Skip to content

Commit 415d5e4

Browse files
NexesenexNexesenex
committed
Refactor furthermore attn.v
And also lower attn_q for IQ2_XS, in order to separate it more for the quite misnamed IQ2_S
1 parent 8c8e43c commit 415d5e4

File tree

1 file changed

+26
-24
lines changed

1 file changed

+26
-24
lines changed

‎src/llama.cpp‎

Lines changed: 26 additions & 24 deletions
Original file line numberDiff line numberDiff line change
@@ -15342,20 +15342,13 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
1534215342
}
1534315343
}
1534415344
} else if (name.find("attn_v.weight") != std::string::npos) {
15345-
if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
15346-
ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
15347-
if (qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q6_K;
15348-
else {
15349-
if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_Q4_K;
15350-
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) new_type = GGML_TYPE_IQ3_XXS;
15351-
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) new_type = GGML_TYPE_IQ3_S;
15352-
}
15353-
++qs.i_attention_wv;
15354-
}
15355-
else if (qs.model.hparams.n_expert >= 4) {
15345+
if (qs.model.hparams.n_expert >= 4) {
1535615346
// for the 8-expert model, bumping this to Q8_0 trades just ~128MB
1535715347
// TODO: explore better strategies
15358-
new_type = GGML_TYPE_Q8_0;
15348+
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
15349+
ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M ||
15350+
ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ) new_type = GGML_TYPE_Q6_K;
15351+
else new_type = GGML_TYPE_Q8_0;
1535915352
}
1536015353
else if (qs.model.hparams.n_gqa() >= 7) {
1536115354
// The Llama 70B models have 8 heads sharing the same attn_v weights (-> GQA 8). As a result, the attn_v.weight tensor is
@@ -15365,20 +15358,28 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
1536515358
if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K || new_type == GGML_TYPE_IQ3_S ||
1536615359
new_type == GGML_TYPE_IQ4_XS) new_type = GGML_TYPE_Q5_K;
1536715360
}
15368-
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) {
15369-
new_type = (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
15361+
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
15362+
new_type = (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) ? GGML_TYPE_IQ4_XS : GGML_TYPE_IQ3_XXS;
1537015363
}
15371-
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
15372-
new_type = GGML_TYPE_Q4_K;
15364+
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
15365+
new_type = (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) ? GGML_TYPE_Q4_K : GGML_TYPE_IQ3_XXS;
15366+
}
15367+
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) {
15368+
new_type = (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) ? GGML_TYPE_Q4_K : GGML_TYPE_IQ3_S;
15369+
}
15370+
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) {
15371+
new_type = (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
1537315372
}
1537415373
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
15375-
new_type = (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) ? GGML_TYPE_Q4_K : !qs.has_imatrix ? GGML_TYPE_IQ3_S : GGML_TYPE_IQ3_XXS;
15374+
new_type = (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) ? GGML_TYPE_Q4_K : GGML_TYPE_IQ3_S;
1537615375
}
1537715376
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
1537815377
new_type = GGML_TYPE_Q4_K;
1537915378
}
1538015379
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_S) new_type = GGML_TYPE_Q4_K;
15381-
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M) new_type = GGML_TYPE_Q5_K;
15380+
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
15381+
new_type = (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
15382+
}
1538215383
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
1538315384
new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
1538415385
}
@@ -15394,9 +15395,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
1539415395
if (qs.model.hparams.n_expert >= 4) {
1539515396
// for the 8-expert model, bumping this to Q8_0 trades just ~128MB
1539615397
// TODO: explore better strategies
15397-
if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
15398-
ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S ||
15399-
ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q6_K;
15398+
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
15399+
ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M ||
15400+
ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ) new_type = GGML_TYPE_Q6_K;
1540015401
else new_type = GGML_TYPE_Q8_0;
1540115402
}
1540215403
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
@@ -15427,16 +15428,17 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
1542715428
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
1542815429
new_type = GGML_TYPE_Q4_K;
1542915430
}
15430-
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S ||
15431+
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S ||
1543115432
ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
1543215433
new_type = GGML_TYPE_Q5_K;
1543315434
}
1543415435
} else if (name.find("attn_q.weight") != std::string::npos) {
1543515436
if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) new_type = GGML_TYPE_IQ3_XXS;
1543615437
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) new_type = GGML_TYPE_IQ2_S;
1543715438
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
15438-
ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
15439+
ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
1543915440
if (qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q3_K;
15441+
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) new_type = GGML_TYPE_IQ2_XXS;
1544015442
}
1544115443
} else if (name.find("ffn_down") != std::string::npos) {
1544215444
auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str());
@@ -15461,7 +15463,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
1546115463
: GGML_TYPE_Q3_K;
1546215464
}
1546315465
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M && (i_layer < n_layer/8 ||
15464-
(qs.model.hparams.n_expert >= 4 && use_more_bits(i_layer, n_layer)))) {
15466+
(qs.model.hparams.n_expert >= 4 && use_more_bits(i_layer, n_layer)))) {
1546515467
new_type = GGML_TYPE_Q4_K;
1546615468
}
1546715469
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {

0 commit comments

Comments
 (0)