diff options
Diffstat (limited to 'src/llama.cpp')
| -rw-r--r-- | src/llama.cpp | 22 |
1 files changed, 19 insertions, 3 deletions
diff --git a/src/llama.cpp b/src/llama.cpp index 4b25d4ec..18c6e111 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -3837,6 +3837,7 @@ struct llama_model_loader { case GGML_TYPE_Q2_K: ftype = LLAMA_FTYPE_MOSTLY_Q2_K; break; case GGML_TYPE_Q3_K: ftype = LLAMA_FTYPE_MOSTLY_Q3_K_M; break; case GGML_TYPE_Q4_K: ftype = LLAMA_FTYPE_MOSTLY_Q4_K_M; break; + case GGML_TYPE_Q4_K_R4: ftype = LLAMA_FTYPE_MOSTLY_Q4_K_R4; break; case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break; case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K; break; case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break; @@ -4546,6 +4547,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "Q3_K - Medium"; case LLAMA_FTYPE_MOSTLY_Q3_K_L: return "Q3_K - Large"; case LLAMA_FTYPE_MOSTLY_Q4_K_S: return "Q4_K - Small"; + case LLAMA_FTYPE_MOSTLY_Q4_K_R4: return "Q4_K_R4"; case LLAMA_FTYPE_MOSTLY_Q4_K_M: return "Q4_K - Medium"; case LLAMA_FTYPE_MOSTLY_Q5_K_S: return "Q5_K - Small"; case LLAMA_FTYPE_MOSTLY_Q5_K_M: return "Q5_K - Medium"; @@ -15786,6 +15788,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (new_type == GGML_TYPE_IQ4_XS_R4) { new_type = GGML_TYPE_IQ4_XS; } + else if (new_type == GGML_TYPE_Q4_K_R4) { + new_type = GGML_TYPE_Q4_K; + } else if (new_type == GGML_TYPE_Q4_0_R4) { new_type = GGML_TYPE_Q4_0; } @@ -15870,6 +15875,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) && use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_R4 && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) { if (qs.model.hparams.n_vocab >= 127999 && (qs.model.type == MODEL_8B || qs.model.type == MODEL_70B)) new_type = GGML_TYPE_Q6_K; @@ -15964,6 +15970,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) { new_type = GGML_TYPE_Q5_K; } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_R4 && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) { + new_type = GGML_TYPE_Q5_K; + } else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_0 || ftype == LLAMA_FTYPE_MOSTLY_Q5_0) && qs.has_imatrix && i_layer < n_layer/8) { // Guard against craziness in the first few ffn_down layers that can happen even with imatrix for Q4_0/Q5_0. @@ -15983,7 +15992,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_K || - ftype == LLAMA_FTYPE_MOSTLY_IQ2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K || + ftype == LLAMA_FTYPE_MOSTLY_IQ2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R4) { new_type = GGML_TYPE_Q5_K; } @@ -16051,8 +16060,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_S || new_type == GGML_TYPE_IQ3_XXS || new_type == GGML_TYPE_IQ1_S || new_type == GGML_TYPE_IQ3_S || new_type == GGML_TYPE_IQ1_M || new_type == GGML_TYPE_IQ4_K || new_type == GGML_TYPE_IQ2_K || - new_type == GGML_TYPE_IQ5_K || new_type == GGML_TYPE_IQ3_K || - new_type == GGML_TYPE_IQ6_K || new_type == GGML_TYPE_IQ4_KS || + new_type == GGML_TYPE_IQ5_K || new_type == GGML_TYPE_IQ3_K || new_type == GGML_TYPE_Q4_K_R4 || + new_type == GGML_TYPE_IQ6_K || new_type == GGML_TYPE_IQ4_KS || new_type == GGML_TYPE_IQ4_XS_R4 || new_type == GGML_TYPE_IQ2_KS || new_type == GGML_TYPE_IQ4_KSS) { int nx = tensor->ne[0]; int ny = tensor->ne[1]; @@ -16085,8 +16094,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_KSS: case GGML_TYPE_IQ4_KS: + case GGML_TYPE_IQ4_XS_R4: case GGML_TYPE_IQ4_XS: new_type = GGML_TYPE_IQ4_NL; break; case GGML_TYPE_IQ4_K: + case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break; case GGML_TYPE_IQ5_K: case GGML_TYPE_Q5_K: new_type = GGML_TYPE_Q6_0; break; @@ -16179,6 +16190,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s 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: default_type = GGML_TYPE_Q4_K; break; + case LLAMA_FTYPE_MOSTLY_Q4_K_R4: default_type = GGML_TYPE_Q4_K_R4; break; case LLAMA_FTYPE_MOSTLY_Q5_K_S: 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; @@ -16581,6 +16593,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q8_0; else chunk_size_multiplier = 4; } + else if (new_type == GGML_TYPE_Q4_K_R4) { + if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_K; + else chunk_size_multiplier = 4; + } else if (new_type == GGML_TYPE_IQ2_BN_R4) { if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ2_BN; else chunk_size_multiplier = 4; |