From f162fcaf8118c5313968c24009574711e2ebfc0d Mon Sep 17 00:00:00 2001 From: Iwan Kawrakow Date: Tue, 27 Feb 2024 10:14:12 +0200 Subject: [PATCH] iq4_xs: revert using IQ3_S for attn_k and attn_v 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. --- examples/quantize/quantize.cpp | 2 +- llama.cpp | 10 ++-------- 2 files changed, 3 insertions(+), 9 deletions(-) diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 2c64946aa..7662ec80c 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -37,7 +37,7 @@ static const std::vector QUANT_OPTIONS = { { "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", }, { "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_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", }, diff --git a/llama.cpp b/llama.cpp index 4c1b70afb..2fe4b88ef 100644 --- a/llama.cpp +++ b/llama.cpp @@ -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_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_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_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) { 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) { if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) { 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) { 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) { auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str()); 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; 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_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || new_type == GGML_TYPE_IQ3_S) { int nx = tensor->ne[0];