vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

iq4_xs: revert using IQ3_S for attn_k and attn_v

Browse source 
PPL vs size is good, but CPU performance suffers: on M2 Max
TG-128 drops to 21.7 t/s from 28.8, and on a Ryzen-7950X
to 14.5 t/s from 15.8 t/s. On CUDA we have 135 t/s when
using IQ3_S vs 133 t/s with pure IQ4_XS.
This commit is contained in:
Iwan Kawrakow 2024-02-27 10:14:12 +02:00
parent 5c2b230552
commit f162fcaf81
2 changed files with 3 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

2
examples/quantize/quantize.cpp
View file

@ -37,7 +37,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", }, { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },
{ "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", }, { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", },
{ "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.50 bpw non-linear quantization", }, { "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.50 bpw non-linear quantization", },
{ "IQ4_XS", LLAMA_FTYPE_MOSTLY_IQ4_XS, " 4.125 bpw non-linear quantization",}, { "IQ4_XS", LLAMA_FTYPE_MOSTLY_IQ4_XS, " 4.25 bpw non-linear quantization", },
{ "Q4_K", LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", }, { "Q4_K", LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", },
{ "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", }, { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", },
{ "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", }, { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", },

10
llama.cpp
View file

@ -2941,7 +2941,7 @@ 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_IQ3_XXS:return "IQ3_XXS - 3.0625 bpw";
case LLAMA_FTYPE_MOSTLY_IQ1_S :return "IQ1_S - 1.5625 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_IQ4_NL: return "IQ4_NL - 4.5 bpw";
case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.125 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.25 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_S: return "IQ3_S - 3.4375 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"; case LLAMA_FTYPE_MOSTLY_IQ3_M: return "IQ3_S mix - 3.66 bpw";
@ -10866,9 +10866,6 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) { else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ2_S; new_type = GGML_TYPE_IQ2_S;
} }
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) {
new_type = GGML_TYPE_IQ3_S;
}
} else if (name.find("attn_q.weight") != std::string::npos) { } else if (name.find("attn_q.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) { if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
new_type = GGML_TYPE_IQ3_XXS; new_type = GGML_TYPE_IQ3_XXS;
@ -10876,9 +10873,6 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) { else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ2_S; new_type = GGML_TYPE_IQ2_S;
} }
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) {
new_type = GGML_TYPE_IQ3_S;
}
} else if (name.find("ffn_down") != std::string::npos) { } else if (name.find("ffn_down") != std::string::npos) {
auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str()); auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str());
int i_layer = info.first, n_layer = info.second; int i_layer = info.first, n_layer = info.second;
@ -10981,7 +10975,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
//} //}
bool convert_incompatible_tensor = false; bool convert_incompatible_tensor = false;
if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K || 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_Q5_K || new_type == GGML_TYPE_Q6_K || new_type == GGML_TYPE_IQ4_XS ||
new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_S || new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_S ||
new_type == GGML_TYPE_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || new_type == GGML_TYPE_IQ3_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 nx = tensor->ne[0];