A fix for the instruction mode stopping.

Author: John

examples/falcon/falcon_main.cpp

@@ -1185,8 +1185,9 @@ fprintf(stderr, "+------------+-------+-------+-------+-------+---------------+-
         #endif
 
         // end of text token or stopword detected in generated content
-        if ((!embd.empty() && embd.back() == falcon_token_eos() && n_past_total > embd_inp.size()) || stopword_fulfilled) 
+        if ((!embd.empty() && embd.back() == falcon_token_eos() && n_consumed >= embd_inp.size()) || stopword_fulfilled) 
         {
+            // printf("embd.size(): %d, embd.back(): %d, n_past_total: %d, embd_inp.size(): %d, n_consumed: %d, n_remain: %d\n",embd.size(), embd.back(), n_past_total, embd_inp.size() , n_consumed, n_remain);
             if (params.instruct) 
             {
                 is_interacting = true;

libfalcon.cpp

@@ -4554,7 +4554,7 @@ int falcon_eval(
         struct falcon_context * ctx,
            const falcon_token * tokens,
                          falcon_evaluation_config &configuration) {
-    //  fprintf(stderr, "falcon_eval: n_tokens=%d, n_past=%d, n_threads=%d\n", n_tokens, n_past, n_threads);
+    //   fprintf(stderr, "falcon_eval: n_tokens=%d, n_past=%d, n_threads=%d\n", configuration.n_tokens, configuration.n_past, configuration.n_threads);
     // fprintf(stderr, "n_ctx=%d, n_embd=%d, n_head=%d, n_layer=%d, n_vocab=%d\n", ctx->model.hparams.n_ctx, ctx->model.hparams.n_embd, ctx->model.hparams.n_head, ctx->model.hparams.n_layer, ctx->model.hparams.n_vocab);
     FALCON_ASSERT(ctx->model.hparams.n_ctx >= (configuration.n_past+configuration.n_tokens)); // kv buffer overflow
     #if defined(GGML_USE_CUBLAS)