diff options
| author | Iwan Kawrakow <iwan.kawrakow@gmail.com> | 2024-06-17 18:41:30 +0300 |
|---|---|---|
| committer | Iwan Kawrakow <iwan.kawrakow@gmail.com> | 2024-06-22 12:02:51 +0300 |
| commit | f6863cfa1bbc5ac42b78837b355e45d82246a472 (patch) | |
| tree | b4718a3e7f7796f2bd5f8f4e6f81a812eeadae32 /llama.cpp | |
| parent | 765622ff8f921319f6b30c556b378d8320500c95 (diff) | |
bitnet: add 2 bpw quantization
The scalar dot product already chieves 37 t/s for TG!
Diffstat (limited to 'llama.cpp')
| -rw-r--r-- | llama.cpp | 7 |
1 files changed, 5 insertions, 2 deletions
@@ -3503,6 +3503,7 @@ struct llama_model_loader { case GGML_TYPE_IQ1_S: ftype = LLAMA_FTYPE_MOSTLY_IQ1_S; break; case GGML_TYPE_IQ1_M: ftype = LLAMA_FTYPE_MOSTLY_IQ1_M; break; case GGML_TYPE_IQ1_BN: ftype = LLAMA_FTYPE_MOSTLY_IQ1_BN; break; + case GGML_TYPE_IQ2_BN: ftype = LLAMA_FTYPE_MOSTLY_IQ2_BN; break; case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break; case GGML_TYPE_IQ4_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS; break; case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break; @@ -4121,6 +4122,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_IQ1_S :return "IQ1_S - 1.5625 bpw"; case LLAMA_FTYPE_MOSTLY_IQ1_M :return "IQ1_M - 1.75 bpw"; case LLAMA_FTYPE_MOSTLY_IQ1_BN :return "IQ1_BN - 1.75 bpw Bitnet"; + case LLAMA_FTYPE_MOSTLY_IQ2_BN :return "IQ2_BN - 2.00 bpw Bitnet"; case LLAMA_FTYPE_MOSTLY_IQ4_NL: return "IQ4_NL - 4.5 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.25 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_S: return "IQ3_S - 3.4375 bpw"; @@ -15473,7 +15475,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) { new_type = GGML_TYPE_IQ3_S; } - else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_BN) { + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_BN || ftype == LLAMA_FTYPE_MOSTLY_IQ2_BN) { new_type = GGML_TYPE_IQ4_NL; } } @@ -15673,7 +15675,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n ++qs.n_k_quantized; } } - if (new_type == GGML_TYPE_IQ1_BN) { + if (new_type == GGML_TYPE_IQ1_BN || new_type == GGML_TYPE_IQ2_BN) { int nx = tensor->ne[0]; if (nx % QK_IQ1BN != 0) { convert_incompatible_tensor = true; @@ -15791,6 +15793,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_IQ1_S: default_type = GGML_TYPE_IQ1_S; break; case LLAMA_FTYPE_MOSTLY_IQ1_M: default_type = GGML_TYPE_IQ1_M; break; case LLAMA_FTYPE_MOSTLY_IQ1_BN: default_type = GGML_TYPE_IQ1_BN; break; + case LLAMA_FTYPE_MOSTLY_IQ2_BN: default_type = GGML_TYPE_IQ2_BN; break; case LLAMA_FTYPE_MOSTLY_IQ4_NL: default_type = GGML_TYPE_IQ4_NL; break; case LLAMA_FTYPE_MOSTLY_IQ4_XS: default_type = GGML_TYPE_IQ4_XS; break; case LLAMA_FTYPE_MOSTLY_IQ3_S: default_type = GGML_TYPE_IQ3_S; break; |