Skip to content

fix(qwen): preserve compact logits indices across graph reuse #426

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
0xShug0 merged 1 commit into from
Sep 3, 2026
Merged

fix(qwen): preserve compact logits indices across graph reuse #426

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
14 changes: 13 additions & 1 deletion src/framework/modules/transformers/qwen_causal_decode_runtime.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -291,10 +291,15 @@ ggml_tensor * make_logits_readback_token_ids(
if (config.logits_readback_token_ids.empty()) {
return nullptr;
}
return ggml_new_tensor_1d(
auto * tensor = ggml_new_tensor_1d(
ctx,
GGML_TYPE_I32,
static_cast<int64_t>(config.logits_readback_token_ids.size()));
// This leaf is populated by the host after graph allocation. Marking it as
// an input keeps the graph allocator from aliasing its storage with a
// temporary that executes before ggml_get_rows consumes the token ids.
ggml_set_input(tensor);
return tensor;
}

core::TensorValue wrap_logits_readback_token_ids(
Expand Down Expand Up @@ -784,6 +789,9 @@ class QwenCausalDecodeRuntime::Impl {
prefill_attention_mask_values_.data(),
0,
prefill_attention_mask_values_.size() * sizeof(ggml_fp16_t));
if (prefill_logits_readback_token_ids_ != nullptr) {
upload_logits_readback_token_ids(prefill_logits_readback_token_ids_, config_);
}
core::set_backend_threads(backend_, threads_);
const ggml_status status = core::compute_backend_graph(backend_, prefill_graph_);
ggml_backend_synchronize(backend_);
Expand Down Expand Up @@ -970,6 +978,10 @@ class QwenCausalDecodeRuntime::Impl {
batched_prefill_attention_mask_values_.data(),
0,
batched_prefill_attention_mask_values_.size() * sizeof(ggml_fp16_t));
if (batched_prefill_logits_readback_token_ids_ != nullptr) {
upload_logits_readback_token_ids(
batched_prefill_logits_readback_token_ids_, config_);
}
core::set_backend_threads(backend_, threads_);
const ggml_status status = core::compute_backend_graph(backend_, batched_prefill_graph_);
ggml_backend_synchronize(backend_);
Expand Down
Loading