Delete GPT-2-Small-1k-opt-v2.py

Testing/GPT-2-Small-1k-opt-v2.py

-import time
-import torch
-import torch.nn as nn
-from transformers import GPT2LMHeadModel, GPT2Tokenizer, DataCollatorForLanguageModeling
-from datasets import load_dataset
-from torch.utils.data import DataLoader
-
-# --- Einstellungen ---
-BATCH_SIZE = 32
-SEQ_LEN = 2048
-NUM_STEPS = 1000
-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
-
-# Datensatz laden
-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()
-    return tokens
-
-tokenized_datasets = dataset.map(tokenize_function, batched=True, num_proc=4, desc="Tokenisierung läuft...")
-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)
-
-# Korrekte FLOP-Berechnung (basierend auf Paper-Formel)
-def compute_flops(batch_size, seq_len, num_layers, hidden_size, num_params):
-    flops_weight_input = 6 * seq_len * num_params
-    flops_attention = 12 * num_layers * hidden_size * seq_len ** 2
-    return (flops_weight_input + flops_attention) * batch_size
-
-# Debugging aktivieren für CUDA-Fehler
-torch.backends.cudnn.deterministic = True
-torch.backends.cudnn.benchmark = False
-
-# Benchmark-Funktion mit genauer Zeitmessung
-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_total = time.time()
-
-    total_forward_time = 0
-    total_backward_time = 0
-
-    for step, batch in enumerate(dataloader):
-        if step >= num_steps:
-            break
-        try:
-            batch = {k: v.to(DEVICE, non_blocking=True) for k, v in batch.items()}
-
-            torch.cuda.synchronize()
-            start_fwd = time.time()
-            loss = model(**batch).loss
-            torch.cuda.synchronize()
-            total_forward_time += time.time() - start_fwd
-
-            torch.cuda.synchronize()
-            start_bwd = time.time()
-            loss.backward()
-            torch.cuda.synchronize()
-            total_backward_time += time.time() - start_bwd
-
-            optimizer.step()
-            optimizer.zero_grad()
-        except RuntimeError as e:
-            print(f"RuntimeError detected: {e}")
-            continue
-
-    total_time = time.time() - start_time_total
-    tokens_per_second = (num_steps * BATCH_SIZE * SEQ_LEN) / total_time
-    flops_per_step = compute_flops(BATCH_SIZE, SEQ_LEN, NUM_LAYERS, HIDDEN_SIZE, model.num_parameters())
-    tflops_per_sec = (flops_per_step * (tokens_per_second / (BATCH_SIZE * SEQ_LEN))) / 1e12
-
-    return tokens_per_second, tflops_per_sec, total_forward_time / num_steps, total_backward_time / num_steps
-
-# FlashAttention-2 korrekt integrieren
-try:
-    from flash_attn.flash_attention import FlashAttention2
-    def replace_attention_with_flash(model):
-        for module in model.modules():
-            if isinstance(module, torch.nn.MultiheadAttention):
-                module.forward = FlashAttention2.apply
-except ImportError:
-    print("FlashAttention-2 nicht installiert. Standard Attention wird verwendet.")
-
-def load_model(attn_type):
-    model = GPT2LMHeadModel.from_pretrained("gpt2", torch_dtype=MIXED_PRECISION, device_map="auto")
-    if attn_type == "flash2":
-        replace_attention_with_flash(model)
-    return model
-
-results = {}
-
-for attn_type in ["standard", "flash2"]:
-    print(f"Teste GPT-2 Small mit {attn_type} Attention...")
-    model = load_model(attn_type)
-    tokens_per_sec, tflops_per_sec, avg_fwd, avg_bwd = benchmark_training(model, dataloader)
-    results[attn_type] = (tokens_per_sec, tflops_per_sec, avg_fwd, avg_bwd)
-    print(f"{attn_type} Attention: {tokens_per_sec:.2f} Tokens/Sek, {tflops_per_sec:.2f} TFLOPS/s")
-    print(f"   Durchschnittliche Forward-Zeit: {avg_fwd:.4f}s, Backward-Zeit: {avg_bwd:.4f}s")
-
-# Ergebnisse ausgeben
-print("--- GPT-2-Small-1k-opt-v2 ---")
-print(f"BATCH_SIZE: {BATCH_SIZE}, SEQ_LEN: {SEQ_LEN}, NUM_STEPS: {NUM_STEPS}")
-print("Endergebnisse:")
-for attn_type, (speed, tflops, avg_fwd, avg_bwd) in results.items():
-    print(f"{attn_type.capitalize()} Attention: {speed:.2f} Tokens/Sek, {tflops:.2f} TFLOPS/s")
-    print(f"   Durchschnittliche Forward-Zeit: {avg_fwd:.4f}s, Backward-Zeit: {avg_bwd:.4f}s")