| """Zeus (#14662) Claim: architecture is ~100M params ("Zeus-100M"). |
| Verifiable from the released config + modeling code (no weights needed). |
| flash_attn is stubbed because it's only used in forward(), not for building layers.""" |
| import sys, types, json, torch |
| for n in ["flash_attn", "flash_attn.bert_padding"]: |
| sys.modules[n] = types.ModuleType(n) |
| sys.modules["flash_attn"].flash_attn_varlen_func = lambda *a, **k: None |
| sys.modules["flash_attn.bert_padding"].unpad_input = lambda *a, **k: None |
| sys.modules["flash_attn.bert_padding"].pad_input = lambda *a, **k: None |
| |
| from zeus.configuration_zeus import ZeusConfig |
| from zeus.modeling_zeus import ZeusForPrediction |
| c = ZeusConfig(**json.load(open("zeus/config.json"))) |
| m = ZeusForPrediction(c) |
| n = sum(p.numel() for p in m.parameters()) |
| print(f"Zeus params: {n:,} ({n/1e6:.1f}M)") |
|
|
