diff options
Diffstat (limited to 'src/llama.cpp')
| -rw-r--r-- | src/llama.cpp | 13 |
1 files changed, 12 insertions, 1 deletions
diff --git a/src/llama.cpp b/src/llama.cpp index 68e59758..f04d7f6f 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -3876,6 +3876,7 @@ struct llama_model_loader { 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_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_Q4_0_4_4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_4; break; @@ -4601,6 +4602,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { 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_IQ5_K_R4: return "IQ5_K_R4 - 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"; case LLAMA_FTYPE_MOSTLY_IQ2_BN: return "IQ2_BN - 2.00 bpw Bitnet"; @@ -15854,6 +15856,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (new_type == GGML_TYPE_IQ4_K_R4) { new_type = GGML_TYPE_IQ4_K; } + else if (new_type == GGML_TYPE_IQ5_K_R4) { + new_type = GGML_TYPE_IQ5_K; + } else if (new_type == GGML_TYPE_Q4_0_R4) { new_type = GGML_TYPE_Q4_0; } @@ -16150,7 +16155,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n 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_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_IQ2_K_R4|| new_type == GGML_TYPE_IQ5_K_R4) { int nx = tensor->ne[0]; int ny = tensor->ne[1]; if (nx % QK_K != 0) { @@ -16193,6 +16198,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n 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_IQ5_K_R4: case GGML_TYPE_Q5_K_R4: case GGML_TYPE_Q5_K: new_type = GGML_TYPE_Q6_0; break; case GGML_TYPE_IQ6_K: @@ -16325,6 +16331,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s 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_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_M: default_type = GGML_TYPE_IQ3_S; break; @@ -16741,6 +16748,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ4_K; else chunk_size_multiplier = 4; } + else if (new_type == GGML_TYPE_IQ5_K_R4) { + if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ5_K; + 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; |