diff options
| author | Kawrakow <iwankawrakow@gmail.com> | 2024-12-03 12:59:22 +0100 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-12-03 12:59:22 +0100 |
| commit | c5bf589367cd609f4c0ff73a6534bbde7902abe8 (patch) | |
| tree | fa17f82c717d535222c1843fc9fca2d66f4d6ea7 /src/llama.cpp | |
| parent | ccec00939a30aa7762a232ac4dcadba985ef9ee4 (diff) | |
Q5_0_R4 (#121)
* Adding q5_0_r4
We get PP-512(LLaMA-3.1-8B) = 256.7 t/s on a Ryzen-7950X.
We even get TG-128 improvement to 11.7 t/s from 11.1 t/s.
* q5_0_r4: NEON
We get PP-512(LLaMA-3.1-8B) = 99.6 t/s on M2-Max,
up from 71.0 t/s for Q5_0. The difference to mainline llama.cpp
is no longer funny: they get 26.5 t/s for Q5_0.
For TG, we are nor able to fully saturate memory bandwidth
and arrive at 22.1 t/s @ 8 threads. Mainline llama.cpp gets
20.6 t/s for Q5_0.
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Diffstat (limited to 'src/llama.cpp')
| -rw-r--r-- | src/llama.cpp | 16 |
1 files changed, 13 insertions, 3 deletions
diff --git a/src/llama.cpp b/src/llama.cpp index 89641f71..51c7d1f8 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_Q5_0_R4: ftype = LLAMA_FTYPE_MOSTLY_Q5_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; @@ -4558,6 +4559,7 @@ 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_Q5_0_R4: return "Q5_0_R4 - 5.5 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"; @@ -15778,6 +15780,9 @@ 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_Q5_0_R4) { + new_type = GGML_TYPE_Q5_0; + } else if (new_type == GGML_TYPE_Q8_0_R4) { new_type = GGML_TYPE_Q8_0; } @@ -16174,6 +16179,7 @@ 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_Q5_0_R4: default_type = GGML_TYPE_Q5_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; @@ -16532,15 +16538,19 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (new_type == GGML_TYPE_Q4_0_8_8) chunk_size_multiplier = 8; else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8) chunk_size_multiplier = 4; } - if (new_type == GGML_TYPE_IQ4_NL_X4) { + else if (new_type == GGML_TYPE_IQ4_NL_X4) { if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ4_NL; else chunk_size_multiplier = 4; } - if (new_type == GGML_TYPE_Q4_0_R4) { + else if (new_type == GGML_TYPE_Q4_0_R4) { 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) { + else if (new_type == GGML_TYPE_Q5_0_R4) { + if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q5_0; + else chunk_size_multiplier = 4; + } + else 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; } |