AMOR-GatedDeltaNet-180M / export_metadata.json
FlyinGodzilla's picture
Provide verified sharded weights and optional inference routers
9e348a1 verified
Raw History Blame Contribute Delete
4 kB
{
"repo_id": "FlyinGodzilla/AMOR-GatedDeltaNet-180M",
"source_filename": "amor_gdn_180m_step_0007503.pt",
"source_sha256": "1e4f788e30601be79ec790a9c896f2a9f455e0fd6fa17d97c646f49ae4eb4f6c",
"step": 7503,
"tokens_seen": 3687948288,
"n_params": 181967832,
"tensor_dtypes": {
"torch.float32": 242
},
"gate_buffers": {
"0": {
"running_threshold": 0.7796082496643066,
"running_std": 0.08262228220701218
},
"1": {
"running_threshold": 0.6976275444030762,
"running_std": 0.12285983562469482
},
"2": {
"running_threshold": 0.7400334477424622,
"running_std": 0.10102786123752594
}
},
"optimizer_included": false,
"precision_changed": false,
"validation": {
"exported_tensors_equal": true,
"strict_model_load": true,
"generation": true,
"cuda": {
"versions": {
"torch": "2.6.0+cu124",
"transformers": "5.0.0",
"mamba-ssm": "2.3.0",
"causal-conv1d": "1.6.0",
"flash-linear-attention": "0.4.2",
"triton": "3.2.0",
"safetensors": "0.7.0"
},
"tokenizer_revision": "5590cc63198b0f111f4289daf006da0a12044173",
"strict_model_load": true,
"generation": true,
"device": "NVIDIA H100 NVL",
"torch": "2.6.0+cu124",
"autocast": "bfloat16",
"prefill_max_abs_diff": 0.0,
"decode_max_abs_diff": 0.0625,
"generated_token_ids": [
264,
1949,
11,
1825,
2592,
11,
1825,
2592
],
"published_cli_passed": true,
"test_scope": "Short CUDA smoke test; not a benchmark rerun",
"comparison_tolerance": {
"rtol": 0.02,
"atol": 0.1
}
},
"router_release": {
"exact_router_export": true,
"base_model_association": true,
"default_forward_max_abs_diff": 0.0,
"modes": [
{
"router": false,
"cases": [
{
"batch": 1,
"length": 16,
"prefill_max_abs_diff": 0.0,
"decode_max_abs_diff": 0.0625
},
{
"batch": 1,
"length": 128,
"prefill_max_abs_diff": 0.0,
"decode_max_abs_diff": 0.0625
},
{
"batch": 2,
"length": 64,
"prefill_max_abs_diff": 0.0625
}
],
"canonical_router_exact": null
},
{
"router": true,
"cases": [
{
"batch": 1,
"length": 16,
"prefill_max_abs_diff": 0.0,
"decode_max_abs_diff": 0.0625
},
{
"batch": 1,
"length": 128,
"prefill_max_abs_diff": 0.0,
"decode_max_abs_diff": 0.0625
},
{
"batch": 2,
"length": 64,
"prefill_max_abs_diff": 0.0
}
],
"canonical_router_exact": true
}
],
"frozen_thresholds_preserved": true,
"training_router_rejected": true,
"gpu": "NVIDIA H100 NVL",
"torch": "2.6.0+cu124",
"generation_cli_modes_passed": [
"entropy",
"router"
]
},
"current_release": {
"strict_model_load": true,
"entropy_and_router_generation_passed": true,
"unit_tests_passed": 17,
"device": "NVIDIA H100 NVL"
}
},
"implementation": "AMOR shared release: native entropy gate and optional fitted inference router",
"router": {
"variant": "distill",
"hidden_dim": 512,
"activation": "silu",
"weights": "router.safetensors",
"config": "router_config.json",
"sha256": "1f5080416aab1f4fda1dd8020a101c671c7812165f81082276e644f5e364d1c1",
"base_weight_sha256": "d50fec918303eb86175795fd4efc7d4d9b27cb11ed77dc7c80d901bd75d57f2f",
"precision": "float32",
"inference_only": true
}
}