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

mpt : do not duplicate token_embd.weight on disk

Browse source 
Previous attempt was #3626
This commit is contained in:
Jared Van Bortel 2024-02-22 16:02:39 -05:00
parent 201294ae17
commit 549fe807fd
2 changed files with 4 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

5
convert-hf-to-gguf.py
View file

@ -618,11 +618,6 @@ class MPTModel(Model):
self.gguf_writer.add_tensor(new_name, data) self.gguf_writer.add_tensor(new_name, data)
# note: MPT output is tied to (same as) wte in original model;
# for easier implementation in llama.cpp it's duplicated in GGUF, though :/
if new_name == "token_embd.weight":
self.gguf_writer.add_tensor("output.weight", data)
class OrionModel(Model): class OrionModel(Model):
def set_vocab(self): def set_vocab(self):

5
llama.cpp
View file

@ -4056,7 +4056,10 @@ static bool llm_load_tensors(
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, false); model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, false);
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); // same as tok_embd, duplicated to allow offloading
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
ml.n_created--; // artificial tensor
ml.size_data += ggml_nbytes(model.output);
} }
for (int i = 0; i < n_layer; ++i) { for (int i = 0; i < n_layer; ++i) {