From bbef28218fe827265716b66977719b9ee2b21165 Mon Sep 17 00:00:00 2001 From: LostRuins <39025047+LostRuins@users.noreply.github.com> Date: Tue, 11 Jul 2023 22:01:08 +0800 Subject: [PATCH] Possible solution to allow K-quants on models with n_vocab!=32000 (#2148) * This allows LLAMA models that were previously incompatible with K quants to function mostly as normal. This happens when a model has a vocab != 32000, e.g 32001 which means it's not divisible by 256 or 64. Since the problematic dimensions only apply for `tok_embeddings.weight` and `output.weight` (dimentions 4096 x n_vocab), we can simply quantize these layers to Q8_0 whereas the majority of the hidden layers are still K-quanted since they have compatible dimensions. * Fix indentation Co-authored-by: Georgi Gerganov * As an alternative, to avoid failing on Metal due to lack of Q8_0 support, instead quantize tok_embeddings.weight to Q4_0 and retain output.weight as F16. This results in a net gain of about 55mb for a 7B model compared to previous approach, but should minimize adverse impact to model quality. --------- Co-authored-by: Georgi Gerganov --- llama.cpp | 18 ++++++++++++++---- 1 file changed, 14 insertions(+), 4 deletions(-) diff --git a/llama.cpp b/llama.cpp index ad7283faf..08ec21ab6 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2454,15 +2454,14 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s } else { new_type = quantized_type; #ifdef GGML_USE_K_QUANTS + bool convert_incompatible_tensor = false; if (quantized_type == GGML_TYPE_Q2_K || quantized_type == GGML_TYPE_Q3_K || quantized_type == GGML_TYPE_Q4_K || quantized_type == GGML_TYPE_Q5_K || quantized_type == GGML_TYPE_Q6_K) { int nx = tensor.ne.at(0); int ny = tensor.ne.at(1); if (nx % QK_K != 0 || ny % QK_K != 0) { - fprintf(stderr, "\n\n========================= Tensor sizes %d x %d are not divisible by %d\n",nx,ny,QK_K); - fprintf(stderr, "This is required to be able to use k-quants for now!\n"); - fprintf(stderr, "========================================================================================\n\n"); - throw std::runtime_error("Unsupported tensor size encountered\n"); + fprintf(stderr, "\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K); + convert_incompatible_tensor = true; } } if (tensor.name == "output.weight") { @@ -2490,6 +2489,17 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K; } + if (convert_incompatible_tensor) { + if (tensor.name == "output.weight") { + new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing. + fprintf(stderr, "F16 will be used for this tensor instead.\n"); + } else if (tensor.name == "tok_embeddings.weight") { + new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing. + fprintf(stderr, "Q4_0 will be used for this tensor instead.\n"); + } else { + throw std::runtime_error("Unsupported tensor size encountered\n"); + } + } #endif float * f32_data;