diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index b989932107dba..219b3c2bd5227 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -54,6 +54,7 @@ static const std::vector QUANT_OPTIONS = { { "F16", LLAMA_FTYPE_MOSTLY_F16, "14.00G, +0.0020 ppl @ Mistral-7B", }, { "BF16", LLAMA_FTYPE_MOSTLY_BF16, "14.00G, -0.0050 ppl @ Mistral-7B", }, { "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", }, + { "CQS", LLAMA_FTYPE_CQS, "Custom Quantization Scheme", }, // Note: Ensure COPY comes after F32 to avoid ftype 0 from matching. { "COPY", LLAMA_FTYPE_ALL_F32, "only copy tensors, no quantizing", }, }; @@ -107,19 +108,35 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp // [[noreturn]] static void usage(const char * executable) { - printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable); + printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--attn-q-type] [--attn-k-type] [--attn-v-type] [--attn-qkv-type] [--attn-output-type] [--ffn-gate-type] [--ffn-down-type] [--ffn-up-type] [--keep-split] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable); printf(" --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n"); printf(" --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n"); printf(" --pure: Disable k-quant mixtures and quantize all tensors to the same type\n"); printf(" --imatrix file_name: use data in file_name as importance matrix for quant optimizations\n"); printf(" --include-weights tensor_name: use importance matrix for this/these tensor(s)\n"); printf(" --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n"); - printf(" --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor\n"); - printf(" --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n"); + printf(" --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor.\n"); + printf(" --token-embedding-type ggml_type: use this ggml_type for the token_embd.weight tensor.\n\n"); + printf("Additional specific tensor quantization types used in the custom quant scheme 'CQS (default is Q2_K):\n"); + printf(" --attn-q-type ggml_type: use this ggml_type for the attn_q.weight tensor.\n"); + printf(" --attn-k-type ggml_type: use this ggml_type for the attn_k.weight tensor.\n"); + printf(" --attn-v-type ggml_type: use this ggml_type for the attn_v.weight tensor.\n"); + printf(" --attn-qkv-type ggml_type: use this ggml_type for the attn_qkv.weight tensor.\n"); + printf(" --attn-output-type ggml_type: use this ggml_type for the attn_output.weight tensor.\n"); + printf(" --ffn-gate-type ggml_type: use this ggml_type for the ffn_gate tensor.\n"); + printf(" --ffn-down-type ggml_type: use this ggml_type for the ffn_down tensor.\n"); + printf(" --ffn-up-type ggml_type: use this ggml_type for the ffn_up tensor.\n\n"); printf(" --keep-split: will generate quantized model in the same shards as input\n"); printf(" --override-kv KEY=TYPE:VALUE\n"); - printf(" Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n"); + printf(" Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n\n"); printf("Note: --include-weights and --exclude-weights cannot be used together\n"); + printf("Note: The token embeddings tensor is loaded in system RAM, even in case of full GPU/VRAM offload.\n"); + printf("Note: The recommanded type for the output tensor is q6_K for the ffn types > iq3_xxs and < q8_0.\n\n"); + printf("Note for the Custom Quant Scheme FTYPE:\n"); + printf(" Write the specific tensor legacy quants as qN_N, the K-Quants as qN_K, the IQ-Quants as iqN_xx.\n"); + printf(" Usually, attn-q-type can be one type below the chosen ffn type, and attn-v-type should be one type above.\n"); + printf(" attn-qkv-type replaces the types attn-q, attn-k and attn-v on some models.\n"); + //TODO: - eventually - harmonize the CAPS writing of the FTYPEs, and non CAPS writing of the GGML_TYPEs. printf("\nAllowed quantization types:\n"); for (auto & it : QUANT_OPTIONS) { if (it.name != "COPY") { @@ -279,6 +296,54 @@ int main(int argc, char ** argv) { } else { usage(argv[0]); } + } else if (strcmp(argv[arg_idx], "--attn-q-type") == 0) { + if (arg_idx < argc-1) { + params.attn_q_type = parse_ggml_type(argv[++arg_idx]); + } else { + usage(argv[0]); + } + } else if (strcmp(argv[arg_idx], "--attn-k-type") == 0) { + if (arg_idx < argc-1) { + params.attn_k_type = parse_ggml_type(argv[++arg_idx]); + } else { + usage(argv[0]); + } + } else if (strcmp(argv[arg_idx], "--attn-v-type") == 0) { + if (arg_idx < argc-1) { + params.attn_v_type = parse_ggml_type(argv[++arg_idx]); + } else { + usage(argv[0]); + } + } else if (strcmp(argv[arg_idx], "--attn-qkv-type") == 0) { + if (arg_idx < argc-1) { + params.attn_qkv_type = parse_ggml_type(argv[++arg_idx]); + } else { + usage(argv[0]); + } + } else if (strcmp(argv[arg_idx], "--attn-output-type") == 0) { + if (arg_idx < argc-1) { + params.attn_output_type = parse_ggml_type(argv[++arg_idx]); + } else { + usage(argv[0]); + } + } else if (strcmp(argv[arg_idx], "--ffn-gate-type") == 0) { + if (arg_idx < argc-1) { + params.ffn_gate_type = parse_ggml_type(argv[++arg_idx]); + } else { + usage(argv[0]); + } + } else if (strcmp(argv[arg_idx], "--ffn-down-type") == 0) { + if (arg_idx < argc-1) { + params.ffn_down_type = parse_ggml_type(argv[++arg_idx]); + } else { + usage(argv[0]); + } + } else if (strcmp(argv[arg_idx], "--ffn-up-type") == 0) { + if (arg_idx < argc-1) { + params.ffn_up_type = parse_ggml_type(argv[++arg_idx]); + } else { + usage(argv[0]); + } } else if (strcmp(argv[arg_idx], "--override-kv") == 0) { if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) { usage(argv[0]); diff --git a/include/llama.h b/include/llama.h index 7cae1bbe2e5b8..4eaf93940514e 100644 --- a/include/llama.h +++ b/include/llama.h @@ -175,6 +175,7 @@ extern "C" { LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35, // except 1d tensors LLAMA_FTYPE_MOSTLY_TQ1_0 = 36, // except 1d tensors LLAMA_FTYPE_MOSTLY_TQ2_0 = 37, // except 1d tensors + LLAMA_FTYPE_CQS = 99, // except 1d tensors LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file }; @@ -360,6 +361,14 @@ extern "C" { enum llama_ftype ftype; // quantize to this llama_ftype enum ggml_type output_tensor_type; // output tensor type enum ggml_type token_embedding_type; // token embeddings tensor type + enum ggml_type attn_q_type; // attention query tensor type + enum ggml_type attn_k_type; // attention key tensor type + enum ggml_type attn_v_type; // attention value tensor type + enum ggml_type attn_qkv_type; // attention query-key-value tensor type + enum ggml_type attn_output_type; // attention output tensor type + enum ggml_type ffn_gate_type; // feedforward network gate type + enum ggml_type ffn_down_type; // feedforward network down type + enum ggml_type ffn_up_type; // feedforward network up type 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 diff --git a/src/llama.cpp b/src/llama.cpp index 01cdf17dcb91b..172a19b73d0b3 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -5317,6 +5317,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: return "Q4_0_4_4"; case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: return "Q4_0_4_8"; case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: return "Q4_0_8_8"; + case LLAMA_FTYPE_CQS: return "Custom Quantization Scheme"; default: return "unknown, may not work"; } @@ -18028,7 +18029,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n } } } else if (name.find("attn_v.weight") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) { + if (ftype == LLAMA_FTYPE_CQS && qs.params->attn_v_type < GGML_TYPE_COUNT) { + new_type = qs.params->attn_v_type; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) { new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && qs.model.hparams.n_gqa() >= 4) { @@ -18066,7 +18070,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n } ++qs.i_attention_wv; } else if (name.find("attn_k.weight") != std::string::npos) { - if (qs.model.hparams.n_expert == 8) { + if (ftype == LLAMA_FTYPE_CQS && qs.params->attn_k_type < GGML_TYPE_COUNT) { + new_type = qs.params->attn_k_type; + } + else if (qs.model.hparams.n_expert == 8) { // for the 8-expert model, bumping this to Q8_0 trades just ~128MB // TODO: explore better strategies new_type = GGML_TYPE_Q8_0; @@ -18078,7 +18085,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n new_type = GGML_TYPE_IQ2_S; } } else if (name.find("attn_q.weight") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) { + if (ftype == LLAMA_FTYPE_CQS && qs.params->attn_q_type < GGML_TYPE_COUNT) { + new_type = qs.params->attn_q_type; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) { new_type = GGML_TYPE_IQ3_XXS; } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) { @@ -18087,7 +18097,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n } 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; - if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; + if (ftype == LLAMA_FTYPE_CQS && qs.params->ffn_down_type < GGML_TYPE_COUNT) { + new_type = qs.params->ffn_down_type; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) { if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K; } @@ -18130,7 +18143,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n } ++qs.i_ffn_down; } else if (name.find("attn_output.weight") != std::string::npos) { - if (arch != LLM_ARCH_FALCON) { + if (ftype == LLAMA_FTYPE_CQS && qs.params->attn_output_type < GGML_TYPE_COUNT) { + new_type = qs.params->attn_output_type; + } + else if (arch != LLM_ARCH_FALCON) { if (qs.model.hparams.n_expert == 8) { if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || @@ -18150,7 +18166,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n } } else if (name.find("attn_qkv.weight") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) { + if (ftype == LLAMA_FTYPE_CQS && qs.params->attn_qkv_type < GGML_TYPE_COUNT) { + new_type = qs.params->attn_qkv_type; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) { new_type = GGML_TYPE_Q4_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K; @@ -18159,7 +18178,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (name.find("ffn_gate") != std::string::npos) { auto info = layer_info(qs.i_ffn_gate, qs.n_ffn_gate, name.c_str()); int i_layer = info.first, n_layer = info.second; - if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) { + if (ftype == LLAMA_FTYPE_CQS && qs.params->ffn_gate_type < GGML_TYPE_COUNT) { + new_type = qs.params->ffn_gate_type; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) { new_type = GGML_TYPE_IQ3_XXS; } ++qs.i_ffn_gate; @@ -18167,7 +18189,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (name.find("ffn_up") != std::string::npos) { auto info = layer_info(qs.i_ffn_up, qs.n_ffn_up, name.c_str()); int i_layer = info.first, n_layer = info.second; - if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) { + if (ftype == LLAMA_FTYPE_CQS && qs.params->ffn_up_type < GGML_TYPE_COUNT) { + new_type = qs.params->ffn_up_type; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (i_layer >= n_layer/8 && i_layer < 7*n_layer/8)) { new_type = GGML_TYPE_IQ3_XXS; } ++qs.i_ffn_up; @@ -18325,6 +18350,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: default_type = GGML_TYPE_Q4_0_4_8; break; case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: default_type = GGML_TYPE_Q4_0_8_8; break; + // Custom Quantization Scheme + case LLAMA_FTYPE_CQS: default_type = GGML_TYPE_Q2_K; break; + default: throw std::runtime_error(format("invalid output file type %d\n", ftype)); } @@ -18583,6 +18611,30 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (params->output_tensor_type < GGML_TYPE_COUNT && strcmp(tensor->name, "output.weight") == 0) { new_type = params->output_tensor_type; } + if (params->attn_q_type < GGML_TYPE_COUNT && strcmp(tensor->name, "attn_q.weight") == 0) { + new_type = params->attn_q_type; + } + if (params->attn_k_type < GGML_TYPE_COUNT && strcmp(tensor->name, "attn_k.weight") == 0) { + new_type = params->attn_k_type; + } + if (params->attn_v_type < GGML_TYPE_COUNT && strcmp(tensor->name, "attn_v.weight") == 0) { + new_type = params->attn_v_type; + } + if (params->attn_qkv_type < GGML_TYPE_COUNT && strcmp(tensor->name, "attn_qkv.weight") == 0) { + new_type = params->attn_qkv_type; + } + if (params->attn_output_type < GGML_TYPE_COUNT && strcmp(tensor->name, "attn_output.weight") == 0) { + new_type = params->attn_output_type; + } + if (params->ffn_gate_type < GGML_TYPE_COUNT && strcmp(tensor->name, "ffn_gate") == 0) { + new_type = params->ffn_gate_type; + } + if (params->ffn_down_type < GGML_TYPE_COUNT && strcmp(tensor->name, "ffn_down") == 0) { + new_type = params->ffn_down_type; + } + if (params->ffn_up_type < GGML_TYPE_COUNT && strcmp(tensor->name, "ffn_up") == 0) { + new_type = params->ffn_up_type; + } // If we've decided to quantize to the same type the tensor is already // in then there's nothing to do. @@ -18993,6 +19045,14 @@ struct llama_model_quantize_params llama_model_quantize_default_params() { /*.ftype =*/ LLAMA_FTYPE_MOSTLY_Q5_1, /*.output_tensor_type =*/ GGML_TYPE_COUNT, /*.token_embedding_type =*/ GGML_TYPE_COUNT, + /*.attn_q_type =*/ GGML_TYPE_COUNT, + /*.attn_k_type =*/ GGML_TYPE_COUNT, + /*.attn_v_type =*/ GGML_TYPE_COUNT, + /*.attn_qkv_type =*/ GGML_TYPE_COUNT, + /*.attn_output_type =*/ GGML_TYPE_COUNT, + /*.ffn_gate_type =*/ GGML_TYPE_COUNT, + /*.ffn_down_type =*/ GGML_TYPE_COUNT, + /*.ffn_up_type =*/ GGML_TYPE_COUNT, /*.allow_requantize =*/ false, /*.quantize_output_tensor =*/ true, /*.only_copy =*/ false,