server: fix bugs when running speculative decoding

README.md

@@ -381,6 +381,7 @@ curl -X POST http://localhost:8080/v1/cancel \
 ```
 
 **9. How to use speculative decoding?**
+
 Please see "[Power prima.cpp with speculative decoding: Further speeds up by up to 80%](https://github.com/Lizonghang/prima.cpp/discussions/29)".
 
 ## ❤️ Acknowledgment

examples/server/server.cpp

@@ -2542,7 +2542,7 @@ struct server_context {
                     llama_batch_add(slot.batch_spec, draft[i], slot.n_past + 1 + i, { slot.id + 1 }, true);
                 }
 
-                llama_decode(ctx, slot.batch_spec);
+                llama_decode(ctx, slot.batch_spec, true);
 
                 // the accepted tokens from the speculation
                 const auto ids = gpt_sampler_sample_and_accept_n(slot.smpl, ctx, draft);