הסקת מסקנות מ-CodeGemma באמצעות JAX ו-Flatx

import os
from google.colab import userdata # `userdata` is a Colab API.

os.environ["KAGGLE_USERNAME"] = userdata.get('KAGGLE_USERNAME')
os.environ["KAGGLE_KEY"] = userdata.get('KAGGLE_KEY')

pip install -q git+https://github.com/google-deepmind/gemma.git

import os
from gemma import params as params_lib
from gemma import sampler as sampler_lib
from gemma import transformer as transformer_lib
import sentencepiece as spm

GEMMA_VARIANT = '2b-pt' # @param ['2b-pt', '7b-it', '7b-pt', '1.1-2b-pt', '1.1-7b-it'] {type:"string"}

import kagglehub

GEMMA_PATH = kagglehub.model_download(f'google/codegemma/flax/{GEMMA_VARIANT}')

Warning: Looks like you're using an outdated `kagglehub` version, please consider updating (latest version: 0.2.7)
Downloading from https://www.kaggle.com/api/v1/models/google/codegemma/flax/2b-pt/3/download...
100%|██████████| 3.67G/3.67G [00:22<00:00, 173MB/s]
Extracting model files...

print('GEMMA_PATH:', GEMMA_PATH)

GEMMA_PATH: /root/.cache/kagglehub/models/google/codegemma/flax/2b-pt/3

CKPT_PATH = os.path.join(GEMMA_PATH, GEMMA_VARIANT[-5:])
TOKENIZER_PATH = os.path.join(GEMMA_PATH, 'spm.model')
print('CKPT_PATH:', CKPT_PATH)
print('TOKENIZER_PATH:', TOKENIZER_PATH)

CKPT_PATH: /root/.cache/kagglehub/models/google/codegemma/flax/2b-pt/3/2b-pt
TOKENIZER_PATH: /root/.cache/kagglehub/models/google/codegemma/flax/2b-pt/3/spm.model

params = params_lib.load_and_format_params(CKPT_PATH)

vocab = spm.SentencePieceProcessor()
vocab.Load(TOKENIZER_PATH)

True

transformer_config = transformer_lib.TransformerConfig.from_params(
    params,
    cache_size=1024
)

transformer = transformer_lib.Transformer(config=transformer_config)

sampler = sampler_lib.Sampler(
    transformer=transformer,
    vocab=vocab,
    params=params['transformer']
)

def function(string):
assert function('asdf') == 'fdsa'

"def function(string):\n"

"assert function('asdf') == 'fdsa'"

"<|fim_prefix|>def function(string):\n<|fim_suffix|>assert function('asdf') == 'fdsa'<|fim_middle|>"

# In the context of a code editor,
# the cursor is the location where the text will be inserted
BEFORE_CURSOR = "<|fim_prefix|>"
AFTER_CURSOR = "<|fim_suffix|>"
AT_CURSOR = "<|fim_middle|>"
FILE_SEPARATOR = "<|file_separator|>"

def format_completion_prompt(before, after):
  print(f"\nORIGINAL PROMPT:\n{before}{after}")
  prompt = f"{BEFORE_CURSOR}{before}{AFTER_CURSOR}{after}{AT_CURSOR}"
  print(f"\nFORMATTED PROMPT:\n{repr(prompt)}")
  return prompt
def format_generated_output(before, after, output):
  print(f"\nGENERATED OUTPUT:\n{repr(output)}")
  formatted_output = f"{before}{output.replace(FILE_SEPARATOR, '')}{after}"
  print(f"\nFILL-IN COMPLETION:\n{formatted_output}")
  return formatted_output

before = "def function(string):\n"
after = "assert function('asdf') == 'fdsa'"
prompt = format_completion_prompt(before, after)

output = sampler(
    [prompt],
    total_generation_steps=100,
    ).text

formatted_output = format_generated_output(before, after, output[0])

ORIGINAL PROMPT:
def function(string):
assert function('asdf') == 'fdsa'

FORMATTED PROMPT:
"<|fim_prefix|>def function(string):\n<|fim_suffix|>assert function('asdf') == 'fdsa'<|fim_middle|>"

GENERATED OUTPUT:
'    return string[::-1]\n\n<|file_separator|>'

FILL-IN COMPLETION:
def function(string):
    return string[::-1]

assert function('asdf') == 'fdsa'

before = "import "
after = """if __name__ == "__main__":\n    sys.exit(0)"""
prompt = format_completion_prompt(before, after)

output = sampler(
    [prompt],
    total_generation_steps=100,
    ).text

formatted_output = format_generated_output(before, after, output[0])

ORIGINAL PROMPT:
import if __name__ == "__main__":
    sys.exit(0)

FORMATTED PROMPT:
'<|fim_prefix|>import <|fim_suffix|>if __name__ == "__main__":\n    sys.exit(0)<|fim_middle|>'

GENERATED OUTPUT:
'sys\n<|file_separator|>'

FILL-IN COMPLETION:
import sys
if __name__ == "__main__":
    sys.exit(0)

before = """import numpy as np
def reflect(matrix):
  # horizontally reflect a matrix
"""
after = ""
prompt = format_completion_prompt(before, after)

output = sampler(
    [prompt],
    total_generation_steps=100,
    ).text

formatted_output = format_generated_output(before, after, output[0])

ORIGINAL PROMPT:
import numpy as np
def reflect(matrix):
  # horizontally reflect a matrix


FORMATTED PROMPT:
'<|fim_prefix|>import numpy as np\ndef reflect(matrix):\n  # horizontally reflect a matrix\n<|fim_suffix|><|fim_middle|>'

GENERATED OUTPUT:
'  return np.flip(matrix, axis=1)\n<|file_separator|>'

FILL-IN COMPLETION:
import numpy as np
def reflect(matrix):
  # horizontally reflect a matrix
  return np.flip(matrix, axis=1)