| { | |
| "parameters": { | |
| "total": 248906240, | |
| "total_human": "248.91M", | |
| "active_per_token": 248906240, | |
| "active_human": "248.91M" | |
| }, | |
| "weights": { | |
| "precision": "bfloat16", | |
| "bytes": 497812480, | |
| "human": "497.81 MB" | |
| }, | |
| "kv_cache": { | |
| "bytes_per_token": 12800, | |
| "per_context_length_batch1": { | |
| "2048": { | |
| "bytes": 26214400, | |
| "human": "26.21 MB" | |
| }, | |
| "4096": { | |
| "bytes": 52428800, | |
| "human": "52.43 MB" | |
| } | |
| }, | |
| "formula": "2 * layers * kv_heads * head_dim * 2B" | |
| }, | |
| "activations_estimate": { | |
| "bytes_per_token": 307200, | |
| "note": "rough planning number (inference, bf16); actual depends on engine, flash-attn, batching" | |
| }, | |
| "training_projection": { | |
| "optimizer": "AdamW, mixed precision", | |
| "bytes_per_param_before_activations": 16, | |
| "state_bytes": 3982499840, | |
| "state_human": "3.98 GB", | |
| "note": "bf16 weights+grads, fp32 master+moments; sharded across data-parallel ranks by ZeRO/FSDP in practice" | |
| }, | |
| "inference_min_footprint": { | |
| "bytes": 547593728, | |
| "human": "547.59 MB", | |
| "note": "weights + ~10% runtime overhead, before KV cache" | |
| } | |
| } |