Files
exp-model-parameter-similarity/results/cpu-qwen35-2b/gain-factorization-summary.json

47 lines
1.4 KiB
JSON

{
"state": "completed",
"records": [],
"interventions": [
{
"layer": 3,
"direction": "q_to_k",
"sequence_nll_deltas": [
3.114725206643243e-07,
-5.23134924712565e-07,
1.7465536439331686e-07,
1.733758607791458e-07
],
"mean_nll_delta": 3.40922052810555e-08,
"sequence_p95_delta": 2.9094994722367315e-07,
"passed": true
},
{
"layer": 3,
"direction": "k_to_q",
"sequence_nll_deltas": [
4.0256782884995346e-07,
-8.45824208006718e-07,
2.1304516684139685e-07,
-3.860568877911419e-07
],
"mean_nll_delta": -1.540670250266274e-07,
"sequence_p95_delta": 3.741394295486699e-07,
"passed": true
}
],
"maximum_score_abs_error": 3.814697265625e-06,
"maximum_probability_abs_error": 8.046627044677734e-07,
"bidirectional_heldout_pass": true,
"elementwise_gain_multiplies_saved_per_token_layer": {
"q_to_k": 1536,
"k_to_q": 384
},
"original_bf16_payload_bytes": 1024,
"factored_lossless_mixed_payload_bytes": 1024,
"fp32_cached_gain_payload_before": 2048,
"fp32_cached_gain_payload_after": 1280,
"claim_limit": "Exact real-arithmetic identity only on non-RoPE coordinates; empirical FP32/NLL checks on tiny text pilot. Not evidence that two existing full layers interchange, not model-wide speedup.",
"elapsed_seconds": 232.08875080291182,
"peak_rss_kib": 83704
}