diff options
Diffstat (limited to 'src/llama.cpp')
| -rw-r--r-- | src/llama.cpp | 21 |
1 files changed, 18 insertions, 3 deletions
diff --git a/src/llama.cpp b/src/llama.cpp index 42193411..37653478 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -3884,6 +3884,7 @@ struct llama_model_loader { case GGML_TYPE_IQ5_K_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ5_K_R4;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; + case GGML_TYPE_IQ3_S_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ3_S_R4;break; case GGML_TYPE_Q4_0_4_4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_4; break; case GGML_TYPE_Q4_0_4_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_8; break; case GGML_TYPE_Q4_0_8_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_8_8; break; @@ -4618,6 +4619,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_IQ2_BN: return "IQ2_BN - 2.00 bpw Bitnet"; case LLAMA_FTYPE_MOSTLY_IQ2_BN_R4:return "IQ2_BN_R4 - 2.00 bpw Bitnet"; case LLAMA_FTYPE_MOSTLY_IQ3_S: return "IQ3_S - 3.4375 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ3_S_R4: return "IQ3_S_R4 - 3.4375 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_M: return "IQ3_S mix - 3.66 bpw"; 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"; @@ -15807,7 +15809,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS_R4) { new_type = !qs.has_output ? GGML_TYPE_IQ4_K_R4 : GGML_TYPE_Q5_K_R4; } - else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_S || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || + else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_S || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_KSS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_KS_R4) && !qs.has_output) { new_type = GGML_TYPE_IQ5_K; } @@ -15871,6 +15873,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (new_type == GGML_TYPE_IQ3_K_R4) { new_type = GGML_TYPE_IQ3_K; } + else if (new_type == GGML_TYPE_IQ3_S_R4) { + new_type = GGML_TYPE_IQ3_S; + } else if (new_type == GGML_TYPE_IQ4_K_R4) { new_type = GGML_TYPE_IQ4_K; } @@ -15955,6 +15960,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S) && qs.model.hparams.n_gqa() >= 2) { new_type = GGML_TYPE_IQ4_K; } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_S_R4 && qs.model.hparams.n_gqa() >= 2) { + new_type = GGML_TYPE_IQ4_K_R4; + } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_K && qs.model.hparams.n_gqa() >= 2) { new_type = GGML_TYPE_IQ4_K; } @@ -16008,6 +16016,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_IQ3_XXS) new_type = GGML_TYPE_IQ3_S; else if (new_type == GGML_TYPE_Q2_K_R4 || new_type == GGML_TYPE_IQ3_XXS_R4) new_type = GGML_TYPE_IQ3_K_R4; else if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_IQ3_S ) new_type = GGML_TYPE_Q4_K; + else if (new_type == GGML_TYPE_IQ3_S_R4) new_type = GGML_TYPE_Q4_K_R4; else if (new_type == GGML_TYPE_Q3_K_R4) new_type = GGML_TYPE_Q4_K_R4; else if (new_type == GGML_TYPE_Q4_K || new_type == GGML_TYPE_IQ4_XS) new_type = GGML_TYPE_Q5_K; else if (new_type == GGML_TYPE_IQ4_NL) new_type = GGML_TYPE_Q5_K; @@ -16119,7 +16128,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n 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_IQ4_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4 || - ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4) { + ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S_R4) { new_type = GGML_TYPE_Q5_K; } } else { @@ -16195,7 +16204,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n new_type == GGML_TYPE_IQ4_K_R4|| new_type == GGML_TYPE_Q8_K_R8 || new_type == GGML_TYPE_IQ3_K_R4|| new_type == GGML_TYPE_IQ2_K_R4|| new_type == GGML_TYPE_IQ5_K_R4|| new_type == GGML_TYPE_IQ4_KS_R4 || new_type == GGML_TYPE_IQ3_XXS_R4 || new_type == GGML_TYPE_IQ2_XXS_R4 || new_type == GGML_TYPE_IQ2_XS_R4 || - new_type == GGML_TYPE_IQ2_S_R4) { + new_type == GGML_TYPE_IQ2_S_R4|| new_type == GGML_TYPE_IQ3_S_R4) { int nx = tensor->ne[0]; int ny = tensor->ne[1]; if (nx % QK_K != 0) { @@ -16223,6 +16232,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n case GGML_TYPE_IQ3_XXS: case GGML_TYPE_IQ3_XXS_R4: case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ3_S_R4: case GGML_TYPE_IQ1_S: case GGML_TYPE_IQ1_M: case GGML_TYPE_Q2_K: @@ -16384,6 +16394,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_IQ5_K_R4:default_type = GGML_TYPE_IQ5_K_R4;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; + case LLAMA_FTYPE_MOSTLY_IQ3_S_R4:default_type = GGML_TYPE_IQ3_S_R4;break; case LLAMA_FTYPE_MOSTLY_IQ3_M: default_type = GGML_TYPE_IQ3_S; break; case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: default_type = GGML_TYPE_Q4_0_4_4; break; case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: default_type = GGML_TYPE_Q4_0_4_8; break; @@ -16825,6 +16836,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ3_XXS; else chunk_size_multiplier = 4; } + else if (new_type == GGML_TYPE_IQ3_S_R4) { + if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ3_S; + else chunk_size_multiplier = 4; + } else if (new_type == GGML_TYPE_BF16_R16) { if (tensor->ne[1] % 16 != 0) new_type = GGML_TYPE_BF16; else chunk_size_multiplier = 16; |