114. Mixed-Precision Adam Memory

In mixed-precision training (fp16/bf16 forward + bwd, fp32 master copy), compute the bytes per parameter for parameters + gradients + optimizer state.

Signature: def mixed_precision_memory_bytes(n_params: int) -> int

The convention (Megatron / DeepSpeed):

• fp16 weights used in forward/backward: 2 bytes
• fp16 gradients: 2 bytes
• fp32 master weights (kept by the optimizer for stable updates): 4 bytes
• fp32 Adam first moment m: 4 bytes
• fp32 Adam second moment v: 4 bytes

Total: 2 + 2 + 4 + 4 + 4 = 16 bytes per parameter.

Return n_params * 16.