The converter's K best paths are rescored with CKIP gpt2-tiny-chinese (GPL-3.0, character-level, ~12M parameters, 22 MB as float16): score = lexicon log-prob + 0.6 x LM log-prob, with the recently committed text as context. Inference is plain Accelerate (Apple Silicon and Intel). Modes off / saver (K=3, 15 ms) / standard (K=10, 40 ms) / quality (K=30, 120 ms) in the input menu. A rerank over its time budget falls back to the lexicon; three in a row pause AI for a minute. With auto-pause on, AI pauses when the Mac is hot or the CPU is busy and drops to saver in Low Power Mode. The model loads in the background and is not loaded at all when off. Evaluation (training/lm/eval_rerank.py): chat 96.6% -> 97.5% chars, Wikipedia 92.9% -> 94.7% chars (about 25-30% fewer wrong characters). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
49 lines
1.9 KiB
Python
49 lines
1.9 KiB
Python
"""Converts the gpt2-tiny-chinese safetensors file to the input method's format.
|
|
|
|
Format (little-endian):
|
|
"AZLM" magic, uint32 header length, JSON header, then float16 tensors back to back.
|
|
Header: {"n_layer", "n_head", "n_embd", "vocab", "eps", "tensors": [[name, [shape], offset]]}
|
|
(offset in float16 elements from the start of the data). Linear weights keep the
|
|
GPT-2 Conv1D layout [in, out]. The attention mask buffers are dropped.
|
|
|
|
Usage: export_lm.py <model dir> <out.bin> (standard library only)
|
|
"""
|
|
import array
|
|
import json
|
|
import struct
|
|
import sys
|
|
|
|
|
|
def main(src, out):
|
|
cfg = json.load(open(f"{src}/config.json"))
|
|
with open(f"{src}/model.safetensors", "rb") as f:
|
|
n = struct.unpack("<Q", f.read(8))[0]
|
|
header = json.loads(f.read(n))
|
|
base = 8 + n
|
|
names = sorted(k for k, v in header.items() if k != "__metadata__" and v["dtype"] == "F32" and v["shape"])
|
|
tensors, blobs, offset = [], [], 0
|
|
for name in names:
|
|
info = header[name]
|
|
a, b = info["data_offsets"]
|
|
f.seek(base + a)
|
|
floats = array.array("f")
|
|
floats.frombytes(f.read(b - a))
|
|
if sys.byteorder != "little":
|
|
floats.byteswap()
|
|
half = struct.pack(f"<{len(floats)}e", *floats)
|
|
tensors.append([name.replace("transformer.", ""), info["shape"], offset])
|
|
blobs.append(half)
|
|
offset += len(floats)
|
|
meta = json.dumps({"n_layer": cfg["n_layer"], "n_head": cfg["n_head"], "n_embd": cfg["n_embd"],
|
|
"vocab": cfg["vocab_size"], "eps": cfg["layer_norm_epsilon"],
|
|
"tensors": tensors}).encode()
|
|
with open(out, "wb") as f:
|
|
f.write(b"AZLM" + struct.pack("<I", len(meta)) + meta)
|
|
for blob in blobs:
|
|
f.write(blob)
|
|
print(f"wrote {out}: {offset} weights")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main(sys.argv[1], sys.argv[2])
|