From bd2d4e393b2b7d2a1b2e201058e26017c9728ead Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Sun, 18 Feb 2024 18:16:55 +0200 Subject: 1.5 bit quantization (#5453) * iq1_s: WIP basics * iq1_s: CUDA is working * iq1_s: scalar CPU dot product * iq1_s: WIP AVX2 dot product - something is not right * Fix tests * Fix shadow warnings * Fix after merge with latest master * iq1_s: AVX2 finally works * iq1_s: ARM_NEON dot product. Works, but not very fast * iq1_s: better grid * iq1_s: use IQ2_XXS for attn_output At a cost of 0.04 extra bpw this gives a big improvement in PPL. * iq1_s: Metal basics Dequantize works, but not dot product * iq1_s: Metal works, but quite slow As usual, Apple Silicon does not like the code I write. * iq1_s: Tests * iq1_s: slightly faster dot product --------- Co-authored-by: Iwan Kawrakow --- llama.cpp | 16 ++++++++++++---- 1 file changed, 12 insertions(+), 4 deletions(-) (limited to 'llama.cpp') diff --git a/llama.cpp b/llama.cpp index 6ac9caa9..5cfebb3b 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2526,6 +2526,7 @@ struct llama_model_loader { case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break; case GGML_TYPE_IQ2_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS; break; case GGML_TYPE_IQ3_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS; break; + case GGML_TYPE_IQ1_S: ftype = LLAMA_FTYPE_MOSTLY_IQ1_S; break; default: { LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max)); @@ -2875,6 +2876,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_IQ2_XS: return "IQ2_XS - 2.3125 bpw"; case LLAMA_FTYPE_MOSTLY_Q3_K_XS:return "Q3_K - Extra small"; case LLAMA_FTYPE_MOSTLY_IQ3_XXS:return "IQ3_XXS - 3.0625 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ1_S :return "IQ1_S - 1.5625 bpw"; default: return "unknown, may not work"; } @@ -10312,20 +10314,20 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) { new_type = GGML_TYPE_Q8_0; } - else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) { + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) { new_type = GGML_TYPE_Q5_K; } else if (new_type != GGML_TYPE_Q8_0) { new_type = GGML_TYPE_Q6_K; } } else if (name == "token_embd.weight") { - if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) { + if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) { new_type = GGML_TYPE_Q2_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) { new_type = GGML_TYPE_Q4_K; } - } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) { + } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) { if (name.find("attn_v.weight") != std::string::npos) { if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q4_K; else new_type = GGML_TYPE_Q2_K; @@ -10335,6 +10337,9 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty if (qs.i_ffn_down < qs.n_ffn_down/8) new_type = GGML_TYPE_Q2_K; ++qs.i_ffn_down; } + else if (name.find("attn_output.weight") != std::string::npos) { + if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) new_type = GGML_TYPE_IQ2_XXS; + } } else if (name.find("attn_v.weight") != std::string::npos) { if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) { new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K; @@ -10468,7 +10473,7 @@ 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) { + new_type == GGML_TYPE_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) { int nx = tensor->ne[0]; int ny = tensor->ne[1]; if (nx % QK_K != 0) { @@ -10483,6 +10488,7 @@ 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_IQ1_S: case GGML_TYPE_Q2_K: new_type = GGML_TYPE_Q4_0; break; case GGML_TYPE_Q3_K: new_type = GGML_TYPE_Q4_1; break; case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break; @@ -10525,6 +10531,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_IQ2_XXS: quantized_type = GGML_TYPE_IQ2_XXS; break; case LLAMA_FTYPE_MOSTLY_IQ2_XS: quantized_type = GGML_TYPE_IQ2_XS; break; 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; default: throw std::runtime_error(format("invalid output file type %d\n", ftype)); } @@ -10698,6 +10705,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s } if ((new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_XS || + new_type == GGML_TYPE_IQ1_S || (new_type == GGML_TYPE_Q2_K && params->ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && strcmp(tensor->name, "token_embd.weight") != 0)) && !imatrix) { LLAMA_LOG_ERROR("\n\n============================================================\n"); LLAMA_LOG_ERROR("Missing importance matrix for tensor %s in a very low-bit quantization\n", tensor->name); -- cgit v1.2.3