Slight reorder of the attn.weight tree

And application of the attn.v.weight logic I used for IQ2 and IQ3, but only when such logic is already implied by the existing quant strategies, as a compromise to not disturb too much Ikawrakow's quant strategies.
This commit is contained in:
Nexes the Old 2024-08-03 01:35:08 +02:00 committed by GitHub
parent 63986631a3
commit 7d337d0f89
No known key found for this signature in database
GPG key ID: B5690EEEBB952194

View file

@ -15370,7 +15370,19 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
}
}
} else if (name.find("attn_v.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) {
if (qs.model.hparams.n_expert >= 8) {
// for the 8-expert model, bumping this to Q8_0 trades just ~128MB
// TODO: explore better strategies
new_type = GGML_TYPE_Q8_0;
}
else if (qs.model.hparams.n_gqa() >= 7) {
// The Llama 70B models have 8 heads sharing the same attn_v weights (-> GQA 8). As a result, the attn_v.weight tensor is
// 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
// nearly negligible increase in model size by quantizing this tensor with more bits.
// That logic applies also to models like Yi 34B (-> GQA 7) and Mistral Large 123B (-> GQA 12).
if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) {
new_type = (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
@ -15389,26 +15401,15 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : 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_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) && qs.model.hparams.n_gqa() >= 4) {
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS) && (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2)) {
new_type = GGML_TYPE_Q5_K;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
if (qs.model.type == MODEL_70B) {
// In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
// 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
// nearly negligible increase in model size by quantizing this tensor with more bits:
if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K;
}
if (qs.model.hparams.n_expert == 8) {
// for the 8-expert model, bumping this to Q8_0 trades just ~128MB
// TODO: explore better strategies
new_type = GGML_TYPE_Q8_0;
}
++qs.i_attention_wv;
} else if (name.find("attn_k.weight") != std::string::npos) {
if (qs.model.hparams.n_expert == 8) {
if (qs.model.hparams.n_expert >= 8) {
// for the 8-expert model, bumping this to Q8_0 trades just ~128MB
// TODO: explore better strategies
new_type = GGML_TYPE_Q8_0;