Skip to content

CUDA: fix crash with partial offloading of MoE #13439

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
May 11, 2025
Merged

CUDA: fix crash with partial offloading of MoE #13439

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 8 additions & 2 deletions ggml/src/ggml-cuda/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1909,13 +1909,19 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);

// 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;

bool use_mul_mat_vec = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
&& src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
&& src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
bool use_mul_mat_q = ggml_is_quantized(src0->type)
bool use_mul_mat_q = ggml_is_quantized(src0->type) && !bad_padding_clear
&& src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;

bool any_gpus_with_slow_fp16 = false;
Expand Down
4 changes: 2 additions & 2 deletions ggml/src/ggml-cuda/mmq.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -91,11 +91,11 @@ void ggml_cuda_mul_mat_q(

// If src0 is a temporary compute buffer, clear any potential padding.
if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
GGML_ASSERT(ggml_is_contiguously_allocated(src0));
GGML_ASSERT(!src0->view_src);
const size_t size_data = ggml_nbytes(src0);
const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0);
if (size_alloc > size_data) {
GGML_ASSERT(ggml_is_contiguously_allocated(src0));
GGML_ASSERT(!src0->view_src);
CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream));
}
}
Expand Down
4 changes: 2 additions & 2 deletions ggml/src/ggml-cuda/mmvq.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -515,11 +515,11 @@ void ggml_cuda_mul_mat_vec_q(

// If src0 is a temporary compute buffer, clear any potential padding.
if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
GGML_ASSERT(ggml_is_contiguously_allocated(src0));
GGML_ASSERT(!src0->view_src);
const size_t size_data = ggml_nbytes(src0);
const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0);
if (size_alloc > size_data) {
GGML_ASSERT(ggml_is_contiguously_allocated(src0));
GGML_ASSERT(!src0->view_src);
CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream));
}
}
Expand Down
Loading