diff --git a/llama.cpp b/llama.cpp index 249442166..110e509cc 100644 --- a/llama.cpp +++ b/llama.cpp @@ -11636,7 +11636,7 @@ static void llama_tensor_dequantize_internal( workers.clear(); } -static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype) { +static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype) { const std::string name = ggml_get_name(tensor); // TODO: avoid hardcoded tensor names - use the TN_* constants @@ -11951,40 +11951,40 @@ static int32_t llama_tensor_quantize_internal(enum ggml_type new_type, const flo } static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) { - ggml_type quantized_type; + ggml_type default_type; llama_ftype ftype = params->ftype; switch (params->ftype) { - case LLAMA_FTYPE_MOSTLY_Q4_0: quantized_type = GGML_TYPE_Q4_0; break; - case LLAMA_FTYPE_MOSTLY_Q4_1: quantized_type = GGML_TYPE_Q4_1; break; - case LLAMA_FTYPE_MOSTLY_Q5_0: quantized_type = GGML_TYPE_Q5_0; break; - case LLAMA_FTYPE_MOSTLY_Q5_1: quantized_type = GGML_TYPE_Q5_1; break; - case LLAMA_FTYPE_MOSTLY_Q8_0: quantized_type = GGML_TYPE_Q8_0; break; - case LLAMA_FTYPE_MOSTLY_F16: quantized_type = GGML_TYPE_F16; break; - case LLAMA_FTYPE_ALL_F32: quantized_type = GGML_TYPE_F32; break; + case LLAMA_FTYPE_MOSTLY_Q4_0: default_type = GGML_TYPE_Q4_0; break; + case LLAMA_FTYPE_MOSTLY_Q4_1: default_type = GGML_TYPE_Q4_1; break; + case LLAMA_FTYPE_MOSTLY_Q5_0: default_type = GGML_TYPE_Q5_0; break; + case LLAMA_FTYPE_MOSTLY_Q5_1: default_type = GGML_TYPE_Q5_1; break; + case LLAMA_FTYPE_MOSTLY_Q8_0: default_type = GGML_TYPE_Q8_0; break; + case LLAMA_FTYPE_MOSTLY_F16: default_type = GGML_TYPE_F16; break; + case LLAMA_FTYPE_ALL_F32: default_type = GGML_TYPE_F32; break; // K-quants case LLAMA_FTYPE_MOSTLY_Q2_K_S: - case LLAMA_FTYPE_MOSTLY_Q2_K: quantized_type = GGML_TYPE_Q2_K; break; - case LLAMA_FTYPE_MOSTLY_IQ3_XS: quantized_type = GGML_TYPE_IQ3_S; break; + case LLAMA_FTYPE_MOSTLY_Q2_K: default_type = GGML_TYPE_Q2_K; break; + case LLAMA_FTYPE_MOSTLY_IQ3_XS: default_type = GGML_TYPE_IQ3_S; break; case LLAMA_FTYPE_MOSTLY_Q3_K_S: case LLAMA_FTYPE_MOSTLY_Q3_K_M: - case LLAMA_FTYPE_MOSTLY_Q3_K_L: quantized_type = GGML_TYPE_Q3_K; break; + case LLAMA_FTYPE_MOSTLY_Q3_K_L: default_type = GGML_TYPE_Q3_K; break; case LLAMA_FTYPE_MOSTLY_Q4_K_S: - case LLAMA_FTYPE_MOSTLY_Q4_K_M: quantized_type = GGML_TYPE_Q4_K; break; + case LLAMA_FTYPE_MOSTLY_Q4_K_M: default_type = GGML_TYPE_Q4_K; break; case LLAMA_FTYPE_MOSTLY_Q5_K_S: - case LLAMA_FTYPE_MOSTLY_Q5_K_M: quantized_type = GGML_TYPE_Q5_K; break; - case LLAMA_FTYPE_MOSTLY_Q6_K: quantized_type = GGML_TYPE_Q6_K; break; - 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_IQ2_S: quantized_type = GGML_TYPE_IQ2_XS; break; - case LLAMA_FTYPE_MOSTLY_IQ2_M: quantized_type = GGML_TYPE_IQ2_S; 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; - case LLAMA_FTYPE_MOSTLY_IQ4_NL: quantized_type = GGML_TYPE_IQ4_NL; break; - case LLAMA_FTYPE_MOSTLY_IQ4_XS: quantized_type = GGML_TYPE_IQ4_XS; break; - case LLAMA_FTYPE_MOSTLY_IQ3_S: quantized_type = GGML_TYPE_IQ3_S; break; - case LLAMA_FTYPE_MOSTLY_IQ3_M: quantized_type = GGML_TYPE_IQ3_S; break; + case LLAMA_FTYPE_MOSTLY_Q5_K_M: default_type = GGML_TYPE_Q5_K; break; + case LLAMA_FTYPE_MOSTLY_Q6_K: default_type = GGML_TYPE_Q6_K; break; + case LLAMA_FTYPE_MOSTLY_IQ2_XXS: default_type = GGML_TYPE_IQ2_XXS; break; + case LLAMA_FTYPE_MOSTLY_IQ2_XS: default_type = GGML_TYPE_IQ2_XS; break; + case LLAMA_FTYPE_MOSTLY_IQ2_S: default_type = GGML_TYPE_IQ2_XS; break; + case LLAMA_FTYPE_MOSTLY_IQ2_M: default_type = GGML_TYPE_IQ2_S; break; + case LLAMA_FTYPE_MOSTLY_IQ3_XXS: default_type = GGML_TYPE_IQ3_XXS; break; + case LLAMA_FTYPE_MOSTLY_IQ1_S: default_type = GGML_TYPE_IQ1_S; break; + case LLAMA_FTYPE_MOSTLY_IQ4_NL: default_type = GGML_TYPE_IQ4_NL; break; + case LLAMA_FTYPE_MOSTLY_IQ4_XS: default_type = GGML_TYPE_IQ4_XS; break; + case LLAMA_FTYPE_MOSTLY_IQ3_S: default_type = GGML_TYPE_IQ3_S; break; + case LLAMA_FTYPE_MOSTLY_IQ3_M: default_type = GGML_TYPE_IQ3_S; break; default: throw std::runtime_error(format("invalid output file type %d\n", ftype)); } @@ -12125,23 +12125,26 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // do not quantize Mamba's small yet 2D weights // NOTE: can't use LLM_TN here because the layer number is not known quantize &= name.find("ssm_conv1d.weight") == std::string::npos; - quantize &= name.find("ssm_x.weight") == std::string::npos; - quantize &= name.find("ssm_dt.weight") == std::string::npos; + quantize &= name.find("ssm_x.weight") == std::string::npos; + quantize &= name.find("ssm_dt.weight") == std::string::npos; enum ggml_type new_type; void * new_data; size_t new_size; if (quantize) { - new_type = quantized_type; - if (!params->pure) { - new_type = get_k_quant_type(qs, new_type, tensor, ftype); + new_type = default_type; + + // get more optimal quantization type based on the tensor shape, layer, etc. + if (!params->pure && ggml_is_quantized(default_type)) { + new_type = llama_tensor_get_type(qs, new_type, tensor, ftype); } // If we've decided to quantize to the same type the tensor is already // in then there's nothing to do. quantize = tensor->type != new_type; } + if (!quantize) { new_type = tensor->type; new_data = tensor->data; @@ -12187,7 +12190,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s f32_data = (float *) f32_conv_buf.data(); } - LLAMA_LOG_INFO("quantizing to %s .. ", ggml_type_name(new_type)); + LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type)); fflush(stdout); if (work.size() < nelements * 4) { @@ -12235,7 +12238,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s LLAMA_LOG_INFO("%s: quant size = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0); if (qs.n_fallback > 0) { - LLAMA_LOG_WARN("%s: WARNING: %d of %d tensor(s) incompatible with k-quants and required fallback quantization\n", + LLAMA_LOG_WARN("%s: WARNING: %d of %d tensor(s) required fallback quantization\n", __func__, qs.n_fallback, qs.n_k_quantized + qs.n_fallback); } } diff --git a/llama.h b/llama.h index c8e05aadd..ccf65ca4e 100644 --- a/llama.h +++ b/llama.h @@ -278,7 +278,7 @@ extern "C" { bool allow_requantize; // allow quantizing non-f32/f16 tensors bool quantize_output_tensor; // quantize output.weight bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored - bool pure; // disable k-quant mixtures and quantize all tensors to the same type + bool pure; // quantize all tensors to the default type void * imatrix; // pointer to importance matrix data } llama_model_quantize_params;