import os from tqdm import tqdm import numpy as np import tiktoken import json import gzip import torch import random torch.set_grad_enabled(False) device = "cuda" def load_exllama(model_dir): from exllamav2 import ExLlamaV2, ExLlamaV2Config, ExLlamaV2Cache, ExLlamaV2Tokenizer from exllamav2.generator import ExLlamaV2DynamicGenerator config = ExLlamaV2Config(model_dir) model = ExLlamaV2(config) model.load() tokenizer = ExLlamaV2Tokenizer(config) return model, tokenizer def load_nanogpt(model_dir, ckpt): import os import pickle from contextlib import nullcontext import torch import tiktoken from model import GPTConfig, GPT ckpt_path = os.path.join(model_dir, ckpt) checkpoint = torch.load(ckpt_path, map_location=device) gptconf = GPTConfig(**checkpoint['model_args']) model = GPT(gptconf) state_dict = checkpoint['model'] unwanted_prefix = '_orig_mod.' for k,v in list(state_dict.items()): if k.startswith(unwanted_prefix): state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k) model.load_state_dict(state_dict) model = model.to(device).eval() return model, tiktoken.get_encoding("gpt2") #model, tokenizer = load_exllama("./Llama-3-8B-Instruct-exl2") #model, tokenizer = load_nanogpt("./atk-fixed-suffix-2-0.0025", "ckpt3000.pt") model, tokenizer = load_nanogpt("./atk-fixed-suffix-2-0.00125", "ckpt3000.pt") def find_closest_tokens(model, tokenizer): weights_ = model.modules[0].embedding.weight.data weights = torch.zeros_like(weights_, device="cuda") weights.copy_(weights_) # some are zero, so we can't normalize easily #weights /= torch.linalg.norm(weights, dim=-1, keepdim=True) vocab_size, dim = weights.shape print("copied") best = torch.zeros(vocab_size, device="cuda", dtype=torch.int32) scores = torch.zeros(vocab_size, device="cuda", dtype=torch.float16) CHUNK_SIZE = 1024 for i in range(0, vocab_size, CHUNK_SIZE): print(i) similarities = (weights @ weights[i:i+CHUNK_SIZE, :].T) # zero similarity to self torch.diagonal(similarities, offset=i, dim1=1, dim2=0).fill_(-float("inf")) score, ix = torch.max(similarities, dim=0) best[i:i+CHUNK_SIZE] = ix scores[i:i+CHUNK_SIZE] = score scores, indices = torch.sort(scores, descending=True) print([ (indices[i].item(), best[indices][i].item(), tokenizer.decode(indices[i:i+1]), tokenizer.decode(best[indices][i:i+1])) for i in range(100) ]) #find_closest_tokens() #best_pair = 28217, 76665 # rare token pair in LLaMA #best_pair = 34966, 70467 # also that #best_pair = 48, 57 # Q, Z in LLaMA - we need to use common tokens or it cannot represent an even mix of them in the logits, but they can't be so common together that a compound token exists best_pair = 49704, 50009 # unused in our GPT-2 training dataset - used for data injection #best_pair = 2, 0 # seem to not form a compound token in GPT-2 tokenizer suffix = 49691 # chosen for data injection variant COUNT = 1000 total_max = 0 total_mean = 0 suffix_len = 512 count_len = 512 for _ in range(COUNT): sequence = torch.randint(low=0, high=2, size=(1024,), device="cuda", dtype=torch.int32) * (best_pair[1] - best_pair[0]) + best_pair[0] sequence[-suffix_len:] = torch.full((suffix_len,), suffix, device="cuda", dtype=torch.int32) sequence2 = sequence.clone() print("---") sequence[suffix_len-1] = best_pair[0] sequence2[suffix_len-1] = best_pair[1] logits = model.forward(torch.stack([sequence, sequence2], dim=0)) if isinstance(logits, tuple): logits = logits[0] #logits = logits.bfloat16() # introduce roundoff error deliberately print("Final logits", logits[:, -5:, :]) #print("Input", tokenizer.decode(sequence.tolist())) #print("Predictions", tokenizer.decode(torch.argmax(logits[0], dim=-1).tolist())) maxdiff = torch.max((logits[0, -count_len:] - logits[1, -count_len:]).flatten(), dim=-1).values.item() meandiff = torch.mean(((logits[0, -count_len:] - logits[1, -count_len:]).abs()).flatten(), dim=-1).item() total_max += maxdiff total_mean += abs(meandiff) print("Max diff", maxdiff) print("Mean diff", meandiff) print("---AVG---") print("Max diff", total_max / COUNT) print("Mean diff", total_mean / COUNT)