The Anatomy of GPU Memory Budgeting: W + G + O + A + B — Managing the 24GB on RTX 4090 at the Atom Level

# === Bellek ölçüm helper'ı — cookbook'un her Lab'ında bulunur ===
import torch
 
def measure_memory(label: str = ""):
    """Mevcut GPU bellek breakdown'unu döner."""
    torch.cuda.synchronize()
    allocated = torch.cuda.memory_allocated() / 1024**3
    reserved = torch.cuda.memory_reserved() / 1024**3
    peak = torch.cuda.max_memory_allocated() / 1024**3
    free, total = torch.cuda.mem_get_info()
    free_gb = free / 1024**3
    total_gb = total / 1024**3
    print(
        f"[mem{':' + label if label else ''}] "
        f"alloc={allocated:6.2f} GB  reserved={reserved:6.2f} GB  "
        f"peak={peak:6.2f} GB  free={free_gb:6.2f}/{total_gb:.1f} GB"
    )
    return {
        "alloc_gb": allocated, "reserved_gb": reserved,
        "peak_gb": peak, "free_gb": free_gb,
    }
 
def estimate_budget(n_params: int, precision: str = "nf4", optimizer: str = "adamw_8bit",
                    batch: int = 1, seq_len: int = 4096, layers: int = 32, hidden: int = 4096,
                    grad_ckpt: bool = True, trainable_fraction: float = 0.007):
    """
    Cookbook bütçe formülü — kâğıt-kalem tahmin yapar.
    trainable_fraction: QLoRA r=32 için ~0.7%, full FT için 1.0.
    """
    bpp = {"fp32": 4, "fp16": 2, "bf16": 2, "fp8": 1, "int8": 1, "nf4": 0.5}[precision]
    opt_mult = {"sgd": 0, "sgd_momentum": 4, "adamw": 8, "adamw_bf16": 8, "adamw_8bit": 2, "lion": 4}[optimizer]
 
    W = n_params * bpp
    G = n_params * trainable_fraction * 2  # grads always bf16
    O = n_params * trainable_fraction * opt_mult
    mult = 10
    A_naive = batch * seq_len * layers * hidden * 2 * mult
    A = A_naive / (layers**0.5) * 2.5 if grad_ckpt else A_naive
    B = 3 * 1024**3
 
    total = W + G + O + A + B
    return {
        "W_gb": W / 1024**3,
        "G_gb": G / 1024**3,
        "O_gb": O / 1024**3,
        "A_gb": A / 1024**3,
        "B_gb": B / 1024**3,
        "total_gb": total / 1024**3,
    }
 
# Llama 3.1 8B QLoRA baseline
budget = estimate_budget(n_params=8_030_000_000)
print(budget)
# {'W_gb': 3.74, 'G_gb': 0.10, 'O_gb': 0.10, 'A_gb': 5.21, 'B_gb': 3.00, 'total_gb': 12.15}