx, y, nb32);
}

template<typename dst_t>
static void dequantize_q4_1_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
const int nb32 = k / 32;
const int nb = (k + 255) / 256;
dequantize_block_q4_1<<<nb, 32, 0, stream>>>(vx, y, nb32);
Expand Down Expand Up @@ -6370,9 +6370,9 @@ static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
int id;
switch (type) {
case GGML_TYPE_Q4_0:
return dequantize_q4_0_cuda;
return dequantize_row_q4_0_cuda;
case GGML_TYPE_Q4_1:
return dequantize_q4_1_cuda;
return dequantize_row_q4_1_cuda;
case GGML_TYPE_Q5_0:
return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
case GGML_TYPE_Q5_1:
Expand Down Expand Up @@ -6407,9 +6407,9 @@ static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
return dequantize_q4_0_cuda;
return dequantize_row_q4_0_cuda;
case GGML_TYPE_Q4_1:
return dequantize_q4_1_cuda;
return dequantize_row_q4_1_cuda;
case GGML_TYPE_Q5_0:
return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
case GGML_TYPE_Q5_1:
Expand Down

0 comments on commit ddb008d

Please sign in to comment.