summaryrefslogtreecommitdiff
path: root/llama.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'llama.cpp')
-rw-r--r--llama.cpp50
1 files changed, 37 insertions, 13 deletions
diff --git a/llama.cpp b/llama.cpp
index 37477e6e..1f6b6cff 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -2545,6 +2545,7 @@ struct llama_model_loader {
case GGML_TYPE_IQ3_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS; break;
case GGML_TYPE_IQ1_S: ftype = LLAMA_FTYPE_MOSTLY_IQ1_S; break;
case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break;
+ case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break;
default:
{
LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@@ -2890,6 +2891,8 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ3_XXS:return "IQ3_XXS - 3.0625 bpw";
case LLAMA_FTYPE_MOSTLY_IQ1_S :return "IQ1_S - 1.5625 bpw";
case LLAMA_FTYPE_MOSTLY_IQ4_NL: return "IQ4_NL - 4.5 bpw";
+ 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";
default: return "unknown, may not work";
}
@@ -10544,6 +10547,12 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : !qs.has_imatrix ? GGML_TYPE_Q3_K : GGML_TYPE_IQ3_XXS;
}
+ else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_S && qs.model.hparams.n_gqa() >= 4) {
+ new_type = GGML_TYPE_Q4_K;
+ }
+ else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
+ new_type = GGML_TYPE_Q4_K;
+ }
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
}
@@ -10575,13 +10584,17 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
new_type = GGML_TYPE_Q8_0;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) {
- new_type = GGML_TYPE_Q2_K;
+ new_type = GGML_TYPE_IQ3_XXS;
+ }
+ } else if (name.find("attn_q.weight") != std::string::npos) {
+ if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) {
+ new_type = GGML_TYPE_IQ3_XXS;
}
} else if (name.find("ffn_down") != std::string::npos) {
auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str());
int i_layer = info.first, n_layer = info.second;
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
- else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) {
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) {
if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && !qs.has_imatrix) {
@@ -10592,6 +10605,10 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
: arch != LLM_ARCH_FALCON || use_more_bits(i_layer, n_layer) ? GGML_TYPE_Q4_K
: GGML_TYPE_Q3_K;
}
+ else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M && (i_layer < n_layer/8 ||
+ (qs.model.hparams.n_expert == 8 && use_more_bits(i_layer, n_layer)))) {
+ new_type = GGML_TYPE_Q4_K;
+ }
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
new_type = arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K;
}
@@ -10623,37 +10640,41 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
if (qs.model.hparams.n_expert == 8) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL ||
- ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
+ ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S ||
+ ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
new_type = GGML_TYPE_Q5_K;
}
} else {
- if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K;
+ if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) new_type = GGML_TYPE_Q3_K;
- else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
- else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ) new_type = GGML_TYPE_Q4_K;
+ else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L ) new_type = GGML_TYPE_Q5_K;
+ else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M ) new_type = GGML_TYPE_Q4_K;
}
} else {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
}
}
else if (name.find("attn_qkv.weight") != std::string::npos) {
- if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
+ if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) {
+ new_type = GGML_TYPE_Q4_K;
+ }
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
}
else if (name.find("ffn_gate") != std::string::npos) {
auto info = layer_info(qs.i_ffn_gate, qs.n_ffn_gate, name.c_str());
int i_layer = info.first, n_layer = info.second;
- if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(i_layer, n_layer)) {
- new_type = GGML_TYPE_Q2_K;
+ if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) {
+ new_type = GGML_TYPE_IQ3_XXS;
}
++qs.i_ffn_gate;
}
else if (name.find("ffn_up") != std::string::npos) {
auto info = layer_info(qs.i_ffn_up, qs.n_ffn_up, name.c_str());
int i_layer = info.first, n_layer = info.second;
- if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(i_layer, n_layer)) {
- new_type = GGML_TYPE_Q2_K;
+ if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) {
+ new_type = GGML_TYPE_IQ3_XXS;
}
++qs.i_ffn_up;
}
@@ -10673,7 +10694,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K ||
new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K ||
new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_XXS ||
- new_type == GGML_TYPE_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) {
+ new_type == GGML_TYPE_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || new_type == GGML_TYPE_IQ3_S) {
int nx = tensor->ne[0];
int ny = tensor->ne[1];
if (nx % QK_K != 0) {
@@ -10688,6 +10709,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ3_XXS:
+ case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K: new_type = GGML_TYPE_IQ4_NL; break;
@@ -10719,7 +10741,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// K-quants
case LLAMA_FTYPE_MOSTLY_Q2_K_S:
case LLAMA_FTYPE_MOSTLY_Q2_K: quantized_type = GGML_TYPE_Q2_K; break;
- case LLAMA_FTYPE_MOSTLY_Q3_K_XS:
+ case LLAMA_FTYPE_MOSTLY_Q3_K_XS: quantized_type = GGML_TYPE_IQ3_S; break;
case LLAMA_FTYPE_MOSTLY_Q3_K_S:
case LLAMA_FTYPE_MOSTLY_Q3_K_M:
case LLAMA_FTYPE_MOSTLY_Q3_K_L: quantized_type = GGML_TYPE_Q3_K; break;
@@ -10733,6 +10755,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_IQ3_XXS: quantized_type = GGML_TYPE_IQ3_XXS; break;
case LLAMA_FTYPE_MOSTLY_IQ1_S: quantized_type = GGML_TYPE_IQ1_S; break;
case LLAMA_FTYPE_MOSTLY_IQ4_NL: quantized_type = GGML_TYPE_IQ4_NL; break;
+ case LLAMA_FTYPE_MOSTLY_IQ3_S: quantized_type = GGML_TYPE_IQ3_S; break;
+ case LLAMA_FTYPE_MOSTLY_IQ3_M: quantized_type = GGML_TYPE_IQ3_S; break;
default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
}
generated by cgit v1.2.3 (git 2.25.1) at 2025-08-18 00:23:01 +0000