Skip to content

Fix attention size computation error in OpenVINO backend for LLM #131

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
wine99 merged 1 commit into from
Apr 13, 2026
+15 −4
Merged

Fix attention size computation error in OpenVINO backend for LLM #131

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
11 changes: 10 additions & 1 deletion ggml/src/ggml-openvino/ggml-decoder.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -286,7 +286,7 @@ std::pair<ModelParams, ComputeParams> GgmlOvDecoder::compute_llm_params(ggml_cgr
for (int i = 0; i < cgraph->n_nodes; i++) {
auto * node = cgraph->nodes[i];
std::string name = std::string(node->name);
if (node->op == GGML_OP_FLASH_ATTN_EXT || (node->op == GGML_OP_SOFT_MAX && node->src[1] != nullptr)) {
if (node->op == GGML_OP_FLASH_ATTN_EXT || (node->op == GGML_OP_SOFT_MAX && node->src[1] != nullptr && node->src[0]->src[1] != nullptr)) {
compute_params.input_len = node->src[0]->ne[1];

auto * q_perm = node->src[0];
Expand Down Expand Up @@ -342,6 +342,15 @@ std::pair<ModelParams, ComputeParams> GgmlOvDecoder::compute_llm_params(ggml_cgr
compute_params.token_len_per_seq = 1;
}
}

if (node->op == GGML_OP_MUL_MAT && node->src[0]->op == GGML_OP_PERMUTE &&
node->src[0]->src[0]->op == GGML_OP_VIEW && is_kvcache(node->src[0]->view_src, node->view_src)) {
if (node->src[1]->op == GGML_OP_PERMUTE && node->src[1]->src[0]->op == GGML_OP_VIEW &&
node->src[1]->src[0]->src[0]->op == GGML_OP_ROPE) {
compute_params.attention_size = node->ne[0];
}
}

// if the node op is TRANSPOSE and its input is PERMUTE and the source of the PERMUTE is VIEW, then get the attention size with the TRANSPOSE node ne[0] (in case no GGML_OP_FLASH_ATTN_EXT)
if (node->op == GGML_OP_TRANSPOSE && node->src[0]->op == GGML_OP_PERMUTE &&
node->src[0]->src[0]->op == GGML_OP_VIEW) {
Expand Down
4 changes: 2 additions & 2 deletions ggml/src/ggml-openvino/ggml-decoder.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -248,8 +248,8 @@ class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder {
}

inline static bool is_kvcache(const ggml_tensor * tensor, const ggml_tensor * op) {
return (op->op == GGML_OP_SET_ROWS && op->src[2] == tensor) ||
tensor->buffer->usage == GGML_BACKEND_BUFFER_USAGE_ANY;
return tensor->buffer->usage == GGML_BACKEND_BUFFER_USAGE_ANY ||
(op->op == GGML_OP_SET_ROWS && op->src[2] == tensor);
}

inline static bool is_kv_idx(const ggml_tensor * tensor, const ggml_tensor * op) {
Expand Down
4 changes: 3 additions & 1 deletion ggml/src/ggml-openvino/openvino/translate_session.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -146,7 +146,9 @@ void add_rope_sin_cos(TensorMap & tensor_map, GgmlDecoder & ggml_model_decoder)

// Create common patterns
void preprocess(TensorMap & tensor_map, GgmlDecoder & ggml_model_decoder) {
add_sliced_mask(tensor_map, ggml_model_decoder);
if (ggml_model_decoder.is_stateful()) {
add_sliced_mask(tensor_map, ggml_model_decoder);
}
add_rope_sin_cos(tensor_map, ggml_model_decoder);
}

Expand Down