CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : fix quantization of shared token_embd (#5944)

Browse Source
This commit is contained in:
compilade
2024-03-08 10:53:37 -05:00
committed by GitHub
parent 76e868821a
commit 515f7d0d4f
1 changed files with 7 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
llama.cpp
View File

@@ -10973,6 +10973,9 @@ struct quantize_state_internal {
bool has_imatrix = false; bool has_imatrix = false;
// used to figure out if a model shares tok_embd with the output weight
bool has_output = false;
quantize_state_internal(const llama_model & model, const llama_model_quantize_params * params) quantize_state_internal(const llama_model & model, const llama_model_quantize_params * params)
: model(model) : model(model)
, params(params) , params(params)
@@ -11070,8 +11073,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
// for arches that share the same tensor between the token embeddings and the output, we quantize the token embeddings // for arches that share the same tensor between the token embeddings and the output, we quantize the token embeddings
// with the quantization of the output tensor // with the quantization of the output tensor
if (name == tn(LLM_TENSOR_OUTPUT, "weight") || if (name == tn(LLM_TENSOR_OUTPUT, "weight") || (!qs.has_output && name == tn(LLM_TENSOR_TOKEN_EMBD, "weight"))) {
(LLM_TENSOR_NAMES.at(arch).find(LLM_TENSOR_OUTPUT) == LLM_TENSOR_NAMES.at(arch).end() && name == "token_embd.weight")) {
int nx = tensor->ne[0]; int nx = tensor->ne[0];
if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) { if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
new_type = GGML_TYPE_Q8_0; new_type = GGML_TYPE_Q8_0;
@@ -11460,6 +11462,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
else if (name.find("ffn_up") != std::string::npos) { else if (name.find("ffn_up") != std::string::npos) {
++qs.n_ffn_up; ++qs.n_ffn_up;
} }
else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
qs.has_output = true;
}
} }
if (qs.n_attention_wv != qs.n_ffn_down || (uint32_t)qs.n_attention_wv != model.hparams.n_layer) { 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", LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_ffn_down = %d, hparams.n_layer = %d\n",