diff options
| author | Kawrakow <iwankawrakow@gmail.com> | 2024-12-03 06:15:29 +0100 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-12-03 06:15:29 +0100 |
| commit | ccec00939a30aa7762a232ac4dcadba985ef9ee4 (patch) | |
| tree | b308bd1de2b982c37fc7542374fb77efa2abb726 /src/llama.cpp | |
| parent | 239a344f9935cb614c6208000edb0815214286a1 (diff) | |
Q8_0_R4 (#120)
* Adding q8_0_r4
We get PP-512(LLaMA-3.1-8B) = 268 t/s on a Ryzen-7950X compared
to 175.6 t/s for Q8_0.
* q8_0_r4: NEON
We get PP-512(LLaMA-3.1-8B) = 112.6 t/s on M2-Max.
* q8_0_r4: Zen4 matrix-vector specialization
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Diffstat (limited to 'src/llama.cpp')
| -rw-r--r-- | src/llama.cpp | 12 |
1 files changed, 11 insertions, 1 deletions
diff --git a/src/llama.cpp b/src/llama.cpp index 9505f56f..89641f71 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -3851,6 +3851,7 @@ struct llama_model_loader { case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break; case GGML_TYPE_IQ4_NL_X4:ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL_X4;break; case GGML_TYPE_Q4_0_R4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_R4; break; + case GGML_TYPE_Q8_0_R4: ftype = LLAMA_FTYPE_MOSTLY_Q8_0_R4; break; case GGML_TYPE_IQ4_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS; break; case GGML_TYPE_IQ4_KS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_KS; break; case GGML_TYPE_IQ4_KSS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_KSS; break; @@ -4557,6 +4558,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_IQ4_NL: return "IQ4_NL - 4.5 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_NL_X4:return "IQ4_NL_X4 - 4.5 bpw"; case LLAMA_FTYPE_MOSTLY_Q4_0_R4: return "Q4_0_R4 - 4.5 bpw"; + case LLAMA_FTYPE_MOSTLY_Q8_0_R4: return "Q8_0_R4 - 8.5 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.25 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_KS: return "IQ4_KS - 4.25 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_KSS: return "IQ4_KSS - 4.0 bpw"; @@ -15745,7 +15747,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n ftype == LLAMA_FTYPE_MOSTLY_IQ4_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_KSS) && !qs.has_output) { new_type = GGML_TYPE_IQ5_K; } - else if (new_type != GGML_TYPE_Q8_0 && new_type != GGML_TYPE_IQ6_K) { + else if (new_type != GGML_TYPE_Q8_0 && new_type != GGML_TYPE_Q8_0_R4 && new_type != GGML_TYPE_IQ6_K) { new_type = GGML_TYPE_Q6_K; } } @@ -15776,6 +15778,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (new_type == GGML_TYPE_Q4_0_R4) { new_type = GGML_TYPE_Q4_0; } + else if (new_type == GGML_TYPE_Q8_0_R4) { + new_type = GGML_TYPE_Q8_0; + } } } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M || @@ -16169,6 +16174,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_IQ4_NL: default_type = GGML_TYPE_IQ4_NL; break; case LLAMA_FTYPE_MOSTLY_IQ4_NL_X4:default_type = GGML_TYPE_IQ4_NL_X4;break; case LLAMA_FTYPE_MOSTLY_Q4_0_R4: default_type = GGML_TYPE_Q4_0_R4; break; + case LLAMA_FTYPE_MOSTLY_Q8_0_R4: default_type = GGML_TYPE_Q8_0_R4; break; case LLAMA_FTYPE_MOSTLY_IQ4_XS: default_type = GGML_TYPE_IQ4_XS; break; case LLAMA_FTYPE_MOSTLY_IQ4_KS: default_type = GGML_TYPE_IQ4_KS; break; case LLAMA_FTYPE_MOSTLY_IQ4_KSS: default_type = GGML_TYPE_IQ4_KSS; break; @@ -16534,6 +16540,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_0; else chunk_size_multiplier = 4; } + if (new_type == GGML_TYPE_Q8_0_R4) { + if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q8_0; + else chunk_size_multiplier = 4; + } LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type)); fflush(stdout); |