bugfix: fix core dump when enabling MTP for GLM4.5. (#302)

DongheJin · web-flow · commit b05f5efc231d · 2025-10-30T22:30:44.000+08:00
diff --git a/xllm/core/runtime/speculative_worker_impl.cpp b/xllm/core/runtime/speculative_worker_impl.cpp
@@ -260,7 +260,7 @@ std::optional<ForwardOutput> SpeculativeWorkerImpl::step_prefill(
         inputs.micro_inputs[i].input_params.embedding_ids.end());
   }
 
-  if (!inputs.concated_sampling_params.selected_token_idxes.defined()) {
+  if (inputs.concated_sampling_params.selected_token_idxes.defined()) {
     embeddings = embeddings.index_select(
         /*dim=*/0, inputs.concated_sampling_params.selected_token_idxes);
     CHECK_EQ(embeddings.size(0), output.sample_output.next_tokens.size(0));

Original file line number	Diff line number	Diff line change
`@@ -260,7 +260,7 @@ std::optional<ForwardOutput> SpeculativeWorkerImpl::step_prefill(`
`260`	`260`	`inputs.micro_inputs[i].input_params.embedding_ids.end());`
`261`	`261`	`}`
`262`	`262`
`263`		`- if (!inputs.concated_sampling_params.selected_token_idxes.defined()) {`
	`263`	`+ if (inputs.concated_sampling_params.selected_token_idxes.defined()) {`
`264`	`264`	`embeddings = embeddings.index_select(`
`265`	`265`	`/dim=/0, inputs.concated_sampling_params.selected_token_idxes);`
`266`	`266`	`CHECK_EQ(embeddings.size(0), output.sample_output.next_tokens.size(0));`