update

crates/llama-cpp-bindings/src/engine.cc

@@ -17,6 +17,9 @@ namespace {
 constexpr size_t N_BATCH = 512;  // # per batch inference.
 constexpr size_t N_CTX = 4096;   // # max kv history.
 
+constexpr size_t DRAFT_N_GRAM_SIZE = 3;
+constexpr size_t DRAFT_N_PRED_TOKENS = 10;
+ 
 struct Request {
   Request(size_t request_id, std::vector<llama_token> input_token_ids) :
     id(request_id),
@@ -34,7 +37,29 @@ struct Request {
   std::string generated_text;
 
 
-  std::vector<llama_token> find_candidate_pred_tokens(size_t ngram_size = 3, size_t n_pred_tokens = 8) {
+  void draft_tokens(int n_draft_quota) {
+    if (n_draft_quota < DRAFT_N_PRED_TOKENS) {
+      n_draft = 0;
+      return;
+    }
+
+    auto draft = find_candidate_pred_tokens(DRAFT_N_GRAM_SIZE, DRAFT_N_PRED_TOKENS);
+    n_draft = draft.size();
+    tokens.insert(tokens.end(), draft.begin(), draft.end());
+  }
+
+  size_t n_past() {
+    return past_tokens.size();
+  }
+
+  void step(llama_token next_token, size_t n_dropped) {
+    past_tokens.insert(past_tokens.end(), tokens.begin(), tokens.end() - n_dropped);
+    tokens.clear();
+    tokens.push_back(next_token);
+  }
+
+ private:
+  std::vector<llama_token> find_candidate_pred_tokens(size_t ngram_size, size_t n_pred_tokens) {
     if (past_tokens.size() < ngram_size) return {};
     std::vector<llama_token> ngram(past_tokens.begin() + past_tokens.size() - ngram_size, past_tokens.end());
 
@@ -46,17 +71,6 @@ struct Request {
     return std::vector<llama_token>(begin, begin + n_pred_tokens);
   }
 
-  size_t n_past() {
-    return past_tokens.size();
-  }
-
-  void step(llama_token next_token, size_t n_dropped) {
-    past_tokens.insert(past_tokens.end(), tokens.begin(), tokens.begin() + tokens.size() - n_dropped);
-    tokens.clear();
-    tokens.push_back(next_token);
-  }
-
- private:
   std::vector<llama_token> past_tokens;
 };
 
@@ -197,6 +211,11 @@ class TextInferenceEngineImpl : public TextInferenceEngine {
     // Insert tokens from ongoing requests to batch.
     for (auto& request : requests_) {
       const size_t n_tokens = batch_.n_tokens;
+
+      // Ensure the draft logits always fall into the same batch.
+      const int n_draft_quota = N_BATCH - (n_tokens + request.tokens.size()) % N_BATCH;
+      request.draft_tokens(n_draft_quota);
+
       for (size_t i = 0; i < request.tokens.size(); ++i) {
         batch_.token[n_tokens + i] = request.tokens[i];
         batch_.pos[n_tokens + i] = request.n_past() + i;
@@ -241,12 +260,9 @@ class TextInferenceEngineImpl : public TextInferenceEngine {
           continue;
         }
 
-        llama_token next_token = -1;
-        int k = -request.n_draft;
-        // FIXME: ensure batching logic always put i_batch - request.n_draft in this batch.
-        for (k = -request.n_draft; k < 1; ++k) {
+        for (int k = -request.n_draft; k < 1; ++k) {
           auto logits = llama_get_logits_ith(ctx, i_batch + k);
-          next_token = std::distance(logits, std::max_element(logits, logits + n_vocab));
+          llama_token next_token = std::distance(logits, std::max_element(logits, logits + n_vocab));
 
           const auto token_str = llama_token_to_piece(ctx, next_token);
           request.generated_text += token_str;
@@ -296,10 +312,6 @@ class TextInferenceEngineImpl : public TextInferenceEngine {
             break;
           }
         }
-
-        auto draft_tokens = request.find_candidate_pred_tokens();
-        request.n_draft = draft_tokens.size();
-        request.tokens.insert(request.tokens.end(), draft_tokens.begin(), draft_tokens.end());
       }
     }