Skip to content

Commit

Permalink
convert : update Falcon script for new HF config (ggerganov#3448)
Browse files Browse the repository at this point in the history 
Also adds Falcon-180B support.
Closes ggerganov#3049

Co-authored-by: jb <jonathan.t.barnard@gmail.com>
  • Loading branch information
@cebtenzzre @logicchains
cebtenzzre and logicchains committed Oct 5, 2023
1 parent 45eba93 commit 48edda3
Showing 1 changed file with 79 additions and 64 deletions.
143 changes: 79 additions & 64 deletions convert-falcon-hf-to-gguf.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,7 @@
from __future__ import annotations

import argparse
import contextlib
import json
import os
import struct
Expand All @@ -20,10 +21,10 @@
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:
Expand Down Expand Up @@ -77,30 +78,36 @@ def parse_args() -> argparse.Namespace:
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
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"])
Expand Down Expand Up @@ -146,8 +153,8 @@ def parse_args() -> argparse.Namespace:
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

Expand All @@ -156,6 +163,10 @@ def parse_args() -> argparse.Namespace:

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)
Expand All @@ -165,60 +176,64 @@ def parse_args() -> argparse.Namespace:
if args.vocab_only:
break
print("gguf: loading model part '" + part_name + "'")
model_part = torch.load(dir_model / part_name, map_location="cpu")

for name in model_part.keys():
data = model_part[name]

old_dtype = data.dtype

# convert any unsupported data types to float32
if data.dtype != torch.float16 and data.dtype != torch.float32:
data = data.to(torch.float32)

# QKV tensor transform
# The original query_key_value tensor contains n_head_kv "kv groups",
# each consisting of n_head/n_head_kv query weights followed by one key
# and one value weight (shared by all query heads in the kv group).
# This layout makes it a big pain to work with in GGML.
# So we rearrange them here,, so that we have n_head query weights
# followed by n_head_kv key weights followed by n_head_kv value weights,
# in contiguous fashion.
# ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py

if "query_key_value" in name:
qkv = data.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head)
k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
data = torch.cat((q,k,v)).reshape_as(data)

data = data.squeeze().numpy()

# map tensor names
new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
if new_name is None:
print("Can not map tensor '" + name + "'")
sys.exit()

n_dims = len(data.shape)
data_dtype = data.dtype

# if f32 desired, convert any float16 to float32
if ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)

# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)

# if f16 desired, convert any float32 2-dim weight tensors to float16
if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)

print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))

gguf_writer.add_tensor(new_name, data)
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.get_tensor(name) if is_safetensors else model_part[name]

old_dtype = data.dtype

# convert any unsupported data types to float32
if data.dtype != torch.float16 and data.dtype != torch.float32:
data = data.to(torch.float32)

# QKV tensor transform
# The original query_key_value tensor contains n_head_kv "kv groups",
# each consisting of n_head/n_head_kv query weights followed by one key
# and one value weight (shared by all query heads in the kv group).
# This layout makes it a big pain to work with in GGML.
# So we rearrange them here,, so that we have n_head query weights
# followed by n_head_kv key weights followed by n_head_kv value weights,
# in contiguous fashion.
# ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py

if "query_key_value" in name:
qkv = data.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head)
k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
data = torch.cat((q,k,v)).reshape_as(data)

data = data.squeeze().numpy()

# map tensor names
new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
if new_name is None:
print("Can not map tensor '" + name + "'")
sys.exit()

n_dims = len(data.shape)
data_dtype = data.dtype

# if f32 desired, convert any float16 to float32
if ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)

# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)

# if f16 desired, convert any float32 2-dim weight tensors to float16
if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)

print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))

gguf_writer.add_tensor(new_name, data)


print("gguf: write header")
Expand Down

0 comments on commit 48edda3

Please sign in to comment.