| { |
| "scheme": "mixed-precision, drift-allocated (Quantize by Drift, Eq.1 balanced drift; exhaustive integer program under a hard byte budget)", |
| "quantizer": "group-wise asymmetric min/max, fp16 scale + fp16 zero per group", |
| "group": 16, |
| "choices_bits": [ |
| 2, |
| 3, |
| 4 |
| ], |
| "budget_real_bits": 5.0, |
| "allocation_units": { |
| "attention.b0": { |
| "bits": 3, |
| "params": 1179648 |
| }, |
| "attention.b1": { |
| "bits": 4, |
| "params": 589824 |
| }, |
| "attention.b2": { |
| "bits": 4, |
| "params": 1179648 |
| }, |
| "attention.b3": { |
| "bits": 4, |
| "params": 589824 |
| }, |
| "embedding.v0": { |
| "bits": 3, |
| "params": 2930688 |
| }, |
| "embedding.v1": { |
| "bits": 2, |
| "params": 2930304 |
| }, |
| "embedding.v2": { |
| "bits": 2, |
| "params": 2929920 |
| }, |
| "embedding.v3": { |
| "bits": 2, |
| "params": 2930304 |
| }, |
| "ffn.b0": { |
| "bits": 3, |
| "params": 2359296 |
| }, |
| "ffn.b1": { |
| "bits": 3, |
| "params": 1179648 |
| }, |
| "ffn.b2": { |
| "bits": 4, |
| "params": 2359296 |
| }, |
| "ffn.b3": { |
| "bits": 3, |
| "params": 1179648 |
| } |
| }, |
| "plan": { |
| "attention.b0": 3, |
| "attention.b1": 4, |
| "attention.b2": 4, |
| "attention.b3": 4, |
| "embedding.v0": 3, |
| "embedding.v1": 2, |
| "embedding.v2": 2, |
| "embedding.v3": 2, |
| "ffn.b0": 3, |
| "ffn.b1": 3, |
| "ffn.b2": 4, |
| "ffn.b3": 3 |
| }, |
| "unit_scheme": "4 layer bands x (attention, ffn) + 4 vocabulary-row bands of the word embedding", |
| "quantized_params": 22338048, |
| "fp16_params_in_payload": 344064, |
| "skipped_1d_params_not_in_payload": true, |
| "real_bits_per_quantized_param": 4.8177, |
| "predicted_mb": 14.14, |
| "packed_on_disk_mb": 14.14, |
| "error_pct": 0.0, |
| "ledger": { |
| "arms": "data/ledger/drift-mixed/arms-sr22m-g16-B5.0.json", |
| "arm": "A1b-drift", |
| "ndcg@10": 74.5589, |
| "real_bits": 4.9873 |
| }, |
| "tensors": [ |
| { |
| "name": "encoder.layer.0.attention.self.query.weight", |
| "unit": "attention.b0", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 0, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.0.attention.self.key.weight", |
| "unit": "attention.b0", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 92160, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.0.attention.self.value.weight", |
| "unit": "attention.b0", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 184320, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.0.attention.output.dense.weight", |
| "unit": "attention.b0", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 276480, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.1.attention.self.query.weight", |
| "unit": "attention.b0", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 368640, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.1.attention.self.key.weight", |
| "unit": "attention.b0", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 460800, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.1.attention.self.value.weight", |
| "unit": "attention.b0", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 552960, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.1.attention.output.dense.weight", |
| "unit": "attention.b0", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 645120, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.2.attention.self.query.weight", |
| "unit": "attention.b1", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 737280, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.2.attention.self.key.weight", |
| "unit": "attention.b1", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 847872, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.2.attention.self.value.weight", |
| "unit": "attention.b1", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 958464, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.2.attention.output.dense.weight", |
| "unit": "attention.b1", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 1069056, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.3.attention.self.query.weight", |
| "unit": "attention.b2", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 1179648, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.3.attention.self.key.weight", |
| "unit": "attention.b2", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 1290240, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.3.attention.self.value.weight", |
| "unit": "attention.b2", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 1400832, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.3.attention.output.dense.weight", |
| "unit": "attention.b2", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 1511424, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.4.attention.self.query.weight", |
| "unit": "attention.b2", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 1622016, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.4.attention.self.key.weight", |
| "unit": "attention.b2", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 1732608, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.4.attention.self.value.weight", |
| "unit": "attention.b2", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 1843200, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.4.attention.output.dense.weight", |
| "unit": "attention.b2", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 1953792, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.5.attention.self.query.weight", |
| "unit": "attention.b3", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 2064384, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.5.attention.self.key.weight", |
| "unit": "attention.b3", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 2174976, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.5.attention.self.value.weight", |
| "unit": "attention.b3", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 2285568, |
| "n": 147456 |
| }, |
| { |
| "name": "encoder.layer.5.attention.output.dense.weight", |
| "unit": "attention.b3", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 2396160, |
| "n": 147456 |
| }, |
| { |
| "name": "embeddings.word_embeddings.weight", |
| "unit": "embedding.v0", |
| "bits": 3, |
| "rows": [ |
| 0, |
| 7630 |
| ], |
| "shape": [ |
| 30522, |
| 384 |
| ], |
| "offset": 2506752, |
| "n": 2929920 |
| }, |
| { |
| "name": "embeddings.token_type_embeddings.weight", |
| "unit": "embedding.v0", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 2, |
| 384 |
| ], |
| "offset": 4337952, |
| "n": 768 |
| }, |
| { |
| "name": "embeddings.word_embeddings.weight", |
| "unit": "embedding.v1", |
| "bits": 2, |
| "rows": [ |
| 7630, |
| 15261 |
| ], |
| "shape": [ |
| 30522, |
| 384 |
| ], |
| "offset": 4338432, |
| "n": 2930304 |
| }, |
| { |
| "name": "embeddings.word_embeddings.weight", |
| "unit": "embedding.v2", |
| "bits": 2, |
| "rows": [ |
| 15261, |
| 22891 |
| ], |
| "shape": [ |
| 30522, |
| 384 |
| ], |
| "offset": 5803584, |
| "n": 2929920 |
| }, |
| { |
| "name": "embeddings.word_embeddings.weight", |
| "unit": "embedding.v3", |
| "bits": 2, |
| "rows": [ |
| 22891, |
| 30522 |
| ], |
| "shape": [ |
| 30522, |
| 384 |
| ], |
| "offset": 7268544, |
| "n": 2930304 |
| }, |
| { |
| "name": "encoder.layer.0.intermediate.dense.weight", |
| "unit": "ffn.b0", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 1536, |
| 384 |
| ], |
| "offset": 8733696, |
| "n": 589824 |
| }, |
| { |
| "name": "encoder.layer.0.output.dense.weight", |
| "unit": "ffn.b0", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 384, |
| 1536 |
| ], |
| "offset": 9102336, |
| "n": 589824 |
| }, |
| { |
| "name": "encoder.layer.1.intermediate.dense.weight", |
| "unit": "ffn.b0", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 1536, |
| 384 |
| ], |
| "offset": 9470976, |
| "n": 589824 |
| }, |
| { |
| "name": "encoder.layer.1.output.dense.weight", |
| "unit": "ffn.b0", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 384, |
| 1536 |
| ], |
| "offset": 9839616, |
| "n": 589824 |
| }, |
| { |
| "name": "encoder.layer.2.intermediate.dense.weight", |
| "unit": "ffn.b1", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 1536, |
| 384 |
| ], |
| "offset": 10208256, |
| "n": 589824 |
| }, |
| { |
| "name": "encoder.layer.2.output.dense.weight", |
| "unit": "ffn.b1", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 384, |
| 1536 |
| ], |
| "offset": 10576896, |
| "n": 589824 |
| }, |
| { |
| "name": "encoder.layer.3.intermediate.dense.weight", |
| "unit": "ffn.b2", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 1536, |
| 384 |
| ], |
| "offset": 10945536, |
| "n": 589824 |
| }, |
| { |
| "name": "encoder.layer.3.output.dense.weight", |
| "unit": "ffn.b2", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 1536 |
| ], |
| "offset": 11387904, |
| "n": 589824 |
| }, |
| { |
| "name": "encoder.layer.4.intermediate.dense.weight", |
| "unit": "ffn.b2", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 1536, |
| 384 |
| ], |
| "offset": 11830272, |
| "n": 589824 |
| }, |
| { |
| "name": "encoder.layer.4.output.dense.weight", |
| "unit": "ffn.b2", |
| "bits": 4, |
| "rows": null, |
| "shape": [ |
| 384, |
| 1536 |
| ], |
| "offset": 12272640, |
| "n": 589824 |
| }, |
| { |
| "name": "encoder.layer.5.intermediate.dense.weight", |
| "unit": "ffn.b3", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 1536, |
| 384 |
| ], |
| "offset": 12715008, |
| "n": 589824 |
| }, |
| { |
| "name": "encoder.layer.5.output.dense.weight", |
| "unit": "ffn.b3", |
| "bits": 3, |
| "rows": null, |
| "shape": [ |
| 384, |
| 1536 |
| ], |
| "offset": 13083648, |
| "n": 589824 |
| }, |
| { |
| "name": "embeddings.position_embeddings.weight", |
| "unit": null, |
| "bits": 16, |
| "rows": null, |
| "shape": [ |
| 512, |
| 384 |
| ], |
| "offset": 13452288, |
| "n": 196608 |
| }, |
| { |
| "name": "pooler.dense.weight", |
| "unit": null, |
| "bits": 16, |
| "rows": null, |
| "shape": [ |
| 384, |
| 384 |
| ], |
| "offset": 13845504, |
| "n": 147456 |
| } |
| ], |
| "reeval": { |
| "n_queries": 4392, |
| "ndcg@10": 74.5685, |
| "ndcg@10_se": 0.4594, |
| "recall@10": 80.8743 |
| } |
| } |
