File size: 1,912 Bytes
46b55ef |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 |
import random
import numpy as np
import torch
# -------------------------
# Replicability Logic
# -------------------------
def set_seed(seed=42):
"""
Sets universal random seeds to ensure deterministic results
across Python, NumPy, and PyTorch.
"""
# Python
random.seed(seed)
# NumPy
np.random.seed(seed)
# PyTorch
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
# Force deterministic algorithms to ensure GPU calculates the exact same gradients every time
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
print(f"Random Seed Set to: {seed}")
# -------------------------
# Hardware Diagnostics
# -------------------------
def get_device_capability():
"""
Diagnostics to ensure the GPU is ready for LLM Fine-Tuning.
Enables TF32 for newer NVIDIA architectures (L4).
"""
# Enable TF32 for modern Tensor Cores
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
if not torch.cuda.is_available():
raise RuntimeError("No GPU found! Go to Runtime > Change runtime type > Select NVIDIA L4.")
device = torch.device('cuda')
# Extract GPU Metadata
gpu_name = torch.cuda.get_device_name(0)
gpu_mem = torch.cuda.get_device_properties(0).total_memory / 1e9
capability = torch.cuda.get_device_capability(0)
bf16_support = torch.cuda.is_bf16_supported()
# Print Status Report
print(f"GPU Detected: {gpu_name}")
print(f" |-- Memory: {gpu_mem:.2f} GB")
print(f" |-- Compute Capability: {capability}")
print(f" |-- BFloat16 Support: {'Yes' if bf16_support else 'No'}")
# Professional Warning System
if "L4" not in gpu_name and "A100" not in gpu_name:
print(f"\nWarning: Using {gpu_name}. Performance may be suboptimal for Gemma fine-tuning.")
return device |