ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-06 12:00:29 +00:00
Code Issues Packages Projects Releases Wiki Activity

Adding iq2_bn_r4

Browse Source 
This Zen4-only implementation achieves PP-512 = 826 t/s (!!!)
for Bitnet-1.58b-3B, up from 620 t/s for iq2_bn.
This commit is contained in:
Iwan Kawrakow
2024-12-05 15:18:33 +02:00
parent f64de08203
commit 0137264c6f
10 changed files with 391 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
src/llama.cpp
View File

@@ -3848,6 +3848,7 @@ struct llama_model_loader {
case GGML_TYPE_IQ1_M: ftype = LLAMA_FTYPE_MOSTLY_IQ1_M; break;
case GGML_TYPE_IQ1_BN: ftype = LLAMA_FTYPE_MOSTLY_IQ1_BN; break;
case GGML_TYPE_IQ2_BN: ftype = LLAMA_FTYPE_MOSTLY_IQ2_BN; break;
case GGML_TYPE_IQ2_BN_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ2_BN_R4;break;
case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break;
case GGML_TYPE_IQ4_NL_X4:ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL_X4;break;
case GGML_TYPE_IQ4_XS_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS_R4;break;
@@ -4576,6 +4577,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ6_K: return "IQ6_K - 6.6 bpw";
case LLAMA_FTYPE_MOSTLY_IQ1_BN: return "IQ1_BN - 1.625 bpw Bitnet";
case LLAMA_FTYPE_MOSTLY_IQ2_BN: return "IQ2_BN - 2.00 bpw Bitnet";
case LLAMA_FTYPE_MOSTLY_IQ2_BN_R4:return "IQ2_BN_R4 - 2.00 bpw Bitnet";
case LLAMA_FTYPE_MOSTLY_IQ3_S: return "IQ3_S - 3.4375 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_M: return "IQ3_S mix - 3.66 bpw";
case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: return "Q4_0_4_4";
@@ -15771,7 +15773,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ3_S;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_BN || ftype == LLAMA_FTYPE_MOSTLY_IQ2_BN) {
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_BN || ftype == LLAMA_FTYPE_MOSTLY_IQ2_BN || ftype == LLAMA_FTYPE_MOSTLY_IQ2_BN_R4) {
new_type = GGML_TYPE_IQ4_NL;
}
else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 ||
@@ -16061,7 +16063,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
++qs.n_k_quantized;
}
}
if (new_type == GGML_TYPE_IQ1_BN || new_type == GGML_TYPE_IQ2_BN) {
if (new_type == GGML_TYPE_IQ1_BN || new_type == GGML_TYPE_IQ2_BN || new_type == GGML_TYPE_IQ2_BN_R4) {
int nx = tensor->ne[0];
if (nx % QK_IQ1BN != 0) {
convert_incompatible_tensor = true;
@@ -16190,6 +16192,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_IQ1_M: default_type = GGML_TYPE_IQ1_M; break;
case LLAMA_FTYPE_MOSTLY_IQ1_BN: default_type = GGML_TYPE_IQ1_BN; break;
case LLAMA_FTYPE_MOSTLY_IQ2_BN: default_type = GGML_TYPE_IQ2_BN; break;
case LLAMA_FTYPE_MOSTLY_IQ2_BN_R4:default_type = GGML_TYPE_IQ2_BN_R4;break;
case LLAMA_FTYPE_MOSTLY_IQ4_NL: default_type = GGML_TYPE_IQ4_NL; break;
case LLAMA_FTYPE_MOSTLY_IQ4_NL_X4:default_type = GGML_TYPE_IQ4_NL_X4;break;
case LLAMA_FTYPE_MOSTLY_IQ4_XS_R4:default_type = GGML_TYPE_IQ4_XS_R4;break;
@@ -16574,6 +16577,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q8_0;
else chunk_size_multiplier = 4;
}
else if (new_type == GGML_TYPE_IQ2_BN_R4) {
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ2_BN;
else chunk_size_multiplier = 4;
}
LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
fflush(stdout);