ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-29 19:01:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

Q2_K_R4 (#136)

Browse Source 
* q2_k_r4: Zen4

PP-512(LLaMA-3.1-8B) = 256 t/s

* q3_k_r4: AVX2

* q2_k_r4: AVX2

We get PP-512(LLaMA-3.1-8B) = 287 t/s.

Also cherry-picked the q3_k_r4 AVX2 adaptation that I somehow
forgot to push upstream.

* q2_k_r4: NEON

We get PP-512(LLaMA-3.1-8B) = 106.2 t/s.
TG-128 is 36.02 t/s, which is ~10% higher than q2_K_S.

* Make sure rows per thread are a multiple of 4

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-12-11 18:16:49 +01:00
committed by GitHub
parent 9469af87f7
commit 8c6b84220d
10 changed files with 448 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
src/llama.cpp
View File

@@ -4545,6 +4545,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_Q6_0: return "Q6_0";
case LLAMA_FTYPE_MOSTLY_Q8_0: return "Q8_0";
case LLAMA_FTYPE_MOSTLY_Q2_K: return "Q2_K - Medium";
case LLAMA_FTYPE_MOSTLY_Q2_K_R4: return "Q2_K_R4";
case LLAMA_FTYPE_MOSTLY_Q2_K_S: return "Q2_K - Small";
case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "Q3_K - Small";
case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "Q3_K - Medium";
@@ -15794,6 +15795,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (new_type == GGML_TYPE_IQ4_XS_R4) {
new_type = GGML_TYPE_IQ4_XS;
}
else if (new_type == GGML_TYPE_Q2_K_R4) {
new_type = GGML_TYPE_Q2_K;
}
else if (new_type == GGML_TYPE_Q3_K_R4) {
new_type = GGML_TYPE_Q3_K;
}
@@ -15859,6 +15863,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && qs.model.hparams.n_gqa() >= 4) {
new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_R4 && qs.model.hparams.n_gqa() >= 4) {
new_type = GGML_TYPE_Q4_K_R4;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ3_K
: !qs.has_imatrix ? GGML_TYPE_IQ3_S : GGML_TYPE_IQ3_XXS;
@@ -15950,6 +15957,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) {
if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_R4) {
if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K_R4;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && !qs.has_imatrix) {
new_type = i_layer < n_layer/8 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
}
@@ -16009,7 +16019,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S ||
ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_K ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_R4 ||
ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R4 || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_R4) {
ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R4 || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_R4 ||
ftype == LLAMA_FTYPE_MOSTLY_Q2_K_R4) {
new_type = GGML_TYPE_Q5_K;
}
} else {
@@ -16079,7 +16090,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
new_type == GGML_TYPE_IQ5_K || new_type == GGML_TYPE_IQ3_K || new_type == GGML_TYPE_Q4_K_R4 ||
new_type == GGML_TYPE_IQ6_K || new_type == GGML_TYPE_IQ4_KS || new_type == GGML_TYPE_IQ4_XS_R4 ||
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_Q5_K_R4 || new_type == GGML_TYPE_Q3_K_R4 || new_type == GGML_TYPE_Q2_K_R4) {
int nx = tensor->ne[0];
int ny = tensor->ne[1];
if (nx % QK_K != 0) {
@@ -16106,6 +16117,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q2_K_R4:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q3_K_R4:
case GGML_TYPE_IQ2_K:
@@ -16204,6 +16216,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// K-quants
case LLAMA_FTYPE_MOSTLY_Q2_K_S:
case LLAMA_FTYPE_MOSTLY_Q2_K: default_type = GGML_TYPE_Q2_K; break;
case LLAMA_FTYPE_MOSTLY_Q2_K_R4: default_type = GGML_TYPE_Q2_K_R4; break;
case LLAMA_FTYPE_MOSTLY_IQ3_XS: default_type = GGML_TYPE_IQ3_S; break;
case LLAMA_FTYPE_MOSTLY_Q3_K_S:
case LLAMA_FTYPE_MOSTLY_Q3_K_M:
@@ -16616,6 +16629,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q8_0;
else chunk_size_multiplier = 4;
}
else if (new_type == GGML_TYPE_Q2_K_R4) {
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q2_K;
else chunk_size_multiplier = 4;
}
else if (new_type == GGML_TYPE_Q3_K_R4) {
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q3_K;
else chunk_size_multiplier = 4;