Simplify IQ4_XSR

But leave in place as a "demo" the more complex template set by Ikawrakow to customize the layers quants, with the added attn_q, attn_k, and attn_output tensors.

Author: Nexesenex

src/llama.cpp

@@ -15508,7 +15508,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 || new_type == GGML_TYPE_Q4_0_8_8) {
                 new_type = GGML_TYPE_Q4_0;
             }
-            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_XSR) new_type = GGML_TYPE_Q8_0;
+            else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_XSR) new_type = GGML_TYPE_IQ4_XS;
         }
     } else if (name.find("attn_v.weight") != std::string::npos) {
         if (qs.model.hparams.n_expert >= 4) {
@@ -15568,8 +15568,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_XSR) {
             if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) {
-                  new_type = qs.i_attention_wv < qs.n_attention_wv/8 ? GGML_TYPE_Q6_K :
-                             use_more_bits(qs.i_attention_wv, qs.n_attention_wv) ? GGML_TYPE_Q5_K : GGML_TYPE_Q5_K;
+                  new_type = qs.i_attention_wv < qs.n_attention_wv/8 ? GGML_TYPE_Q5_K :
+                             use_more_bits(qs.i_attention_wv, qs.n_attention_wv) ? GGML_TYPE_Q6_K : GGML_TYPE_Q5_K;
             }
         }
         ++qs.i_attention_wv;
@@ -15622,7 +15622,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_XSR) {
             if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) {
                   new_type = qs.i_attention_wk < qs.n_attention_wk/8 ? GGML_TYPE_Q5_K :
-                             use_more_bits(qs.i_attention_wk, qs.n_attention_wk) ? GGML_TYPE_IQ4_XS : GGML_TYPE_IQ4_XS;
+                             use_more_bits(qs.i_attention_wk, qs.n_attention_wk) ? GGML_TYPE_Q5_K : GGML_TYPE_Q5_K;
             }
         }
         ++qs.i_attention_wk;
@@ -15704,7 +15704,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_XSR) {
             if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) {
-                 new_type = i_layer < n_layer/8 ? GGML_TYPE_Q5_K :
+                 new_type = i_layer < n_layer/8 ? GGML_TYPE_IQ4_XS :
                             use_more_bits(i_layer, n_layer) ? GGML_TYPE_IQ4_XS : GGML_TYPE_IQ4_XS;
             }
         }
@@ -15737,7 +15737,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
                 else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XL) new_type = GGML_TYPE_IQ4_XS;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_XSR) {
                     if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) {
-                        new_type = qs.i_attention_wo < qs.n_attention_wo/8 ? GGML_TYPE_Q5_K :
+                        new_type = qs.i_attention_wo < qs.n_attention_wo/8 ? GGML_TYPE_IQ4_XS :
                                    use_more_bits(qs.i_attention_wo, qs.n_attention_wo) ? GGML_TYPE_IQ4_XS : GGML_TYPE_IQ4_XS;
                     }
                 }
@@ -15783,7 +15783,6 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XL && (use_more_bits(i_layer, n_layer))) new_type = GGML_TYPE_IQ3_XXS;
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (use_more_bits(i_layer, n_layer))) new_type = GGML_TYPE_IQ3_S;
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XL && (use_more_bits(i_layer, n_layer))) new_type = GGML_TYPE_IQ4_XS;
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_XSR && (i_layer < n_layer/8)) new_type = GGML_TYPE_Q5_K;
         ++qs.i_ffn_gate;
     }
     else if (name.find("ffn_up") != std::string::npos) {
@@ -15799,7 +15798,6 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XL && (use_more_bits(i_layer, n_layer))) new_type = GGML_TYPE_IQ3_XXS;
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (use_more_bits(i_layer, n_layer))) new_type = GGML_TYPE_IQ3_S;
         else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XL && (use_more_bits(i_layer, n_layer))) new_type = GGML_TYPE_IQ4_XS;
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_XSR && (i_layer < n_layer/8)) new_type = GGML_TYPE_Q5_K;
         ++qs.i_ffn_up;
     }