v4.2.0 — HF-Reference Fidelity + Reasoning API (7B chapter complete)

v4.2.0 — Reference Fidelity

The 7B chapter close-out. ATLAS output is now byte-identical to the HuggingFace transformers reference for OLMo-3-7B-Think.

Three silent inference bugs fixed (77fdc34)

Found by differential testing against HF transformers (token-level logit comparison):

1. YaRN correction range — missing 2π wavelength factor + wrong ramp; attn_factor must be applied squared
2. Layer-type RoPE split — YaRN only on the 8 full-attention layers; the 24 sliding-window layers use plain θ=500K
3. QK-norm scope — RMSNorm over the full 4096-dim Q/K vector, not per-head (the per-head variant destroyed retrieval >128 tokens)

Benchmark impact (live A100 endpoint)

Bench	Before	After
GSM8K (25)	12%	88% (Ai2 ref: 85–91)
MMLU (100, direct)	22%	54%
HumanEval (15, exec-verified)	—	73.3% pass@1

Serving upgrades (17c4c9c)

• Context 4K → 16K, max output 3,584
• Official OLMo chat template + <think> primer + Ai2 system prompt server-side
• Reasoning API: message.reasoning + streaming delta.reasoning (OpenRouter convention)
• Sampling defaults per Ai2 reference (rep_penalty 1.15 → 1.0), context-overflow guard
• Bearer-key auth on /v1/*; production endpoint behind Cloudflare at https://atlas.thebeastagi.com
• compute-sanitizer memcheck clean; exact CPU/GPU long-sequence parity test added

Also in this release window

• OpenRouter launch prep: /privacy policy page + provider description knob (3d1a502)
• GPU kernel-launch failure detection — no more silent garbage (bb8227a, 543edd1)
• AdamW step_gpu host-pointer no-op fixed with systemic CUDA error propagation (062674d, closes #19)

627 tests · CI green. Next: the 32B chapter (#26), batched prefill (#22), GEMV occupancy (#23), 32K context (#24).