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, 30 insertions, 0 deletions

diff --git a/ggml.c b/ggml.c
index d129df50..91adbb0a 100644
--- a/ggml.c
+++ b/ggml.c
@@ -690,6 +690,18 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
+    [GGML_TYPE_IQ4_NL] = {
+        .type_name                = "iq4_nl",
+        .blck_size                = QK4_NL,
+        .type_size                = sizeof(block_iq4_nl),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq4_nl,
+        .from_float               = quantize_row_iq4_nl,
+        .from_float_reference     = (ggml_from_float_t)quantize_row_iq4_nl_reference,
+        .vec_dot                  = ggml_vec_dot_iq4_nl_q8_0,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+    },
     [GGML_TYPE_Q8_K] = {
         .type_name                = "q8_K",
         .blck_size                = QK_K,
@@ -2291,6 +2303,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_IQ2_XS:        wtype = GGML_TYPE_IQ2_XS;   break;
         case GGML_FTYPE_MOSTLY_IQ3_XXS:       wtype = GGML_TYPE_IQ3_XXS;  break;
         case GGML_FTYPE_MOSTLY_IQ1_S:         wtype = GGML_TYPE_IQ1_S;    break;
+        case GGML_FTYPE_MOSTLY_IQ4_NL:        wtype = GGML_TYPE_IQ4_NL;   break;
         case GGML_FTYPE_UNKNOWN:              wtype = GGML_TYPE_COUNT; break;
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
     }
@@ -7702,6 +7715,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ4_NL:
             {
                 ggml_compute_forward_add_q_f32(params, src0, src1, dst);
             } break;
@@ -7970,6 +7984,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ4_NL:
             {
                 ggml_compute_forward_add1_q_f32(params, src0, src1, dst);
             } break;
@@ -8091,6 +8106,7 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ4_NL:
         default:
             {
                 GGML_ASSERT(false);
@@ -10858,6 +10874,7 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ4_NL:
             {
                 ggml_compute_forward_out_prod_q_f32(params, src0, src1, dst);
             } break;
@@ -11039,6 +11056,7 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ4_NL:
         default:
             {
                 GGML_ASSERT(false);
@@ -11237,6 +11255,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ4_NL:
             {
                 ggml_compute_forward_get_rows_q(params, src0, src1, dst);
             } break;
@@ -11911,6 +11930,7 @@ static void ggml_compute_forward_alibi(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
@@ -11989,6 +12009,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
@@ -19455,6 +19476,15 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
                 result = quantize_iq1_s(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
                 GGML_ASSERT(result == row_size * nrows);
             } break;
+        case GGML_TYPE_IQ4_NL:
+            {
+                GGML_ASSERT(start % QK4_NL == 0);
+                GGML_ASSERT(start % n_per_row == 0);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_iq4_nl(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
+            } break;
         case GGML_TYPE_F16:
             {
                 size_t elemsize = sizeof(ggml_fp16_t);