ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-28 17:14:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

IQ3_K_R4 (#145)

Browse Source 
* iq3_k_r4 WIP

* iq3_k_r4: Zen4

* iq3_k_r4: AVX2

* iq3_k_r4: NEON

* iq3_k_r4: faster matrix x vector multiplication on NEON

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-12-17 07:51:11 +01:00
committed by GitHub
parent 1714e46f13
commit d69344f8ea
10 changed files with 482 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
src/llama.cpp
View File

@@ -3867,6 +3867,7 @@ struct llama_model_loader {
case GGML_TYPE_IQ4_KSS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_KSS; break;
case GGML_TYPE_IQ2_K: ftype = LLAMA_FTYPE_MOSTLY_IQ2_K; break;
case GGML_TYPE_IQ3_K: ftype = LLAMA_FTYPE_MOSTLY_IQ3_K; break;
case GGML_TYPE_IQ3_K_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ3_K_R4;break;
case GGML_TYPE_IQ4_K: ftype = LLAMA_FTYPE_MOSTLY_IQ4_K; break;
case GGML_TYPE_IQ4_K_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ4_K_R4;break;
case GGML_TYPE_IQ5_K: ftype = LLAMA_FTYPE_MOSTLY_IQ5_K; break;
@@ -4585,6 +4586,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ4_KSS: return "IQ4_KSS - 4.0 bpw";
case LLAMA_FTYPE_MOSTLY_IQ2_K: return "IQ2_K - 2.375 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_K: return "IQ3_K - 3.4325 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_K_R4: return "IQ3_K_R4 - 3.4325 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_KL: return "IQ3_KL - 4 bpw";
case LLAMA_FTYPE_MOSTLY_IQ4_K: return "IQ4_K - 4.5 bpw";
case LLAMA_FTYPE_MOSTLY_IQ4_K_R4: return "IQ4_K_R4 - 4.5 bpw";
@@ -15763,7 +15765,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ||
ftype == LLAMA_FTYPE_MOSTLY_IQ1_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_KS) {
ftype == LLAMA_FTYPE_MOSTLY_IQ2_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4) {
new_type = !qs.has_output ? GGML_TYPE_IQ4_K : GGML_TYPE_Q5_K;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_S || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS ||
@@ -15820,6 +15822,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (new_type == GGML_TYPE_Q8_K_R8) {
new_type = GGML_TYPE_Q8_0;
}
else if (new_type == GGML_TYPE_IQ3_K_R4) {
new_type = GGML_TYPE_IQ3_K;
}
else if (new_type == GGML_TYPE_IQ4_K_R4) {
new_type = GGML_TYPE_IQ4_K;
}
@@ -15892,6 +15897,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_K && qs.model.hparams.n_gqa() >= 2) {
new_type = GGML_TYPE_IQ4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4 && qs.model.hparams.n_gqa() >= 2) {
new_type = GGML_TYPE_IQ4_K_R4;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_KL) {
new_type = qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ5_K : GGML_TYPE_IQ4_K;
}
@@ -16039,7 +16047,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_K ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_R4 ||
ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R4 || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_R4 ||
ftype == LLAMA_FTYPE_MOSTLY_Q2_K_R4|| ftype == LLAMA_FTYPE_MOSTLY_IQ4_K_R4) {
ftype == LLAMA_FTYPE_MOSTLY_Q2_K_R4|| ftype == LLAMA_FTYPE_MOSTLY_IQ4_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4) {
new_type = GGML_TYPE_Q5_K;
}
} else {
@@ -16110,7 +16118,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
new_type == GGML_TYPE_IQ6_K || new_type == GGML_TYPE_IQ4_KS || new_type == GGML_TYPE_IQ4_XS_R4 ||
new_type == GGML_TYPE_IQ2_KS || new_type == GGML_TYPE_IQ4_KSS || new_type == GGML_TYPE_Q6_K_R4 ||
new_type == GGML_TYPE_Q5_K_R4 || new_type == GGML_TYPE_Q3_K_R4 || new_type == GGML_TYPE_Q2_K_R4 ||
new_type == GGML_TYPE_IQ4_K_R4|| new_type == GGML_TYPE_Q8_K_R8) {
new_type == GGML_TYPE_IQ4_K_R4|| new_type == GGML_TYPE_Q8_K_R8 || new_type == GGML_TYPE_IQ3_K_R4) {
int nx = tensor->ne[0];
int ny = tensor->ne[1];
if (nx % QK_K != 0) {
@@ -16142,6 +16150,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
case GGML_TYPE_Q3_K_R4:
case GGML_TYPE_IQ2_K:
case GGML_TYPE_IQ3_K:
case GGML_TYPE_IQ3_K_R4:
case GGML_TYPE_IQ4_KSS:
case GGML_TYPE_IQ4_KS:
case GGML_TYPE_IQ4_XS_R4:
@@ -16277,6 +16286,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_IQ4_KSS: default_type = GGML_TYPE_IQ4_KSS; break;
case LLAMA_FTYPE_MOSTLY_IQ2_K: default_type = GGML_TYPE_IQ2_K; break;
case LLAMA_FTYPE_MOSTLY_IQ3_K: default_type = GGML_TYPE_IQ3_K; break;
case LLAMA_FTYPE_MOSTLY_IQ3_K_R4:default_type = GGML_TYPE_IQ3_K_R4;break;
case LLAMA_FTYPE_MOSTLY_IQ3_KL: default_type = GGML_TYPE_IQ3_K; break;
case LLAMA_FTYPE_MOSTLY_IQ4_K: default_type = GGML_TYPE_IQ4_K; break;
case LLAMA_FTYPE_MOSTLY_IQ4_K_R4:default_type = GGML_TYPE_IQ4_K_R4;break;
@@ -16685,6 +16695,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ2_BN;
else chunk_size_multiplier = 4;
}
else if (new_type == GGML_TYPE_IQ3_K_R4) {
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ3_K;
else chunk_size_multiplier = 4;
}
else if (new_type == GGML_TYPE_IQ4_K_R4) {
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ4_K;
else chunk_size_multiplier = 4;