refactor: switch back to llama batch interface

TabbyML · Oct 28, 2023 · 2a23daf · 2a23daf
1 parent 43cc5f3
commit 2a23daf
Showing 1 changed file with 19 additions and 2 deletions.
diff --git a/crates/llama-cpp-bindings/src/engine.cc b/crates/llama-cpp-bindings/src/engine.cc
@@ -21,6 +21,11 @@ class TextInferenceEngineImpl : public TextInferenceEngine {
   TextInferenceEngineImpl(owned<llama_model> model, owned<llama_context> ctx) :
     model_(std::move(model)),
     ctx_(std::move(ctx)) {
+      batch_ = llama_batch_init(N_BATCH, 0, 1);
+  }
+
+  ~TextInferenceEngineImpl() override {
+    llama_batch_free(batch_);
   }
 
   void start(rust::Slice<const uint32_t> input_token_ids) override {
@@ -52,7 +57,7 @@ class TextInferenceEngineImpl : public TextInferenceEngine {
   uint32_t sample() const {
     auto* ctx = ctx_.get();
 
-    auto logits = llama_get_logits_ith(ctx, 0);
+    auto logits = llama_get_logits_ith(ctx, batch_.n_tokens - 1);
     auto n_vocab = llama_n_vocab(llama_get_model(ctx));
 
     // Greedy sampling (always select the highest logit).
@@ -64,9 +69,19 @@ class TextInferenceEngineImpl : public TextInferenceEngine {
       n_past_ = 0;
     }
 
+    batch_.n_tokens = size;
+    for (size_t i = 0; i < size; ++i) {
+      batch_.token[i] = data[i];
+      batch_.pos[i] = n_past_ + i;
+      batch_.n_seq_id[i] = 1;
+      batch_.seq_id[i][0] = 0;
+      batch_.logits[i] = false;
+    }
+    batch_.logits[size - 1] = true;
+
     auto* ctx = ctx_.get();
     llama_kv_cache_tokens_rm(ctx, n_past_, -1);
-    if (llama_decode(ctx, llama_batch_get_one(data, size, n_past_, 0))) {
+    if (llama_decode(ctx, batch_)) {
       throw std::runtime_error("Failed to eval");
     }
 
@@ -76,6 +91,8 @@ class TextInferenceEngineImpl : public TextInferenceEngine {
   size_t n_past_;
   owned<llama_model> model_;
   owned<llama_context> ctx_;
+
+  llama_batch batch_;
 };
 
 static int g_llama_cpp_log_level = 0;