ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-11 08:50:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

Added Johannes' changes, still getting NaNs with quantized k-cache.

Browse Source 
Also getting NaN's on Johannes's mainline branch.
This commit is contained in:
Iwan Kawrakow
2024-10-24 10:44:43 +03:00
parent b61cf7d0d7
commit d64d602faa
2 changed files with 10 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
ggml/src/ggml-cuda.cu
View File

@@ -1170,8 +1170,8 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
GGML_ASSERT(ggml_backend_buffer_is_cuda(src->buffer));
char * src_ptr = (char *) src->data;
char * dst_ptr = (char *) dst;
const char * src_ptr = (const char *) src->data;
char * dst_ptr = (char *) dst;
const int64_t ne0 = src->ne[0];
const int64_t nb0 = src->nb[0];
@@ -1182,7 +1182,7 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
const int64_t ts = ggml_type_size(type);
const int64_t rs = ggml_row_size(type, ne0);
const int64_t bs = ggml_blck_size(type);
int64_t i1_diff = i1_high - i1_low;
const int64_t i1_diff = i1_high - i1_low;
const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3;
if (nb0 == ts && nb1 == rs) {
@@ -1532,10 +1532,14 @@ static void ggml_cuda_op_mul_mat(
if (src0_is_contiguous) {
dev[id].src0_dd = split ? (char *) src0_extra->data_device[id] : (char *) src0->data;
} else {
dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), ggml_nbytes(src0));
// If src0 is not contiguous it will be copied to a temporary buffer, it may then be necessary to clear padding.
const size_t nbytes_data = ggml_nbytes(src0);
const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), nbytes_data + nbytes_padding);
CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
}
// If src0 is on a temporary compute buffers (partial offloading) there may be some padding that needs to be cleared:
// If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared:
if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
const int64_t nbytes_data = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);

4
ggml/src/ggml-cuda/mmq.cu
View File

@@ -8,8 +8,6 @@ void ggml_cuda_op_mul_mat_q(
const int64_t ne00 = src0->ne[0];
const int64_t nb01 = src0->nb[1];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
GGML_ASSERT(ne10 % QK8_1 == 0);
@@ -17,7 +15,7 @@ void ggml_cuda_op_mul_mat_q(
const int64_t ne0 = dst->ne[0];
const int64_t row_diff = row_high - row_low;
const int64_t stride00 = nb01 / ggml_type_size(src0->type);
const int64_t stride00 = ne00 / ggml_blck_size(src0->type);
int id = ggml_cuda_get_device();
const int compute_capability = ggml_cuda_info().devices[id].cc;