vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : remove ffn tensor counting + add sanity check

Browse source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-04-03 11:03:27 +03:00
parent 19dafafd5f
commit 3779b984ac
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 3 additions and 17 deletions
Download patch file Download diff file Expand all files Collapse all files

20
llama.cpp
View file

@ -13441,29 +13441,15 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// TODO: avoid hardcoded tensor names - use the TN_* constants
if (name.find("attn_v.weight") != std::string::npos || name.find("attn_qkv.weight") != std::string::npos) {
++qs.n_attention_wv;
} else if (name.find("ffn_down") != std::string::npos) {
++qs.n_ffn_down;
} else if (name.find("ffn_gate") != std::string::npos) {
++qs.n_ffn_gate;
} else if (name.find("ffn_up") != std::string::npos) {
++qs.n_ffn_up;
} else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
qs.has_output = true;
}
}
// REVIEW: i do not undertand why there is logic for counting the number of layers by counting the number of tensors
// instead of just using the n_layer metadata
// without this change, it would require different logic for merged experts and split experts models,
// as split expert models end with a ffn_* count n_expert times higher than the real number of layers,
// which then is corrected in layer_info by dividing the value by n_expert
// this code needs to be refactored
qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
//if (qs.n_ffn_down )
//if (qs.n_attention_wv != qs.n_ffn_down || (uint32_t) qs.n_attention_wv != model.hparams.n_layer) {
// LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_ffn_down = %d, hparams.n_layer = %d\n",
// __func__, qs.n_attention_wv, qs.n_ffn_down, model.hparams.n_layer);
//}
// sanity checks
GGML_ASSERT(qs.n_attention_wv == (int)model.hparams.n_layer && "n_attention_wv != n_layer is unexpected");
size_t total_size_org = 0;
size_t total_size_new = 0;