completions.py

#%%
import time
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, PreTrainedModel, PreTrainedTokenizer, PreTrainedTokenizerFast, BatchEncoding
from transformers.generation.utils import GenerateOutput
from typing import cast
from dataclasses import dataclass

type Tokenizer = PreTrainedTokenizer | PreTrainedTokenizerFast

@dataclass
class Word:
    tokens: list[int]
    text: str
    logprob: float
    context: list[int]

def starts_with_space(token: str) -> bool:
    return token.startswith(chr(9601)) or token.startswith(chr(288))

def split_into_words(token_probs: list[tuple[int, float]], tokenizer: Tokenizer) -> list[Word]:
    words: list[Word] = []
    current_word: list[int] = []
    current_log_probs: list[float] = []
    current_word_first_token_index: int = 0
    all_tokens: list[int] = [token_id for token_id, _ in token_probs]

    def append_current_word():
        if current_word:
            words.append(Word(current_word,
                              tokenizer.decode(current_word),
                              sum(current_log_probs),
                              all_tokens[:current_word_first_token_index]))

    for i, (token_id, logprob) in enumerate(token_probs):
        token: str = tokenizer.convert_ids_to_tokens([token_id])[0]
        if not starts_with_space(token) and token.isalpha():
            current_word.append(token_id)
            current_log_probs.append(logprob)
        else:
            append_current_word()
            current_word = [token_id]
            current_log_probs = [logprob]
            current_word_first_token_index = i

    append_current_word()

    return words

def load_model_and_tokenizer(model_name: str, device: torch.device) -> tuple[PreTrainedModel, Tokenizer]:
    tokenizer: Tokenizer = AutoTokenizer.from_pretrained(model_name, padding_side="left")
    tokenizer.pad_token = tokenizer.eos_token
    model: PreTrainedModel = AutoModelForCausalLM.from_pretrained(model_name)
    model.to(device)
    return model, tokenizer

def tokenize(input_text: str, tokenizer: Tokenizer, device: torch.device) -> BatchEncoding:
    return tokenizer(input_text, return_tensors="pt").to(device)

def calculate_log_probabilities(model: PreTrainedModel, tokenizer: Tokenizer, inputs: BatchEncoding) -> list[tuple[int, float]]:
    input_ids = inputs["input_ids"]
    attention_mask = inputs["attention_mask"]
    with torch.no_grad():
        outputs = model(input_ids=input_ids, attention_mask=attention_mask, labels=input_ids)
    # B x T x V
    logits: torch.Tensor = outputs.logits[:, :-1, :]
    # B x T x V
    log_probs: torch.Tensor = torch.log_softmax(logits, dim=-1)
    # T - 1
    token_log_probs: torch.Tensor = log_probs[0, range(log_probs.shape[1]), input_ids[0][1:]]
    # T - 1
    tokens: torch.Tensor = input_ids[0][1:]
    return list(zip(tokens.tolist(), token_log_probs.tolist()))

def prepare_inputs(contexts: list[list[int]], tokenizer: Tokenizer, device: torch.device) -> BatchEncoding:
    texts = [tokenizer.decode(context, skip_special_tokens=True) for context in contexts]
    return tokenizer(texts, return_tensors="pt", padding=True).to(device)

def generate_outputs(model: PreTrainedModel, inputs: BatchEncoding, num_samples: int = 5) -> GenerateOutput | torch.LongTensor:
    input_ids = inputs["input_ids"]
    attention_mask = inputs["attention_mask"]
    with torch.no_grad():
        outputs = model.generate(
            input_ids=input_ids,
            attention_mask=attention_mask,
            max_new_tokens=4,
            num_return_sequences=num_samples,
            temperature=1.0,
            top_k=50,
            top_p=0.95,
            do_sample=True
            # num_beams=num_samples
        )
    return outputs

def extract_replacements(outputs: GenerateOutput | torch.LongTensor, tokenizer: Tokenizer, num_inputs: int, input_len: int, num_samples: int = 5) -> list[list[str]]:
    all_new_words = []
    for i in range(num_inputs):
        replacements = []
        for j in range(num_samples):
            generated_ids = outputs[i * num_samples + j][input_len:]
            new_word = tokenizer.convert_ids_to_tokens(generated_ids.tolist())[0]
            if starts_with_space(new_word):
                replacements.append(new_word[1:])
        all_new_words.append(replacements)
    return all_new_words

#%%

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# model_name = "mistralai/Mistral-7B-v0.1"
model_name = "unsloth/Llama-3.2-1B"
model, tokenizer = load_model_and_tokenizer(model_name, device)

#%%
input_text = "He asked me to prostrate myself before the king, but I rifused."
inputs: BatchEncoding = tokenize(input_text, tokenizer, device)

#%%
token_probs: list[tuple[int, float]] = calculate_log_probabilities(model, tokenizer, inputs)

#%%
words = split_into_words(token_probs, tokenizer)
log_prob_threshold = -5.0
low_prob_words = [word for word in words if word.logprob < log_prob_threshold]

#%%
contexts = [word.context for word in low_prob_words]
inputs = prepare_inputs(contexts, tokenizer, device)
input_ids = inputs["input_ids"]

#%%
num_samples = 5
start_time = time.time()
outputs = generate_outputs(model, inputs, num_samples)
end_time = time.time()
print(f"Total time taken for replacements: {end_time - start_time:.4f} seconds")

#%%
replacements_batch = extract_replacements(outputs, tokenizer, input_ids.shape[0], input_ids.shape[1], num_samples)

#%%
for word, replacements in zip(low_prob_words, replacements_batch):
    print(f"Original word: {word.text}, Log Probability: {word.logprob:.4f}")
    print(f"Proposed replacements: {replacements}")

# %%
generated_ids = outputs[:, input_ids.shape[-1]:]
for g in generated_ids:
    print(tokenizer.convert_ids_to_tokens(g.tolist()))

# %%