IQ3_S: a much better alternative to Q3_K (llama/5676)

* 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

* Resurrecting iq3_xs

After all the experimentation, nothing was better than this.

* Minor PPL improvement via a block scale fudge factor

* Minor improvement via 3 neighbours

* iq3_xs: working scalar and AVX2 dot products

* iq3_xs: ARM_NEON dot product - works but extremely slow (10 t/s)

* iq3_xs: working Metal implementation

* Adding IQ3_M - IQ3_XS mix with mostly Q4_K

* iiq3_xs: a 3.4375 bpw variant

* iq3_xs: make CUDA work for new version

* iq3_xs: make scalar and AVX2 work for new version

* iq3_s: make ARM_NEON work with new version

* iq3_xs: make new version work on metal

Performance is very similar to Q3_K_S

* iq3_xs: tiny Metal speed improvement

* iq3_xs: tiny Metal speed improvement

* Fix stupid warning

* Q3_K_XS now uses a mix of IQ3_XS and IQ3_XXS

* iq3_xs: rename to iq3_s

* iq3_s: make tests pass

* Move Q3_K_XS mix to 3.25 bpw

* Attempt to fix failing tests

* Another attempt to fix the Windows builds

* Attempt to fix ROCm

* ROCm again

* iq3_s: partial fix for QK_K = 64

* iq3_s: make it work on metal for QK_K = 64

Pleasent surprise: the coding was super-block size independent,
so all it took was to delete some QK_K == 256 guards.

* Will this fix ROCm?

---------

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

ggml.c

@@ -682,6 +682,18 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
+    [GGML_TYPE_IQ3_S] = {
+        .type_name                = "iq3_s",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq3_s),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq3_s,
+        .from_float               = quantize_row_iq3_s,
+        .from_float_reference     = (ggml_from_float_t)quantize_row_iq3_s_reference,
+        .vec_dot                  = ggml_vec_dot_iq3_s_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+    },
     [GGML_TYPE_IQ1_S] = {
         .type_name                = "iq1_s",
         .blck_size                = QK_K,
@@ -2308,6 +2320,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         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_MOSTLY_IQ3_S:         wtype = GGML_TYPE_IQ3_S;    break;
         case GGML_FTYPE_UNKNOWN:              wtype = GGML_TYPE_COUNT; break;
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
     }
@@ -7742,6 +7755,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ3_S:
             {
                 ggml_compute_forward_add_q_f32(params, dst);
             } break;
@@ -8021,6 +8035,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ3_S:
             {
                 ggml_compute_forward_add1_q_f32(params, dst);
             } break;
@@ -8145,6 +8160,7 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ3_S:
         default:
             {
                 GGML_ASSERT(false);
@@ -11043,6 +11059,7 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ3_S:
             {
                 ggml_compute_forward_out_prod_q_f32(params, dst);
             } break;
@@ -11231,6 +11248,7 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ3_S:
         default:
             {
                 GGML_ASSERT(false);
@@ -11433,6 +11451,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ3_S:
             {
                 ggml_compute_forward_get_rows_q(params, dst);
             } break;
@@ -12133,6 +12152,7 @@ static void ggml_compute_forward_alibi(
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ3_S:
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
@@ -12216,6 +12236,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ3_S:
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
@@ -19467,6 +19488,7 @@ void ggml_quantize_init(enum ggml_type type) {
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ1_S:   iq2xs_init_impl(type); break;
         case GGML_TYPE_IQ3_XXS: iq3xs_init_impl(256); break;
+        case GGML_TYPE_IQ3_S:   iq3xs_init_impl(512); break;
         default: // nothing
             break;
     }
@@ -19741,6 +19763,15 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
                 result = quantize_iq3_xxs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
                 GGML_ASSERT(result == row_size * nrows);
             } break;
+        case GGML_TYPE_IQ3_S:
+            {
+                GGML_ASSERT(start % QK_K == 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_iq3_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_IQ1_S:
             {
                 GGML_ASSERT(start % QK_K == 0);