Codegen,codegen2,replit等代码补全模型的尝试

1
2
3
4
5
6

cd converter/
conda activate textgen
MODEL="codegen2-3_7B"
NUM_GPUS=1
python3 huggingface_gptj_convert.py -in_file ~/ai/text-generation-webui/models/michaelfeil_codegen2-3_7B-gptj -saved_dir ../models/${MODEL}-${NUM_GPUS}gpu/fastertransformer/1 -infer_gpu_num ${NUM_GPUS}
cp models/codegen2-3_7B-1gpu/fastertransformer/config.pbtxt ../models/codegen2-3_7B-1gpu/fastertransformer/

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

checkpoint = "codet5p-2b"
device = "cuda" # for GPU usage or "cpu" for CPU usage

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForSeq2SeqLM.from_pretrained(checkpoint,
                                              torch_dtype=torch.float16,
                                              trust_remote_code=True).to(device)

encoding = tokenizer("def print_hello_world():", return_tensors="pt").to(device)
encoding['decoder_input_ids'] = encoding['input_ids'].clone()
outputs = model.generate(**encoding, max_length=15)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

1
2
3
4
5
6

# 来点新内容尝试

encoding = tokenizer("### Instruction:\n\nwrite a bubble sort method in csharp, the argument is `int[]`\n\n### Response:", return_tensors="pt").to(device)


encoding = tokenizer("using System.Collections;\nusing System.Collections.Generic;\nusing UnityEngine;\nusing NaughtyAttributes;\nusing UnityQuery;\n\npublic class CapturedEnergyBallTest : KXMonoBehaviour\n{\n    public float destroyTime = 10f;\n\n    public Transform startPosTarget;\n\n    // Start is called before the first frame update\n    void Start()\n    {\n\n    }\n\n    // Update is called once per frame\n    void Update()\n    {\n        \n    }\n\n    [Button]\n    public void MakeOne(){\n        var effectPrefab = BattleController.Instance.EffectPrefabMgr.capturedEnergyBallEffect;\n        var effect = EffectPrefabMgr.InsCommonEffect(effectPrefab);\n        effect.transform.position = startPosTarget.position;\n\n        var capturedEnergyBall = effect.GetComponent<CapturedEnergyBall>();\n        capturedEnergyBall.handTarget = this.CachedTransform;\n\n        // destroy effect in 12s", return_tensors="pt").to(device)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "teknium_Replit-v2-CodeInstruct-3B"
tokenizer = AutoTokenizer.from_pretrained(checkpoint, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
  checkpoint,
  trust_remote_code=True,
  torch_dtype=torch.bfloat16,
)
model.cuda()

inputs = tokenizer("### Instruction:\n\nwrite a bubble sort method in csharp, the argument is `int[]`\n\n### Response:\n", return_tensors="pt").to("cuda")
tokens = model.generate(
  **inputs,
  max_new_tokens=160, 
  do_sample=True, 
  use_cache=True, temperature=0.2, 
  top_p=0.9
)
print(tokenizer.decode(tokens[0], skip_special_tokens=True))

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

# version 1
from transformers import AutoModelForCausalLM, AutoConfig
import torch

checkpoint = "replit_replit-code-v1-3b"
config = AutoConfig.from_pretrained(
    checkpoint,
    trust_remote_code=True
)
config.attn_config['attn_impl'] = 'triton'

model = AutoModelForCausalLM.from_pretrained(checkpoint, config=config, trust_remote_code=True)
model.to(device='cuda:0', dtype=torch.bfloat16)

# forward pass
x = torch.tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]])
x = x.to(device='cuda:0')
y = model(x)

from transformers import AutoTokenizer
# load tokenizer
tokenizer = AutoTokenizer.from_pretrained(checkpoint, trust_remote_code=True)

# single input encoding + generation
x = tokenizer.encode('def hello():\n  print("hello world")\n', return_tensors='pt')
y = model.generate(x)

# decoding, clean_up_tokenization_spaces=False to ensure syntactical correctness
generated_code = tokenizer.decode(y[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
print(generated_code)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16

# version 2
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "replit_replit-code-v1-3b"
tokenizer = AutoTokenizer.from_pretrained(checkpoint, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(checkpoint, trust_remote_code=True)
model.cuda()

x = tokenizer.encode('### Instruction:\n\nwrite a bubble sort method in csharp, the argument is `int[]`\n\n### Response:\n', return_tensors='pt').to("cuda")
x = tokenizer.encode('def fibonacci(n): ', return_tensors='pt').to("cuda")
x = tokenizer.encode('using System.Collections;\nusing System.Collections.Generic;\nusing UnityEngine;\nusing NaughtyAttributes;\nusing UnityQuery;\n\npublic class CapturedEnergyBallTest : KXMonoBehaviour\n{\n    public float destroyTime = 10f;\n\n    public Transform startPosTarget;\n\n    // Start is called before the first frame update\n    void Start()\n    {\n\n    }\n\n    // Update is called once per frame\n    void Update()\n    {\n        \n    }\n\n    [Button]\n    public void MakeOne(){\n        var effectPrefab = BattleController.Instance.EffectPrefabMgr.capturedEnergyBallEffect;\n        var effect = EffectPrefabMgr.InsCommonEffect(effectPrefab);\n        effect.transform.position = startPosTarget.position;\n\n        var capturedEnergyBall = effect.GetComponent<CapturedEnergyBall>();\n        capturedEnergyBall.handTarget = this.CachedTransform;\n\n        // destroy effect in 12s', return_tensors='pt').to("cuda")
y = model.generate(x, max_new_tokens=120, do_sample=True, top_p=0.95, top_k=4, temperature=0.2, num_return_sequences=1, eos_token_id=tokenizer.eos_token_id)

# decoding, clean_up_tokenization_spaces=False to ensure syntactical correctness
generated_code = tokenizer.decode(y[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
print(generated_code)