Correction too small tensor embeddings to quantize

IQ2_XS doesn't seem to work as such, back to IQ2_S
ggerganov · Aug 21, 2024 · 644aa9f · 644aa9f
1 parent 32f6ead
commit 644aa9f
Showing 1 changed file with 2 additions and 5 deletions.
diff --git a/src/llama.cpp b/src/llama.cpp
@@ -15964,11 +15964,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
                 new_type = GGML_TYPE_Q4_K;
                 }
             }
-            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
-                if (qs.model.hparams.n_vocab >= 127999) new_type = GGML_TYPE_IQ2_XS;
-                else new_type = GGML_TYPE_IQ2_S;
-            }
-            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XL || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M ||
+                     ftype == LLAMA_FTYPE_MOSTLY_IQ1_XL || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
                 new_type = GGML_TYPE_IQ2_S;
             }
             else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {