IQ4_NL: 4-bit non-linear quants with blocks of 32 (#5590)

* iq4_nl: squash commits for easier rebase

* Basics (quantize, dequantize)
* CUDA dequantize and dot product
* Slightly faster CUDA dot product (120 t/s)
* Switch to 6-bit scales
* Scalar dot product
* AVX2 dot product
* ARM_NEON dot product
* Works on metal, but still slow
* Slightly better Metal dot product
* Another small Metal improvement
* Metal dot product is getting there
* Faster CUDA dot product
* Add 1/8 ffn_down layers as Q5_K when no imatrix has been provided
* Report the actual bpw
* Add _xs mix that is 4.05 bpw for non-MoE models
* Remove IQ4_XS for now, slightly adjust kvalues_iq4nl
* AVX2 dot product uses Q8_0 instead of Q8_K
* Add to test-backend-ops
* Minor fix
* Also use use Q5_K for attn_output in MoE models
* Fixes after merging latest master
* Switching to blocks of 32
* AVX2 for blocks of 32
* Scaler dot product for blocks of 32
* ARM_NEON dot product for blocks of 32
* Metal kernels for blocks of 32
* Slightly faster Metal kernels

* iq4_nl: Fix after merging with master

* iq4_nl: another fix after merging with master

* Use IQ4_NL instead of Q4_K when using k-quants is not possible

* Fix typo that makes several tests fail

* It was the ggml_vdotq thing missed inside the brackets

---------

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

1 files changed, 13 insertions, 4 deletions

diff --git a/llama.cpp b/llama.cpp
index 4296eca3..3748d5ea 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -2527,6 +2527,7 @@ struct llama_model_loader {
                 case GGML_TYPE_IQ2_XS:  ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS;  break;
                 case GGML_TYPE_IQ3_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS; break;
                 case GGML_TYPE_IQ1_S:   ftype = LLAMA_FTYPE_MOSTLY_IQ1_S;   break;
+                case GGML_TYPE_IQ4_NL:  ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL;  break;
                 default:
                     {
                         LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@@ -2877,6 +2878,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q3_K_XS:return "Q3_K - Extra small";
         case LLAMA_FTYPE_MOSTLY_IQ3_XXS:return "IQ3_XXS - 3.0625 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ1_S  :return "IQ1_S - 1.5625 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ4_NL: return "IQ4_NL - 4.5 bpw";
 
         default: return "unknown, may not work";
     }
@@ -10354,6 +10356,9 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
             new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL && qs.model.hparams.n_gqa() >= 4) {
+            new_type = GGML_TYPE_Q5_K;
+        }
         else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
                 use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
@@ -10406,6 +10411,9 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
                 if (use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
             }
         }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL && !qs.has_imatrix) {
+            if (i_layer < n_layer/8) new_type = GGML_TYPE_Q5_K;
+        }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) {
             new_type = GGML_TYPE_Q5_K;
@@ -10422,7 +10430,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         if (arch != LLM_ARCH_FALCON) {
             if (qs.model.hparams.n_expert == 8) {
                 if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K   || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
-                    ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ||
+                    ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M  || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL  ||
                     ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
                     new_type = GGML_TYPE_Q5_K;
                 }
@@ -10489,8 +10497,8 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
             case GGML_TYPE_IQ2_XS:
             case GGML_TYPE_IQ3_XXS:
             case GGML_TYPE_IQ1_S:
-            case GGML_TYPE_Q2_K: new_type = GGML_TYPE_Q4_0; break;
-            case GGML_TYPE_Q3_K: new_type = GGML_TYPE_Q4_1; break;
+            case GGML_TYPE_Q2_K:
+            case GGML_TYPE_Q3_K: new_type = GGML_TYPE_IQ4_NL; break;
             case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break;
             case GGML_TYPE_Q5_K: new_type = GGML_TYPE_Q5_1; break;
             case GGML_TYPE_Q6_K: new_type = GGML_TYPE_Q8_0; break;
@@ -10531,7 +10539,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_IQ2_XXS: quantized_type = GGML_TYPE_IQ2_XXS; break;
         case LLAMA_FTYPE_MOSTLY_IQ2_XS:  quantized_type = GGML_TYPE_IQ2_XS;  break;
         case LLAMA_FTYPE_MOSTLY_IQ3_XXS: quantized_type = GGML_TYPE_IQ3_XXS; break;
-        case LLAMA_FTYPE_MOSTLY_IQ1_S:   quantized_type = GGML_TYPE_IQ1_S  ; break;
+        case LLAMA_FTYPE_MOSTLY_IQ1_S:   quantized_type = GGML_TYPE_IQ1_S;   break;
+        case LLAMA_FTYPE_MOSTLY_IQ4_NL:  quantized_type = GGML_TYPE_IQ4_NL;  break;
 
         default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
     }