Adding IQ4_KSS: 4.0 bpw quants (#89)

* iq4_kss: WIP

* iq4_kss: CUDA dequantize works

So we can run perplexity. Sadly, the result does not look good
on the bpw vs quantization error plot.

* iq4_kss: slightly better quantization

* iq4_kss: another small quantization improvement

* iq4_kss: CUDA works

TG-128 performance is very decent with 131 t/s for LLaMA-3.1-8B.
In comparison, we have 123 t/s for q4_0 and 128 t/s for iq4_ks.
I.e., the reduced model size more than offsets the additional
bit fiddling required for iq4_kss.

* iq4_kss: new bit arrangement - CUDA and Zen4 work

Did not lose performance on CUDA. Zen4 is decent, but not great:
PP-512(LLaMA-3.1-8B) = 163 t/s.
TG-128 is of course better than other 4-bit quants due to smaller model size.
We get 14.5 t/s @ 8 threads.

* iq4_kss: ARM_NEON. Predictably very slow

* iq4_kss: Metal

PP is not too bad - just 10% slower than q4_0.
But TG is 30% slower, i.e., predictably bad.

* iq4_kss: somewhat faster Metal dot product

45.75 t/s -> 48.75 t/s.
Still 22% slower than q4_0

* iq4_kss: AVX2

Bad, but better than I expected.
PP-512(LLaMA-3.1-8B) = 167 t/s on the Ryzen-5950X.
I.e., with 32 AVX2 threads we get the performance of
16 Zen4 threads.

* iq4_kss: very slightly faster Metal dot product

48.7 t/s -> 49.3 t/s

---------

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

1 files changed, 22 insertions, 0 deletions

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index a9f795ae..35ed68d0 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1100,6 +1100,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
         .row_meta_size            = 4,
     },
+    [GGML_TYPE_IQ4_KSS] = {
+        .type_name                = "iq4_kss",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq4_kss),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq4_kss,
+        .from_float               = quantize_row_iq4_kss,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq4_kss_ref,
+        .vec_dot                  = vec_dot_iq4_kss_q8_k,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+        .row_meta_size            = 4,
+    },
     [GGML_TYPE_Q8_K] = {
         .type_name                = "q8_K",
         .blck_size                = QK_K,
@@ -3918,6 +3931,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_IQ4_NL:        wtype = GGML_TYPE_IQ4_NL;   break;
         case GGML_FTYPE_MOSTLY_IQ4_XS:        wtype = GGML_TYPE_IQ4_XS;   break;
         case GGML_FTYPE_MOSTLY_IQ4_KS:        wtype = GGML_TYPE_IQ4_KS;   break;
+        case GGML_FTYPE_MOSTLY_IQ4_KSS:       wtype = GGML_TYPE_IQ4_KSS;  break;
         case GGML_FTYPE_MOSTLY_IQ2_K:         wtype = GGML_TYPE_IQ2_K;    break;
         case GGML_FTYPE_MOSTLY_IQ2_KS:        wtype = GGML_TYPE_IQ2_KS;   break;
         case GGML_FTYPE_MOSTLY_IQ3_K:         wtype = GGML_TYPE_IQ3_K;    break;
@@ -10419,6 +10433,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
+        case GGML_TYPE_IQ4_KSS:
         case GGML_TYPE_IQ2_K:
         case GGML_TYPE_IQ2_KS:
         case GGML_TYPE_IQ3_K:
@@ -10809,6 +10824,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
+        case GGML_TYPE_IQ4_KSS:
         case GGML_TYPE_IQ2_K:
         case GGML_TYPE_IQ2_KS:
         case GGML_TYPE_IQ3_K:
@@ -10949,6 +10965,7 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
+        case GGML_TYPE_IQ4_KSS:
         case GGML_TYPE_IQ2_K:
         case GGML_TYPE_IQ2_KS:
         case GGML_TYPE_IQ3_K:
@@ -14135,6 +14152,7 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
+        case GGML_TYPE_IQ4_KSS:
         case GGML_TYPE_IQ2_K:
         case GGML_TYPE_IQ2_KS:
         case GGML_TYPE_IQ3_K:
@@ -14515,6 +14533,7 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
+        case GGML_TYPE_IQ4_KSS:
         case GGML_TYPE_IQ2_K:
         case GGML_TYPE_IQ2_KS:
         case GGML_TYPE_IQ3_K:
@@ -14789,6 +14808,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
+        case GGML_TYPE_IQ4_KSS:
         case GGML_TYPE_IQ2_K:
         case GGML_TYPE_IQ2_KS:
         case GGML_TYPE_IQ3_K:
@@ -15390,6 +15410,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
+        case GGML_TYPE_IQ4_KSS:
         case GGML_TYPE_IQ2_K:
         case GGML_TYPE_IQ2_KS:
         case GGML_TYPE_IQ3_K:
@@ -22208,6 +22229,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_IQ4_NL:  result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_XS:  result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_KS:  result = quantize_iq4_ks (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ4_KSS: result = quantize_iq4_kss(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_K:   result = quantize_iq2_k  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_KS:  result = quantize_iq2_ks (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ3_K:   result = quantize_iq3_k  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;