IQ1_S_R4: better 1.5 bpw quants (#185)

* iq1_s_r4: basics - quantize/dequantize

* iq1_s_r4: gemm/gemv works on AVX2/Zen4

* Don't forget to make sure we have a multiple of 4 rows per thread

* iq1_s_r4: this is better

* iq1_s_r4: fix Zen4 after AVX2 changes

* iq1_s_r4: NEON gemm/gemv

* iq1_s_r4: more bits for shared experts

With this mix we arrive at PPL(512) = 9.4140
for Deepseek-Lite using 1.766 bpw for the repeating layers.

On the Ryzen-7950X we get PP-512 = 494 t/s and
TG-128 = 52 t/s @ 16 threads.

* Forgotten counter increment

* iq1_s_r4: slightly faster AVX2/Zen4 gemm/gemv

* Compiler warnings

---------

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

1 files changed, 25 insertions, 2 deletions

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index b3c8a951..64b7d3ce 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1176,13 +1176,26 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_iq1_s),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_iq1_s,
-        .from_float               = NULL,
-        .from_float_ref           = NULL,
+        .from_float               = quantize_row_iq1_s,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq1_s_ref,
         .vec_dot                  = ggml_vec_dot_iq1_s_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
         .row_meta_size            = 0,
     },
+    [GGML_TYPE_IQ1_S_R4] = {
+        .type_name                = "iq1_s_r4",
+        .blck_size                = 32,
+        .type_size                = sizeof(block_iq1_s_r4)/4,
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq1_s_r4,
+        .from_float               = quantize_row_iq1_s_r4,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq1_s_r4_ref,
+        .vec_dot                  = vec_dot_iq1_s_r4_q8_k,
+        .vec_dot_type             = GGML_TYPE_Q8_1_X4,
+        .nrows                    = 1,
+        .row_meta_size            = 2,
+    },
     [GGML_TYPE_IQ1_M] = {
         .type_name                = "iq1_m",
         .blck_size                = QK_K,
@@ -4387,6 +4400,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_IQ3_S_R4:      wtype = GGML_TYPE_IQ3_S_R4; break;
         case GGML_FTYPE_MOSTLY_IQ2_S:         wtype = GGML_TYPE_IQ2_S;    break;
         case GGML_FTYPE_MOSTLY_IQ2_S_R4:      wtype = GGML_TYPE_IQ2_S_R4; break;
+        case GGML_FTYPE_MOSTLY_IQ1_S_R4:      wtype = GGML_TYPE_IQ1_S_R4; break;
         case GGML_FTYPE_MOSTLY_Q4_0_4_4:      wtype = GGML_TYPE_Q4_0_4_4; break;
         case GGML_FTYPE_MOSTLY_Q4_0_4_8:      wtype = GGML_TYPE_Q4_0_4_8; break;
         case GGML_FTYPE_MOSTLY_Q4_0_8_8:      wtype = GGML_TYPE_Q4_0_8_8; break;
@@ -10934,6 +10948,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ3_S_R4:
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ2_S_R4:
+        case GGML_TYPE_IQ1_S_R4:
         case GGML_TYPE_Q4_0_4_4:
         case GGML_TYPE_Q4_0_4_8:
         case GGML_TYPE_Q4_0_8_8:
@@ -11402,6 +11417,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ3_S_R4:
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ2_S_R4:
+        case GGML_TYPE_IQ1_S_R4:
         case GGML_TYPE_Q4_0_4_4:
         case GGML_TYPE_Q4_0_4_8:
         case GGML_TYPE_Q4_0_8_8:
@@ -11567,6 +11583,7 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ3_S_R4:
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ2_S_R4:
+        case GGML_TYPE_IQ1_S_R4:
         case GGML_TYPE_Q4_0_4_4:
         case GGML_TYPE_Q4_0_4_8:
         case GGML_TYPE_Q4_0_8_8:
@@ -14805,6 +14822,7 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ3_S_R4:
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ2_S_R4:
+        case GGML_TYPE_IQ1_S_R4:
         case GGML_TYPE_Q4_0_4_4:
         case GGML_TYPE_Q4_0_4_8:
         case GGML_TYPE_Q4_0_8_8:
@@ -15210,6 +15228,7 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ3_S_R4:
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ2_S_R4:
+        case GGML_TYPE_IQ1_S_R4:
         case GGML_TYPE_Q4_0_4_4:
         case GGML_TYPE_Q4_0_4_8:
         case GGML_TYPE_Q4_0_8_8:
@@ -15509,6 +15528,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ3_S_R4:
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ2_S_R4:
+        case GGML_TYPE_IQ1_S_R4:
         case GGML_TYPE_Q4_0_4_4:
         case GGML_TYPE_Q4_0_4_8:
         case GGML_TYPE_Q4_0_8_8:
@@ -16137,6 +16157,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ3_S_R4:
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ2_S_R4:
+        case GGML_TYPE_IQ1_S_R4:
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_Q8_K64:
         case GGML_TYPE_Q8_K16:
@@ -22893,6 +22914,7 @@ void ggml_quantize_init(enum ggml_type type) {
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ1_S:
         case GGML_TYPE_IQ1_M:   iq2xs_init_impl(type); break;
+        case GGML_TYPE_IQ1_S_R4:iq2xs_init_impl(GGML_TYPE_IQ1_S); break;
         case GGML_TYPE_IQ3_XXS_R4:
         case GGML_TYPE_IQ3_XXS: iq3xs_init_impl(256); break;
         case GGML_TYPE_IQ3_S_R4:
@@ -22975,6 +22997,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_IQ3_S_R4:result = quantize_iq3_s_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_S:   result = quantize_iq2_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_S_R4:result = quantize_iq2_s_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ1_S_R4:result = quantize_iq1_s_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ1_S:   result = quantize_iq1_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ1_M:   result = quantize_iq1_m  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ1_BN:  result = quantize_iq1_bn (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;