ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-27 00:24:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq3_xxs_r4: 1st shot on Zen4

Browse Source 
PP-512: 107 t/s -> 137 t/s
TG-128(1 thread): 2.64 t/s -> 3.44 t/s
This commit is contained in:
Iwan Kawrakow
2024-12-19 12:19:34 +02:00
parent baa9ed4a5e
commit 19acd7ce1e
10 changed files with 252 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
src/llama.cpp
View File

@@ -3854,6 +3854,7 @@ struct llama_model_loader {
case GGML_TYPE_IQ2_KS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_KS; break;
case GGML_TYPE_IQ2_S: ftype = LLAMA_FTYPE_MOSTLY_IQ2_S; break;
case GGML_TYPE_IQ3_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS; break;
case GGML_TYPE_IQ3_XXS_R4: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4; break;
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;
@@ -4583,6 +4584,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ2_M: return "IQ2_M - 2.7 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_XS: return "IQ3_XS - 3.3 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_XXS: return "IQ3_XXS - 3.0625 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4: return "IQ3_XXS_R4 - 3.0625 bpw";
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_IQ4_NL: return "IQ4_NL - 4.5 bpw";
@@ -15792,7 +15794,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ||
ftype == LLAMA_FTYPE_MOSTLY_IQ1_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4) {
ftype == LLAMA_FTYPE_MOSTLY_IQ2_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4 ||
ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4) {
new_type = !qs.has_output ? GGML_TYPE_IQ4_K : GGML_TYPE_Q5_K;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_S || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS ||
@@ -15818,6 +15821,9 @@ 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_IQ3_XXS_R4) {
new_type = GGML_TYPE_IQ3_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_BN || ftype == LLAMA_FTYPE_MOSTLY_IQ2_BN || ftype == LLAMA_FTYPE_MOSTLY_IQ2_BN_R4) {
new_type = GGML_TYPE_IQ4_NL;
}
@@ -15930,6 +15936,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
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;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4) {
new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K_R4 : qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ3_K_R4
: !qs.has_imatrix ? GGML_TYPE_IQ3_K_R4 : GGML_TYPE_IQ3_XXS_R4;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S) && qs.model.hparams.n_gqa() >= 2) {
new_type = GGML_TYPE_IQ4_K;
}
@@ -15984,6 +15994,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
}
else if (qs.model.hparams.n_gqa() >= 4) {
if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_IQ3_XXS) new_type = GGML_TYPE_IQ3_S;
else if (new_type == GGML_TYPE_Q2_K_R4 || new_type == GGML_TYPE_IQ3_XXS_R4) new_type = GGML_TYPE_IQ3_K_R4;
else if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_IQ3_S ) new_type = GGML_TYPE_Q4_K;
else if (new_type == GGML_TYPE_Q3_K_R4) new_type = GGML_TYPE_Q4_K_R4;
else if (new_type == GGML_TYPE_Q4_K || new_type == GGML_TYPE_IQ4_XS) new_type = GGML_TYPE_Q5_K;
@@ -16003,7 +16014,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
new_type = GGML_TYPE_IQ3_XXS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4) {
new_type = GGML_TYPE_IQ2_S;
}
} else if (name.find("attn_q.weight") != std::string::npos) {
@@ -16011,7 +16022,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
new_type = GGML_TYPE_IQ3_XXS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4) {
new_type = GGML_TYPE_IQ2_S;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) {
@@ -16032,6 +16043,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
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;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 && !qs.has_imatrix) {
new_type = i_layer < n_layer/8 ? GGML_TYPE_Q4_K_R4 : GGML_TYPE_IQ3_K_R4;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = i_layer < n_layer/16 ? GGML_TYPE_Q5_K
: arch != LLM_ARCH_FALCON || use_more_bits(i_layer, n_layer) ? GGML_TYPE_Q4_K
@@ -16093,12 +16107,13 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
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_Q2_K_R4|| ftype == LLAMA_FTYPE_MOSTLY_IQ4_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4 ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4) {
ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4) {
new_type = GGML_TYPE_Q5_K;
}
} else {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) new_type = GGML_TYPE_IQ3_S;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4) new_type = GGML_TYPE_IQ3_K_R4;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ) new_type = GGML_TYPE_Q4_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L ) new_type = GGML_TYPE_Q5_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M ) new_type = GGML_TYPE_IQ4_K;
@@ -16166,7 +16181,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
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_Q2_K_R4 ||
new_type == GGML_TYPE_IQ4_K_R4|| new_type == GGML_TYPE_Q8_K_R8 || new_type == GGML_TYPE_IQ3_K_R4||
new_type == GGML_TYPE_IQ2_K_R4|| new_type == GGML_TYPE_IQ5_K_R4|| new_type == GGML_TYPE_IQ4_KS_R4) {
new_type == GGML_TYPE_IQ2_K_R4|| new_type == GGML_TYPE_IQ5_K_R4|| new_type == GGML_TYPE_IQ4_KS_R4 ||
new_type == GGML_TYPE_IQ3_XXS_R4) {
int nx = tensor->ne[0];
int ny = tensor->ne[1];
if (nx % QK_K != 0) {
@@ -16189,6 +16205,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
case GGML_TYPE_IQ2_KS:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ3_XXS_R4:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
@@ -16320,6 +16337,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_IQ2_S: default_type = GGML_TYPE_IQ2_XS; break;
case LLAMA_FTYPE_MOSTLY_IQ2_M: default_type = GGML_TYPE_IQ2_S; break;
case LLAMA_FTYPE_MOSTLY_IQ3_XXS: default_type = GGML_TYPE_IQ3_XXS; break;
case LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4: default_type = GGML_TYPE_IQ3_XXS_R4; break;
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;
@@ -16769,6 +16787,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ4_KS;
else chunk_size_multiplier = 4;
}
else if (new_type == GGML_TYPE_IQ3_XXS_R4) {
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ3_XXS;
else chunk_size_multiplier = 4;
}
else if (new_type == GGML_TYPE_BF16_R16) {
if (tensor->ne[1] % 16 != 0) new_type = GGML_TYPE_BF16;
else chunk_size_multiplier = 16;