ggml : importance matrix support for legacy quants (#4969)

* imatrix: adding support for legacy quants

* imatrix: guard Q4_0/Q5_0 against ffn_down craziness

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

1 files changed, 10 insertions, 0 deletions

diff --git a/llama.cpp b/llama.cpp
index 46c4d11c..765d20dd 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -8374,6 +8374,8 @@ struct quantize_state_internal {
     int n_k_quantized     = 0;
     int n_fallback        = 0;
 
+    bool has_imatrix      = false;
+
     quantize_state_internal(const llama_model & model, const llama_model_quantize_params * params)
         : model(model)
         , params(params)
@@ -8546,6 +8548,13 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) {
             new_type = GGML_TYPE_Q5_K;
         }
+        else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_0 || ftype == LLAMA_FTYPE_MOSTLY_Q5_0)
+                && qs.has_imatrix && i_layer < n_layer/8) {
+            // Guard against craziness in the first few ffn_down layers that can happen even with imatrix for Q4_0/Q5_0.
+            // We only do it when an imatrix is provided because a) we want to make sure that one can always get the
+            // same quantization as before imatrix stuff, and b) Q4_1/Q5_1 do go crazy on ffn_down without an imatrix.
+            new_type = ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q5_1;
+        }
         ++qs.i_feed_forward_w2;
     } else if (name.find("attn_output.weight") != std::string::npos) {
         if (arch != LLM_ARCH_FALCON) {
@@ -8669,6 +8678,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         imatrix_data = static_cast<const std::unordered_map<std::string, std::vector<float>>*>(params->imatrix);
         if (imatrix_data) {
             LLAMA_LOG_INFO("================================ Have weights data with %d entries\n",int(imatrix_data->size()));
+            qs.has_imatrix = true;
         }
     }