llama : adjust default context size + print warnings

ggml-ci

Author: ggerganov

common/common.h

@@ -155,7 +155,7 @@ struct common_sampler_params {
 
 struct common_params {
     int32_t n_predict             =    -1; // new tokens to predict
-    int32_t n_ctx                 =     0; // context size
+    int32_t n_ctx                 =  4096; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                =     0; // number of tokens to keep from initial prompt

src/llama.cpp

@@ -19440,12 +19440,26 @@ struct llama_context * llama_new_context_with_model(
         cparams.causal_attn = params.attention_type == LLAMA_ATTENTION_TYPE_CAUSAL;
     }
 
-    LLAMA_LOG_INFO("%s: n_ctx      = %u\n",     __func__, cparams.n_ctx);
-    LLAMA_LOG_INFO("%s: n_batch    = %u\n",     __func__, cparams.n_batch);
-    LLAMA_LOG_INFO("%s: n_ubatch   = %u\n",     __func__, cparams.n_ubatch);
-    LLAMA_LOG_INFO("%s: flash_attn = %d\n",     __func__, cparams.flash_attn);
-    LLAMA_LOG_INFO("%s: freq_base  = %.1f\n",   __func__, cparams.rope_freq_base);
-    LLAMA_LOG_INFO("%s: freq_scale = %g\n",     __func__, cparams.rope_freq_scale);
+    const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
+
+    LLAMA_LOG_INFO("%s: n_seq_max     = %u\n",     __func__, cparams.n_seq_max);
+    LLAMA_LOG_INFO("%s: n_ctx         = %u\n",     __func__, cparams.n_ctx);
+    LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n",     __func__, n_ctx_per_seq);
+    LLAMA_LOG_INFO("%s: n_batch       = %u\n",     __func__, cparams.n_batch);
+    LLAMA_LOG_INFO("%s: n_ubatch      = %u\n",     __func__, cparams.n_ubatch);
+    LLAMA_LOG_INFO("%s: flash_attn    = %d\n",     __func__, cparams.flash_attn);
+    LLAMA_LOG_INFO("%s: freq_base     = %.1f\n",   __func__, cparams.rope_freq_base);
+    LLAMA_LOG_INFO("%s: freq_scale    = %g\n",     __func__, cparams.rope_freq_scale);
+
+    if (n_ctx_per_seq < hparams.n_ctx_train) {
+        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
+                __func__, n_ctx_per_seq, hparams.n_ctx_train);
+    }
+
+    if (n_ctx_per_seq > hparams.n_ctx_train) {
+        LLAMA_LOG_WARN("%s: n_ctx_pre_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
+                __func__, n_ctx_per_seq, hparams.n_ctx_train);
+    }
 
     ctx->abort_callback      = params.abort_callback;
     ctx->abort_callback_data = params.abort_callback_data;