root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 10:54:36 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

convert : update Falcon script for new HF config (#3448)

Browse Source 
Also adds Falcon-180B support.
Closes #3049

Co-authored-by: jb <jonathan.t.barnard@gmail.com>
This commit is contained in:
cebtenzzre 2023-10-05 15:00:34 -04:00 committed by GitHub
parent 45eba9369f
commit 48edda30ee
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 66 additions and 51 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

39
convert-falcon-hf-to-gguf.py
View File

@ -4,6 +4,7 @@
from __future__ import annotations
import argparse
import contextlib
import json
import os
import struct
@ -20,10 +21,10 @@ if 'NO_LOCAL_GGUF' not in os.environ:
import gguf
def count_model_parts(dir_model: Path) -> int:
def count_model_parts(dir_model: Path, prefix: str) -> int:
num_parts = 0
for filename in os.listdir(dir_model):
if filename.startswith("pytorch_model-"):
if filename.startswith(prefix):
num_parts += 1
if num_parts > 0:
@ -77,20 +78,26 @@ print("gguf: loading model "+dir_model.name)
with open(dir_model / "config.json", "r", encoding="utf-8") as f:
hparams = json.load(f)
if hparams["architectures"][0] != "RWForCausalLM":
if hparams["architectures"][0] != "FalconForCausalLM":
print("Model architecture not supported: " + hparams["architectures"][0])
sys.exit(1)
# get number of model parts
num_parts = count_model_parts(dir_model)
num_parts = count_model_parts(dir_model, "model-00")
if num_parts:
is_safetensors = True
from safetensors import safe_open
else:
is_safetensors = False
num_parts = count_model_parts(dir_model, "pytorch_model-")
ARCH=gguf.MODEL_ARCH.FALCON
gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
print("gguf: get model metadata")
block_count = hparams["n_layer"]
block_count = hparams["num_hidden_layers"]
gguf_writer.add_name("Falcon")
gguf_writer.add_context_length(2048) # not in config.json
@ -98,9 +105,9 @@ gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform
gguf_writer.add_embedding_length(hparams["hidden_size"])
gguf_writer.add_feed_forward_length(4 * hparams["hidden_size"])
gguf_writer.add_block_count(block_count)
gguf_writer.add_head_count(hparams["n_head"])
if "n_head_kv" in hparams:
gguf_writer.add_head_count_kv(hparams["n_head_kv"])
gguf_writer.add_head_count(hparams["num_attention_heads"])
if "num_kv_heads" in hparams:
gguf_writer.add_head_count_kv(hparams["num_kv_heads"])
else:
gguf_writer.add_head_count_kv(1)
gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
@ -146,8 +153,8 @@ special_vocab.add_to_gguf(gguf_writer)
tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
# params for qkv transform
n_head = hparams["n_head"]
n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
n_head = hparams["num_attention_heads"]
n_head_kv = hparams["num_kv_heads"] if "num_kv_heads" in hparams else 1
head_dim = hparams["hidden_size"] // n_head
@ -156,6 +163,10 @@ print("gguf: get tensor metadata")
if num_parts == 0:
part_names = iter(("pytorch_model.bin",))
elif is_safetensors:
part_names = (
f"model-{n:05}-of-{num_parts:05}.safetensors" for n in range(1, num_parts + 1)
)
else:
part_names = (
f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
@ -165,10 +176,14 @@ for part_name in part_names:
if args.vocab_only:
break
print("gguf: loading model part '" + part_name + "'")
model_part = torch.load(dir_model / part_name, map_location="cpu")
if is_safetensors:
ctx = safe_open(dir_model / part_name, framework="pt", device="cpu")
else:
ctx = contextlib.nullcontext(torch.load(dir_model / part_name, map_location="cpu"))
with ctx as model_part:
for name in model_part.keys():
data = model_part[name]
data = model_part.get_tensor(name) if is_safetensors else model_part[name]
old_dtype = data.dtype