diff options
Diffstat (limited to 'src/llama.cpp')
| -rw-r--r-- | src/llama.cpp | 19 |
1 files changed, 17 insertions, 2 deletions
diff --git a/src/llama.cpp b/src/llama.cpp index 6ecf0452..9356c639 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -3866,6 +3866,7 @@ struct llama_model_loader { case GGML_TYPE_IQ2_K: ftype = LLAMA_FTYPE_MOSTLY_IQ2_K; break; case GGML_TYPE_IQ3_K: ftype = LLAMA_FTYPE_MOSTLY_IQ3_K; break; case GGML_TYPE_IQ4_K: ftype = LLAMA_FTYPE_MOSTLY_IQ4_K; break; + case GGML_TYPE_IQ4_K_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ4_K_R4;break; case GGML_TYPE_IQ5_K: ftype = LLAMA_FTYPE_MOSTLY_IQ5_K; break; case GGML_TYPE_IQ6_K: ftype = LLAMA_FTYPE_MOSTLY_IQ6_K; break; case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break; @@ -4582,6 +4583,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_IQ3_K: return "IQ3_K - 3.4325 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_KL: return "IQ3_KL - 4 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_K: return "IQ4_K - 4.5 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ4_K_R4: return "IQ4_K_R4 - 4.5 bpw"; case LLAMA_FTYPE_MOSTLY_IQ5_K: return "IQ5_K - 5.5 bpw"; case LLAMA_FTYPE_MOSTLY_IQ6_K: return "IQ6_K - 6.6 bpw"; case LLAMA_FTYPE_MOSTLY_IQ1_BN: return "IQ1_BN - 1.625 bpw Bitnet"; @@ -15810,6 +15812,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (new_type == GGML_TYPE_Q6_K_R4) { new_type = GGML_TYPE_Q6_K; } + else if (new_type == GGML_TYPE_IQ4_K_R4) { + new_type = GGML_TYPE_IQ4_K; + } else if (new_type == GGML_TYPE_Q4_0_R4) { new_type = GGML_TYPE_Q4_0; } @@ -15894,6 +15899,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_K && qs.model.hparams.n_gqa() >= 2) { new_type = GGML_TYPE_IQ5_K; } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_K_R4 && qs.model.hparams.n_gqa() >= 2) { + new_type = GGML_TYPE_IQ5_K; + } 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; @@ -16020,7 +16028,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n 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_Q4_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R4 || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_R4 || - ftype == LLAMA_FTYPE_MOSTLY_Q2_K_R4) { + ftype == LLAMA_FTYPE_MOSTLY_Q2_K_R4|| ftype == LLAMA_FTYPE_MOSTLY_IQ4_K_R4) { new_type = GGML_TYPE_Q5_K; } } else { @@ -16090,7 +16098,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n 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 || new_type == GGML_TYPE_Q6_K_R4 || - new_type == GGML_TYPE_Q5_K_R4 || new_type == GGML_TYPE_Q3_K_R4 || new_type == GGML_TYPE_Q2_K_R4) { + new_type == GGML_TYPE_Q5_K_R4 || new_type == GGML_TYPE_Q3_K_R4 || new_type == GGML_TYPE_Q2_K_R4 || + new_type == GGML_TYPE_IQ4_K_R4) { int nx = tensor->ne[0]; int ny = tensor->ne[1]; if (nx % QK_K != 0) { @@ -16127,6 +16136,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n 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_IQ4_K_R4: case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break; case GGML_TYPE_IQ5_K: @@ -16255,6 +16265,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_IQ3_K: default_type = GGML_TYPE_IQ3_K; break; case LLAMA_FTYPE_MOSTLY_IQ3_KL: default_type = GGML_TYPE_IQ3_K; break; case LLAMA_FTYPE_MOSTLY_IQ4_K: default_type = GGML_TYPE_IQ4_K; break; + case LLAMA_FTYPE_MOSTLY_IQ4_K_R4:default_type = GGML_TYPE_IQ4_K_R4;break; case LLAMA_FTYPE_MOSTLY_IQ5_K: default_type = GGML_TYPE_IQ5_K; break; case LLAMA_FTYPE_MOSTLY_IQ6_K: default_type = GGML_TYPE_IQ6_K; break; case LLAMA_FTYPE_MOSTLY_IQ3_S: default_type = GGML_TYPE_IQ3_S; break; @@ -16653,6 +16664,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ2_BN; else chunk_size_multiplier = 4; } + else if (new_type == GGML_TYPE_IQ4_K_R4) { + if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ4_K; + else chunk_size_multiplier = 4; + } LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type)); fflush(stdout); |