Bitnet changes (#106)

* Adapting iq2_bn to work without separate scale tensors

Why? It is becoming burdensome to maintain the special Bitnet
conversion in convert_hf_to_gguf.py, so I thnk it is better
to make iq1_bn and iq2_bn just work with the mainline
conversion script (which does not generate scales).

* Adapting iq1_bn to work without separate scale tensors

* Adapting iq2_bn: CUDA dequantize

* Adapting iq2_bn: CUDA works

* Adapting iq1_bn: CUDA works

* Adapting iq1_bn, iq2_bn: NEON

* Adapting iq1_bn, iq2_bn: Metal

Dequantize works, but there is still something wrong
with the dot products.

* WIP

Absoolutely don't see what is wrong with the iq1_bn and iq2_bn
vector dot product kernels.

* Remove iq1_tn and iq2_tn - Part 1

Now that iq1_bn and iq2_bn have per row scales, there is no
reason to also have iq1_tn and iq2_tn.

* Remove iq1_tn and iq2_tn - Part 2

* Bitnet: use the standard llm_build_kv to build self attention

My main motivation was to enable FA. But FA does not work anyway
because head size is 100 for the Botnet ternary models
(and I had forgotten this little detail).

* Revert "Avoid rebuild of GGML graph for each token (#98)"

This reverts commit f2d315b46f7aacc7df4b86bd8acba387b30e11ca.
As far as I can tell, the commit breaks Metal TG.

---------

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

1 files changed, 1 insertions, 45 deletions

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 35ed68d0..5570b1fc 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1016,7 +1016,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .vec_dot                  = ggml_vec_dot_iq1_bn_q8_K64,
         .vec_dot_type             = GGML_TYPE_Q8_K64,
         .nrows                    = 1,
-        .row_meta_size            = 0,
+        .row_meta_size            = 2,
     },
     [GGML_TYPE_IQ2_BN] = {
         .type_name                = "iq2_bn",
@@ -1029,34 +1029,8 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .vec_dot                  = ggml_vec_dot_iq2_bn_q8_K64,
         .vec_dot_type             = GGML_TYPE_Q8_K64,
         .nrows                    = 1,
-        .row_meta_size            = 0,
-    },
-    [GGML_TYPE_IQ2_TN] = {
-        .type_name                = "iq2_tn",
-        .blck_size                = QK_K,
-        .type_size                = sizeof(block_iq2_tn),
-        .is_quantized             = true,
-        .to_float                 = (ggml_to_float_t) dequantize_row_iq2_tn,
-        .from_float               = quantize_row_iq2_tn,
-        .from_float_ref           = (ggml_from_float_t)quantize_row_iq2_tn_ref,
-        .vec_dot                  = vec_dot_iq2_tn_q8_k,
-        .vec_dot_type             = GGML_TYPE_Q8_K,
-        .nrows                    = 1,
         .row_meta_size            = 4,
     },
-    [GGML_TYPE_IQ1_TN] = {
-        .type_name                = "iq1_tn",
-        .blck_size                = QK_K,
-        .type_size                = sizeof(block_iq1_tn),
-        .is_quantized             = true,
-        .to_float                 = (ggml_to_float_t) dequantize_row_iq1_tn,
-        .from_float               = quantize_row_iq1_tn,
-        .from_float_ref           = (ggml_from_float_t)quantize_row_iq1_tn_ref,
-        .vec_dot                  = vec_dot_iq1_tn_q8_k,
-        .vec_dot_type             = GGML_TYPE_Q8_K64,
-        .nrows                    = 1,
-        .row_meta_size            = 2,
-    },
     [GGML_TYPE_IQ4_NL] = {
         .type_name                = "iq4_nl",
         .blck_size                = QK4_NL,
@@ -3926,8 +3900,6 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_IQ1_M:         wtype = GGML_TYPE_IQ1_M;    break;
         case GGML_FTYPE_MOSTLY_IQ1_BN:        wtype = GGML_TYPE_IQ1_BN;   break;
         case GGML_FTYPE_MOSTLY_IQ2_BN:        wtype = GGML_TYPE_IQ2_BN;   break;
-        case GGML_FTYPE_MOSTLY_IQ2_TN:        wtype = GGML_TYPE_IQ2_TN;   break;
-        case GGML_FTYPE_MOSTLY_IQ1_TN:        wtype = GGML_TYPE_IQ1_TN;   break;
         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;
@@ -10428,8 +10400,6 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
-        case GGML_TYPE_IQ2_TN:
-        case GGML_TYPE_IQ1_TN:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
@@ -10819,8 +10789,6 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
-        case GGML_TYPE_IQ2_TN:
-        case GGML_TYPE_IQ1_TN:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
@@ -10960,8 +10928,6 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
-        case GGML_TYPE_IQ2_TN:
-        case GGML_TYPE_IQ1_TN:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
@@ -14147,8 +14113,6 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
-        case GGML_TYPE_IQ2_TN:
-        case GGML_TYPE_IQ1_TN:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
@@ -14528,8 +14492,6 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
-        case GGML_TYPE_IQ2_TN:
-        case GGML_TYPE_IQ1_TN:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
@@ -14803,8 +14765,6 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
-        case GGML_TYPE_IQ2_TN:
-        case GGML_TYPE_IQ1_TN:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
@@ -15405,8 +15365,6 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ1_BN:
         case GGML_TYPE_IQ2_BN:
-        case GGML_TYPE_IQ2_TN:
-        case GGML_TYPE_IQ1_TN:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ4_KS:
@@ -22224,8 +22182,6 @@ size_t ggml_quantize_chunk(
         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;
         case GGML_TYPE_IQ2_BN:  result = quantize_iq2_bn (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_IQ2_TN:  result = quantize_iq2_tn (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_IQ1_TN:  result = quantize_iq1_tn (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         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;