CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-03 09:22:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

mpt : do not duplicate token_embd.weight on disk (#5670)

Browse Source
This commit is contained in:
Jared Van Bortel
2024-02-22 17:05:23 -05:00
committed by GitHub
parent 96633eeca1
commit 15499eb942
2 changed files with 4 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -622,11 +622,6 @@ 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):

6
llama.cpp
View File

@@ -509,7 +509,6 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
{ {
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" }, { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" }, { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
@@ -4056,7 +4055,10 @@ 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) {