github/gpt4all
1
0
Fork 0
mirror of https://github.com/nomic-ai/gpt4all.git synced 2025-06-26 07:23:38 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: convert script working

Browse Source
This commit is contained in:
Zach Nussbaum 2023-05-07 18:18:04 -04:00 committed by Adam Treat
parent d928540a08
commit 28f0f76b9f
1 changed files with 16 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
convert_hf_to_ggml.py
View File

@ -73,23 +73,28 @@ print("Model loaded: ", model_name)
fname_out = dir_out + f"/ggml-model-{model_name.split('/')[-1]}-{ftype_str[ftype]}.bin" fname_out = dir_out + f"/ggml-model-{model_name.split('/')[-1]}-{ftype_str[ftype]}.bin"
fout = open(fname_out, "wb") fout = open(fname_out, "wb")
vocab = tokenizer.vocab
hparams["multiple_of"] = 1 hparams["multiple_of"] = 1
fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex fout.write(struct.pack("i", 0x67676d6d)) # magic: ggml in hex
fout.write(struct.pack("i", hparams["vocab_size"])) fout.write(struct.pack("i", hparams["vocab_size"]))
# fout.write(struct.pack("i", hparams["seq_length"])) fout.write(struct.pack("i", hparams["max_seq_len"]))
fout.write(struct.pack("i", hparams["d_model"])) fout.write(struct.pack("i", hparams["d_model"]))
fout.write(struct.pack("i", hparams["multiple_of"]))
fout.write(struct.pack("i", hparams["n_heads"])) fout.write(struct.pack("i", hparams["n_heads"]))
fout.write(struct.pack("i", hparams["n_layers"])) fout.write(struct.pack("i", hparams["n_layers"]))
# n_rot (unused)
fout.write(struct.pack("i", 0))
fout.write(struct.pack("i", ftype)) fout.write(struct.pack("i", ftype))
# # Is this correct?? # # Is this correct??
# dot_token = tokenizer.encode(".")[0] # dot_token = tokenizer.encode(".")[0]
# for i in range(hparams["vocab_size"]): # write tokens to ggml file
# text = tokenizer.decode([i]).encode('utf-8') fout.write(struct.pack("i", hparams["vocab_size"]))
# fout.write(struct.pack("i", len(text)))
# fout.write(text) for i in range(hparams["vocab_size"]):
text = tokenizer.decode([i]).encode('utf-8')
fout.write(struct.pack("i", len(text)))
fout.write(text)
list_vars = model.state_dict() list_vars = model.state_dict()
for name in list_vars.keys(): for name in list_vars.keys():
@ -107,21 +112,16 @@ for name in list_vars.keys():
new_name = name.split("Wqkv.weight")[0] new_name = name.split("Wqkv.weight")[0]
for (data, name) in [(query, new_name + "q_proj_w"), (key, new_name + "k_proj_w"), (value, new_name + "v_proj_w")]: for (data, name) in [(query, new_name + "q_proj.weight"), (key, new_name + "k_proj.weight"), (value, new_name + "v_proj.weight")]:
print(f"Processing variable: {name} with shape: {data.shape}") print(f"Processing variable: {name} with shape: {data.shape}")
n_dims = len(data.shape); n_dims = len(data.shape);
# ftype == 0 -> float32, ftype == 1 -> float16 # ftype == 0 -> float32, ftype == 1 -> float16
ftype_cur = 0; ftype_cur = 0;
if ftype != 0: if ftype != 0:
if name[-7:] == ".weight" and n_dims == 2:
print(" Converting to float16") print(" Converting to float16")
data = data.astype(np.float16) data = data.astype(np.float16)
ftype_cur = 1 ftype_cur = 1
else:
print(" Converting to float32")
data = data.astype(np.float32)
ftype_cur = 0
else: else:
if data.dtype != np.float32: if data.dtype != np.float32:
print(" Converting to float32") print(" Converting to float32")
@ -137,7 +137,6 @@ for name in list_vars.keys():
# data # data
data.tofile(fout) data.tofile(fout)
fout.write(struct.pack("i", len(data)))
else: else: