tinystories-1bit-llm / results.json
euclidstellar's picture
Upload results.json with huggingface_hub
e1c06f1 verified
Raw
History Blame Contribute Delete
3.12 kB
{
"arms": {
"fp32": {
"val": 2.0553244292736053,
"ppl": 7.809371050272343,
"params": 11159360
},
"qat_ternary": {
"val": 2.1759746789932253,
"ppl": 8.81076860951355,
"params": 11159360
},
"ptq_ternary": {
"val": 5.022902528444926,
"ppl": 151.8514178229577,
"params": 11159360
},
"fp16_d128": {
"val": 2.360668977101644,
"ppl": 10.598038926854809,
"params": 2104448
},
"qat_ternary_embed": {
"val": 2.310712154706319,
"ppl": 10.081601759246057,
"params": 11159360,
"embed_mode": "ternary",
"qat_recovery_of_ptq_embed_loss": 0.760477491978586
}
},
"baselines": {
"uniform": 8.3178,
"unigram": 6.038,
"bigram_bound": 3.614
},
"ladder_fp32": {
"rung2": 5.3828,
"rung3": 3.9106,
"rung4": 3.715,
"rung5": 3.2948,
"rung6": 2.4304,
"rung7": 2.0651
},
"config": {
"vocab": 4096,
"d": 320,
"n_layer": 8,
"n_head": 8,
"mult": 4,
"block_size": 256,
"weight_mode": "ternary",
"act_bits": 8
},
"training": {
"tokens": 19996672,
"steps": 2441,
"batch": 32,
"ctx": 256,
"lr": 0.001,
"warmup": 100,
"seed": 1337,
"optimizer": "AdamW(0.9,0.95) wd=0.1 clip=1.0",
"schedule": "linear warmup, cosine to 10%"
},
"param_split": {
"ternary": 9830400,
"full_precision": 1328960
},
"packed_bytes": {
"ternary_body": 1947601,
"fp16_rest": 2657920,
"total": 4605521
},
"key_findings": {
"qat_gap_to_fp32_nats": 0.12065024971962002,
"qat_recovered_fraction_of_ptq_loss": 0.9593438670566713,
"equal_memory_win_vs_fp16_nats": 0.18469429810841875,
"ternarization_relative_error": 0.531,
"weight_zero_fraction": 0.314,
"qat_wallclock_multiplier": 1.1170731707317074
},
"note": "checkpoints store fp32 MASTER weights. ternary quantization happens in the forward pass. the 4.61 MB figure is the packed inference size.",
"pareto": {
"all_fp16": {
"bytes": 22318720,
"MB": 22.32,
"share_1p58bit": 0.0,
"share_fp16": 0.0016,
"compression_vs_fp16": 1.0,
"val": 2.0553244292736053
},
"ternary_body_fp16_embed": {
"bytes": 4605521,
"MB": 4.61,
"share_1p58bit": 0.4229,
"share_fp16": 0.0079,
"compression_vs_fp16": 4.85,
"val": 2.1759746789932253
},
"ternary_body_int8_embed": {
"bytes": 3294801,
"MB": 3.29,
"share_1p58bit": 0.5911,
"share_fp16": 0.0111,
"compression_vs_fp16": 6.77,
"val": 2.1795
},
"ternary_body_ternary_embed": {
"bytes": 2243762,
"MB": 2.24,
"share_1p58bit": 0.9837,
"share_fp16": 0.0163,
"compression_vs_fp16": 9.95,
"val": 2.310712154706319
}
},
"param_buckets": {
"ternary_body": 9830400,
"embedding_tied": 1310720,
"norms_always_fp16": 18240
},
"correction": "an earlier table bucketed embedding+norms as 'full precision' then quantized the bucket; the true fp16 share of the ternary-embed model is 1.63%, not 11.9%"
}