Files
2024-09-25 18:04:29 +08:00

323 lines
9.8 KiB
Python

import json
import os
from abc import ABC, abstractmethod
from typing import List
from warnings import warn
import torch
from stop_sequencer import StopSequencer
from tqdm import tqdm
from transformers import AutoModelForCausalLM, AutoTokenizer
from vllm import LLM, SamplingParams
EOS = [
"<|endoftext|>",
"<|endofmask|>",
"</s>",
"\nif __name__",
"\ndef main(",
"\nprint(",
]
def extra_eos_for_direct_completion(dataset) -> List[str]:
print(f"eos: {dataset = }")
if dataset.lower() == "humaneval":
return ["\ndef ", "\nclass ", "\nimport ", "\nfrom ", "\nassert "]
elif dataset.lower() == "mbpp":
return ['\n"""', "\nassert"]
raise ValueError(f"Unknown dataset: {dataset}")
# some random words which serves as the splitter
_MAGIC_SPLITTER_ = "-[[]]-this-is-really-our-highest-priority-[[]]-"
def make_chat_prompt(
task_prompt: str,
instruction_prefix: str,
response_prefix: str,
tokenizer: AutoTokenizer,
chat_mode=False,
) -> str:
# if tokenizer.chat_template is None:
if not chat_mode:
return task_prompt
assert instruction_prefix is not None, "Instruction prefix is required!"
assert response_prefix is not None, "Response prefix is required!"
task_prompt = f"""\
{instruction_prefix}
```
{task_prompt.strip()}
```
"""
response = f"""\
{response_prefix}
```python
{_MAGIC_SPLITTER_}
```
"""
task_prompt = tokenizer.apply_chat_template(
[
{"role": "user", "content": task_prompt},
{"role": "assistant", "content": response},
],
tokenize=False,
).split(_MAGIC_SPLITTER_)[0]
return task_prompt
class DecoderBase(ABC):
def __init__(
self,
name: str,
batch_size: int = 1,
temperature: float = 0.8,
max_new_tokens: int = 768,
dtype: str = "bfloat16", # default
trust_remote_code: bool = True,
instruction_prefix: str = None,
response_prefix: str = None,
chat_mode=False,
) -> None:
print("Initializing a decoder model: {} ...".format(name))
self.name = name
self.batch_size = batch_size
self.temperature = temperature
self.eos = EOS
self.skip_special_tokens = False
self.max_new_tokens = max_new_tokens
self.dtype = dtype
self.trust_remote_code = trust_remote_code
self.instruction_prefix = instruction_prefix
self.response_prefix = response_prefix
self.chat_mode = chat_mode
print(f"{self.chat_mode = }")
@abstractmethod
def codegen(self, prompt: str, do_sample: bool = True, num_samples: int = 200) -> List[str]:
pass
@abstractmethod
def is_direct_completion(self) -> bool:
pass
def __repr__(self) -> str:
return self.name
def __str__(self) -> str:
return self.name
class VllmDecoder(DecoderBase):
def __init__(self, name: str, dataset: str, tp: int, **kwargs) -> None:
super().__init__(name, **kwargs)
kwargs = {
"tensor_parallel_size": int(os.getenv("VLLM_N_GPUS", tp)),
"dtype": self.dtype,
"trust_remote_code": self.trust_remote_code,
"gpu_memory_utilization": 0.95,
"enforce_eager": True,
"distributed_executor_backend": "ray",
}
self.tokenizer = AutoTokenizer.from_pretrained(self.name)
if not self.chat_mode:
self.eos += extra_eos_for_direct_completion(dataset)
self.llm = LLM(model=name, max_model_len=2048, **kwargs)
def is_direct_completion(self) -> bool:
return not self.chat_mode
def codegen(self, prompts: List[str], do_sample: bool = True, num_samples: int = 200, use_tqdm=True) -> List[str]:
if do_sample:
assert self.temperature > 0, "Temperature must be greater than 0!"
vllm_outputs = self.llm.generate(
prompts,
SamplingParams(
temperature=self.temperature,
max_tokens=self.max_new_tokens,
top_p=0.95 if do_sample else 1.0,
stop=self.eos,
),
use_tqdm=use_tqdm,
)
gen_strs = [x.outputs[0].text.replace("\t", " ") for x in vllm_outputs]
return gen_strs
class GeneralVllmDecoder(VllmDecoder):
def __init__(self, name: str, no_batching: bool, **kwargs) -> None:
super().__init__(name, **kwargs)
self.no_batching = no_batching
self.eos += ["\n```\n", "```"]
print(f"EOS strings: {self.eos}")
def codegen(self, prompts: List[str], do_sample: bool = True, num_samples: int = 200) -> List[str]:
if self.no_batching:
print(f"{self.no_batching = }, disable continous batching. This may be slow.")
generated = []
for prompt in tqdm(prompts):
chat_prompt = make_chat_prompt(
prompt,
instruction_prefix=self.instruction_prefix,
response_prefix=self.response_prefix,
tokenizer=self.tokenizer,
chat_mode=self.chat_mode,
)
generated.append(VllmDecoder.codegen(self, chat_prompt, do_sample, num_samples, use_tqdm=False)[0])
else:
print(f"{self.no_batching = }, enable continous batching.")
chat_prompts = [
make_chat_prompt(
prompt,
instruction_prefix=self.instruction_prefix,
response_prefix=self.response_prefix,
tokenizer=self.tokenizer,
chat_mode=self.chat_mode,
)
for prompt in prompts
]
generated = VllmDecoder.codegen(self, chat_prompts, do_sample, num_samples)
return generated
class HfTorchDecoder(DecoderBase):
def __init__(self, name: str, dataset: str, **kwargs):
super().__init__(name=name, **kwargs)
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
kwargs = {}
kwargs["device_map"] = "auto"
kwargs["trust_remote_code"] = self.trust_remote_code
# string to torch dtype
kwargs["torch_dtype"] = getattr(torch, self.dtype)
self.skip_special_tokens = True
print(f"{kwargs = }")
self.tokenizer = AutoTokenizer.from_pretrained(name)
if not self.chat_mode:
self.eos += extra_eos_for_direct_completion(dataset)
self.model = AutoModelForCausalLM.from_pretrained(name, **kwargs)
self.model = self.model.to(self.device)
def is_direct_completion(self) -> bool:
return not self.chat_mode
@torch.inference_mode()
def codegen(self, prompt: str, do_sample: bool = True, num_samples: int = 200) -> List[str]:
if self.temperature == 0:
assert not do_sample
assert num_samples == 1
input_tokens = self.tokenizer.encode(prompt, return_tensors="pt").to(self.device)
kwargs = {}
if do_sample:
kwargs["top_p"] = 0.95
kwargs["temperature"] = self.temperature
stop_sequencer = StopSequencer(
self.model,
model_type="causal", # or seq2seq
tokenizer=self.tokenizer,
)
model = stop_sequencer.register_stop_texts(
stop_texts=self.eos,
input_length=input_tokens.size(-1),
)
outputs = model.generate(
input_tokens,
max_new_tokens=self.max_new_tokens,
do_sample=do_sample,
num_return_sequences=min(self.batch_size, num_samples),
pad_token_id=self.tokenizer.eos_token_id,
**kwargs,
)
gen_strs = self.tokenizer.batch_decode(
outputs[:, input_tokens.size(-1) :],
skip_special_tokens=self.skip_special_tokens,
)
outputs = []
# removes eos tokens.
for output in gen_strs:
min_index = 10000
for eos in self.eos:
if eos in output:
min_index = min(min_index, output.index(eos))
outputs.append(output[:min_index].replace("\t", " "))
return outputs
class GenenralHfTorchDecoder(HfTorchDecoder):
def __init__(self, name: str, **kwargs):
super().__init__(name=name, **kwargs)
self.eos += ["\n```\n", "```"]
print(f"EOS strings: {self.eos}")
self.tokenizer = AutoTokenizer.from_pretrained(self.name)
def codegen(self, prompt: str, do_sample: bool = True, num_samples: int = 200) -> List[str]:
prompt = make_chat_prompt(
prompt,
self.instruction_prefix,
self.response_prefix,
self.tokenizer,
chat_mode=self.chat_mode,
)
return HfTorchDecoder.codegen(self, prompt, do_sample, num_samples)
def make_model(
model: str,
dataset: str,
backend: str = "vllm",
batch_size: int = 1,
temperature: float = 0.0,
tp=1,
instruction_prefix=None,
response_prefix=None,
no_batching=False, # no_batching -> disable continous batching in vLLM
chat_mode=False,
):
if backend == "vllm":
return GeneralVllmDecoder(
name=model,
batch_size=batch_size,
temperature=temperature,
dataset=dataset,
tp=tp,
instruction_prefix=instruction_prefix,
response_prefix=response_prefix,
no_batching=no_batching,
chat_mode=chat_mode,
)
elif backend == "hf":
return GenenralHfTorchDecoder(
name=model,
batch_size=batch_size,
temperature=temperature,
dataset=dataset,
instruction_prefix=instruction_prefix,
response_prefix=response_prefix,
chat_mode=chat_mode,
)