SlayerLab

/

NERGAL

Like 3

Follow

Fabryka AI 78

Same weights (model.safetensors unchanged, 1d42c344…), rules, threshold and default outputs; hybrid.json only bumps version.

• Nergal.predict_many(texts) / scrub_many(texts): windows batched across texts (≤ 32,768 padded tokens, ≤ 128 rows, groups of 64). predict / scrub are the one-text case.
• dtype="float16" (CUDA/MPS, opt-in; CLI --dtype): cast at load time.
• Encoding.count from cached unit pieces; identical windows.
• from_pretrained default device: CUDA → MPS → CPU.

RTX 4090 throughput (kchar/s): 1.0.3-style 11.0 → predict_many fp32 21.3 → fp16 39.4 (79.6 with 3 processes).

Gate (1,685 labelled dev rows vs cached published-weight spans): fp32 0 span changes at 0.95; fp16 +2 spans on gold, 0 removed, 841-dev union numbers identical. Details in CHANGELOG.md.

Branch: release/1.1.0 (fdc79fc).