LLM-Inference/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-08-12 14:24:03 +04:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: clear MMQ row padding on partially offloaded quantized weights (#2292)

Browse Source 
Co-authored-by: Joel Farthing <262452229+joelfarthing@users.noreply.github.com>
This commit is contained in:
Joel Farthing
2026-08-11 09:07:01 +02:00
committed by GitHub
co-authored by Joel Farthing
parent b382ebd848
commit 26ceed9d40
1 changed files with 10 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda.cu
+10 -3
View File
@@ -2042,7 +2042,7 @@ static void ggml_cuda_op_mul_mat(
// 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_data = ggml_nbytes(src0);
const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
}
@@ -2647,8 +2647,11 @@ static int ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
// If src0 is a temporary compute buffer it may have some padding that needs to be cleared for mul_mat_vec_q or mul_mat_q.
// But if src0 is also a view of another tensor then this cannot be done safely because it may overwrite valid tensor data.
// Therefore, in such cases use cuBLAS.
const bool bad_padding_clear = ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE
&& ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) && src0->view_src;
const size_t src0_nbytes = ggml_nbytes(src0);
const size_t src0_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0);
const bool src0_padded = ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE
&& src0_nbytes != src0_alloc;
const bool bad_padding_clear = src0_padded && src0->view_src;
bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
@@ -2672,6 +2675,10 @@ static int ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
if ((use_mul_mat_vec_q || use_mul_mat_q) && src1->ne[2]*src1->ne[3] == 1) {
// This return does not go through ggml_cuda_op_mul_mat, which is where the padding is otherwise cleared.
if (src0_padded) {
CUDA_CHECK(cudaMemsetAsync((char *) src0->data + src0_nbytes, 0, src0_alloc - src0_nbytes, ctx.stream()));
}
return ggml_cuda_mul_mat_q(ctx, src0, src1, dst, cgraph, node_n, use_mul_mat_vec_q);
}