ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-12 23:10:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Q4_0_R4 (#119)

Browse Source 
* Adding iq4_0_r4 - q4_0 repacked

We get PP-512(LLaMA-3.1-8B) = 278 t/s on a Ryzen-7950X CPU,
so ~5-6% faster than iq4_nl_x4.

* q4_0_r4: NEON

Here we get 115.8 t/s, so also ~5% better than iq4_nl_x4.

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-12-02 17:01:48 +01:00
committed by GitHub
parent 72d94fbf22
commit 61304f5c04
9 changed files with 304 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
src/llama.cpp
View File

@@ -3850,6 +3850,7 @@ struct llama_model_loader {
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_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_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;
@@ -4555,6 +4556,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ1_M: return "IQ1_M - 1.75 bpw";
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_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";
@@ -15771,6 +15773,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (new_type == GGML_TYPE_IQ4_NL_X4) {
new_type = GGML_TYPE_IQ4_NL;
}
else if (new_type == GGML_TYPE_Q4_0_R4) {
new_type = GGML_TYPE_Q4_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 ||
@@ -15941,6 +15946,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
// same quantization as before imatrix stuff, and b) Q4_1/Q5_1 do go crazy on ffn_down without an imatrix.
new_type = ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q5_1;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_0_R4 && qs.has_imatrix && i_layer < n_layer/8) {
new_type = GGML_TYPE_IQ4_NL_X4;
}
++qs.i_ffn_down;
} else if (name.find("attn_output.weight") != std::string::npos) {
if (qs.params->attn_output_type < GGML_TYPE_COUNT) new_type = qs.params->attn_output_type;
@@ -16160,6 +16168,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
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_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_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;
@@ -16521,6 +16530,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ4_NL;
else chunk_size_multiplier = 4;
}
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;
}
LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
fflush(stdout);