From 998b635a17adc33dfccf8083417efc721351f0f3 Mon Sep 17 00:00:00 2001 From: Iwan Kawrakow Date: Sun, 14 Jan 2024 08:39:10 +0200 Subject: [PATCH] Fix ffn_down quantization mix for MoE models In #4872 I did not consider the part where every third tensor is quantized with more bits. Fir MoE this leads to tensors of the same layer being quantized with different number of bits, which is not considered as a possibility in the inference implementation (it is assumed all experts use the same quantization). --- llama.cpp | 19 +++++++++++-------- 1 file changed, 11 insertions(+), 8 deletions(-) diff --git a/llama.cpp b/llama.cpp index 66494974a..28de3c741 100644 --- a/llama.cpp +++ b/llama.cpp @@ -8462,13 +8462,16 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty new_type = GGML_TYPE_Q8_0; } } else if (name.find("ffn_down") != std::string::npos) { + const int n_expert = std::max(1, (int)qs.model.hparams.n_expert); + const int i_layer = qs.i_feed_forward_w2 / n_expert; + const int n_layer = qs.i_feed_forward_w2 / n_expert; if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) { - if (qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) new_type = GGML_TYPE_Q4_K; + if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { - new_type = qs.i_feed_forward_w2 < qs.n_feed_forward_w2/16 ? GGML_TYPE_Q5_K - : arch != LLM_ARCH_FALCON || use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q4_K + new_type = i_layer < n_layer/16 ? GGML_TYPE_Q5_K + : 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_Q3_K_L) { @@ -8476,14 +8479,14 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty } else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) { if (arch == LLM_ARCH_FALCON) { - new_type = qs.i_feed_forward_w2 < qs.n_feed_forward_w2/16 ? GGML_TYPE_Q6_K : - use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; + new_type = i_layer < n_layer/16 ? GGML_TYPE_Q6_K : + use_more_bits(i_layer, n_layer) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; } else { - if (use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; + if (use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K; } } - else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) { + else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(i_layer, n_layer)) new_type = GGML_TYPE_Q6_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) { new_type = GGML_TYPE_Q5_K; } ++qs.i_feed_forward_w2;