Delete GPT-2-Small-1k-opt.py

Testing/GPT-2-Small-1k-opt.py

-import time
-import torch
-import torch.nn as nn
-from transformers import GPT2LMHeadModel, GPT2Tokenizer, DataCollatorForLanguageModeling, TrainingArguments, Trainer
-from datasets import load_dataset
-from torch.utils.data import DataLoader
-
-# --- Einstellungen ---
-BATCH_SIZE = 16
-SEQ_LEN = 1024
-NUM_STEPS = 100
-DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
-MIXED_PRECISION = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
-
-# GPT-2 Small Model Parameter
-NUM_LAYERS = 12
-HIDDEN_SIZE = 768
-NUM_HEADS = 12
-HEAD_DIM = HIDDEN_SIZE // NUM_HEADS
-
-dataset = load_dataset("wikitext", "wikitext-103-v1", streaming=False)
-tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
-tokenizer.pad_token = tokenizer.eos_token
-
-def tokenize_function(examples):
-    tokens = tokenizer(examples["text"], padding="max_length", truncation=True, max_length=SEQ_LEN)
-    tokens["labels"] = tokens["input_ids"].copy()  # Labels = Shifted input_ids
-    return tokens
-
-tokenized_datasets = dataset.map(tokenize_function, batched=True)
-tokenized_datasets.set_format(type="torch", columns=["input_ids", "attention_mask", "labels"])
-
-data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
-dataloader = DataLoader(tokenized_datasets["train"], batch_size=BATCH_SIZE, collate_fn=data_collator)
-
-def compute_flops(batch_size, seq_len, num_layers, hidden_size):
-    return 6 * num_layers * (hidden_size ** 2) * seq_len * batch_size
-
-def benchmark_training(model, dataloader, num_steps=NUM_STEPS):
-    model.to(DEVICE)
-    model.train()
-    optimizer = torch.optim.AdamW(model.parameters(), lr=5e-4)
-    
-    torch.cuda.synchronize()
-    start_time_fwd = time.time()
-
-    for step, batch in enumerate(dataloader):
-        if step >= num_steps:
-            break
-        batch = {k: v.to(DEVICE) for k, v in batch.items()}
-        torch.cuda.synchronize()
-        forward_time = time.time() - start_time_fwd
-        loss = model(**batch).loss
-        torch.cuda.synchronize()
-        start_time_bwd = time.time()
-        loss.backward()
-        torch.cuda.synchronize()
-        backward_time = time.time() - start_time_bwd
-        optimizer.step()
-        optimizer.zero_grad()
-
-    total_time = forward_time + backward_time
-    tokens_per_second = (num_steps * BATCH_SIZE * SEQ_LEN) / total_time
-    flops_per_step = compute_flops(BATCH_SIZE, SEQ_LEN, NUM_LAYERS, HIDDEN_SIZE)
-    tflops_per_sec = (flops_per_step * (tokens_per_second / (BATCH_SIZE * SEQ_LEN))) / 1e12
-    return tokens_per_second, tflops_per_sec
-
-results = {}
-
-for attn_type in ["standard", "flash2"]:
-    print(f"Teste GPT-2 Small mit {attn_type} Attention...")
-    if attn_type == "standard":
-        model = GPT2LMHeadModel.from_pretrained("gpt2", torch_dtype=MIXED_PRECISION, device_map="auto")
-    elif attn_type == "flash2":
-        model = GPT2LMHeadModel.from_pretrained("gpt2", attn_implementation="flash_attention_2", torch_dtype=MIXED_PRECISION, device_map="auto")
-    
-    tokens_per_sec, tflops_per_sec = benchmark_training(model, dataloader)
-    results[attn_type] = (tokens_per_sec, tflops_per_sec)
-    print(f"{attn_type} Attention: {tokens_per_sec:.2f} Tokens pro Sekunde, {tflops_per_sec:.2f} TFLOPS/s")
-
-print("--- GPT-2-Small-1k-opt ---")
-print(f"BATCH_SIZE: {BATCH_SIZE}, SEQ_LEN: {SEQ_LEN}, NUM_STEPS: {NUM_STEPS}")
-print("Endergebnisse:")
-for attn_type, (speed, tflops) in results.items():
-    print(f"{attn_type.capitalize()} Attention: {speed:.2f} Tokens pro Sekunde, {tflops:.2f} TFLOPS/s")