mirror
/
whisper.cpp
mirror of https://github.com/ggerganov/whisper.cpp
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

ggml : allow fill node alloc inplace (llama/17870)

This commit is contained in:
Sigbjørn Skjæret 2025-12-09 12:23:47 +01:00 committed by Georgi Gerganov
parent 79d86a5c2c
commit ba463fb577
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
2 changed files with 2 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
ggml/src/ggml-alloc.c
View File

@ -25,6 +25,7 @@ static bool ggml_is_view(const struct ggml_tensor * t) {
// ops that return true for this function must not use restrict pointers for their backend implementations // ops that return true for this function must not use restrict pointers for their backend implementations
bool ggml_op_can_inplace(enum ggml_op op) { bool ggml_op_can_inplace(enum ggml_op op) {
switch (op) { switch (op) {
case GGML_OP_FILL:
case GGML_OP_SCALE: case GGML_OP_SCALE:
case GGML_OP_DIAG_MASK_ZERO: case GGML_OP_DIAG_MASK_ZERO:
case GGML_OP_DIAG_MASK_INF: case GGML_OP_DIAG_MASK_INF:

2
ggml/src/ggml-cuda/fill.cu
View File

@ -4,7 +4,7 @@
#define CUDA_FILL_BLOCK_SIZE 256 #define CUDA_FILL_BLOCK_SIZE 256
template <typename T> template <typename T>
static __global__ void fill_kernel(T * __restrict__ dst, const int64_t k, const T value) { static __global__ void fill_kernel(T * dst, const int64_t k, const T value) {
const int64_t i = (int64_t)blockDim.x * blockIdx.x + threadIdx.x; const int64_t i = (int64_t)blockDim.x * blockIdx.x + threadIdx.x;
if (i >= k) { if (i >= k) {
return; return;