Can 70B fit in 96 GiB?
Audit weight memory, KV cache, and runtime overhead before treating unified memory as a no-quantization guarantee.
Illustrative transformer assumptions
weights = parameters × 10^9 × bytes/parameter ÷ 2^30
KV = 2 × layers × context × KV heads × head dim × KV bytes ÷ 2^30
total = (weights + KV) × (1 + overhead)
KV = 2 × layers × context × KV heads × head dim × KV bytes ÷ 2^30
total = (weights + KV) × (1 + overhead)