bitgpu 0.16.0

f16 activations for the Qwen3.5 hybrid decode path and the Bonsai-27B as a first-class committed model.

Highlights

• f16 activation-compute (activation: 'f16') now covers the hybrid (Bonsai-27B) decode path, not just the dense models. Projection GEMVs read f16 norm outputs while recurrence/attention/residual stay f32 — opt-in, greedy-exact, auto-falls-back to f32 where shader-f16 is missing. A real decode win on packed-f16 GPUs (NVIDIA/AMD), ~neutral on Apple.
• Bonsai-27B is now committed like the dense models — models/bonsai-27b-gguf/{manifest.json, aux.bin} (851 tensors, 64 layers). No more in-browser fromGguf shortcut in the demo: it loads through the same committed-manifest path as 1.7B/4B/8B, weights stream from the Hub.
• convert-gguf.py now handles the Qwen3.5 hybrid arch (qwen35), mirroring the in-engine buildQwen35Manifest: the 3:1 linear/full split, the gated-DeltaNet tensor names, and the arch.hybrid section. One converter for all four models; its output is deep-equal to fromGguf's in-browser parse (aux byte-identical).
• Demo defaults activation: 'f16' (safe everywhere) and adds the Bonsai-27B option.
• Docs: README covers activation: 'f16' guidance and the hybrid's long-context durability edge; models/README lists the 27B.

Verified

• Hybrid gate under f16-act: greedy-exact (cos 1.0 vs golden).
• Real Bonsai-27B forward under f16-act: argmax 25358.
• convert-gguf.py qwen35 output: deep-equal to fromGguf, aux byte-identical.
• Dense 8B: PACKAGE OK, bit-exact, greedy 13.12 tok/s.

Full parity: the dense 1.7B/4B/8B paths are byte-identical to 0.15.0.

bitgpu 0.15.0

Full Changelog: v0.14.0...v0.15.0

bitgpu 0.14.0

Full Changelog: v0.13.1...v0.14.0

bitgpu 0.13.1

Full Changelog: v0.13.0...v0.13.1

bitgpu 0.13.0

Full Changelog: v0.12.0...v0.13.0

bitgpu 0.12.0

Full Changelog: v0.11.0...v0.12.0

bitgpu 0.11.0

Full Changelog: v0.10.0...v0.11.0

bitgpu 0.10.0

Full Changelog: v0.9.0...v0.10.0

bitgpu 0.9.0

Full Changelog: v0.8.0...v0.9.0

bitgpu 0.8.0

Full Changelog: v0.7.0...v0.8.0