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

convert-llama-h5-to-gguf.py : support alt ctx param name

Browse source
This commit is contained in:
klosax 2023-08-16 17:10:29 +02:00 • committed by GitHub
parent 6412e97427
commit f3e90f27de
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 10 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

11
convert-llama-h5-to-gguf.py
View file

@ -95,12 +95,21 @@ if "_name_or_path" in hparams:
else: else:
hf_repo="" hf_repo=""
if "max_sequence_length" in hparams:
ctx_length = hparams["max_sequence_length"]
elif "max_position_embeddings" in hparams:
ctx_length = hparams["max_position_embeddings"]
else:
print("gguf: can not find ctx length parameter.")
sys.exit()
gguf_writer.add_architecture(llm_arch) gguf_writer.add_architecture(llm_arch)
gguf_writer.add_name(last_dir) gguf_writer.add_name(last_dir)
gguf_writer.add_file_type("All tensors F32" if ftype == 0 else "Most tensors F16, some F32") gguf_writer.add_file_type("All tensors F32" if ftype == 0 else "Most tensors F16, some F32")
gguf_writer.add_source_hf_repo(hf_repo) gguf_writer.add_source_hf_repo(hf_repo)
gguf_writer.add_tensor_data_layout(llm_arch, "Meta AI original pth") gguf_writer.add_tensor_data_layout(llm_arch, "Meta AI original pth")
gguf_writer.add_context_length(llm_arch, hparams["max_position_embeddings"]) gguf_writer.add_context_length(llm_arch, ctx_length)
gguf_writer.add_embedding_length(llm_arch, hparams["hidden_size"]) gguf_writer.add_embedding_length(llm_arch, hparams["hidden_size"])
gguf_writer.add_block_count(llm_arch, block_count) gguf_writer.add_block_count(llm_arch, block_count)
gguf_writer.add_feed_forward_length(llm_arch, hparams["intermediate_size"]) gguf_writer.add_feed_forward_length(llm_arch, hparams["intermediate_size"])