v0.9.0

Hexagon HTP

Microsoft-signed HTP catalog (llama.cpp @ e467e5). No cert import required on Windows on Snapdragon.

What's Changed

• feat(sdk): add geniex_set_log_level and derive the QNN log level from it by David Qian (@Davidqian123) in #1510
• feat(bench): add --chat-template and default --log to error by David Qian (@Davidqian123) in #1513
• chore(sdk): bump geniex-qairt to latest main by David Qian (@Davidqian123) in #1516
• fix(sdk): fix HTP fastrpc mapping failures on IQ9/IQ8 NPU by Perry Cheng (@zhycheng614) in #1503
• feat(server): add Jev-shaped local decision endpoint by RemiliaForever (@RemiliaForever) in #1497
• fix(android): propagate apply_chat_template failures as exception by Gerard Louis Andres Recinto (@gerardrecinto) in #1518
• fix(sdk): stop pinning GPT-OSS experts to CPU by RemiliaForever (@RemiliaForever) in #1523
• chore(sdk): bump geniex-qairt for QAIRT tool-call support by RemiliaForever (@RemiliaForever) in #1525
• docs(docs): document reasoning_format for chat completions by RemiliaForever (@RemiliaForever) in #1528
• feat(server): expose OpenAI Decisions API by RemiliaForever (@RemiliaForever) in #1526
• fix(sdk): warm up llama.cpp context at load so HTP mapping stays out of TTFT by Mengsheng Wu (@mengshengwu) in #1529
• fix(sdk): enable flash attention on GPU only for GPT-OSS by zdc (@Hahahaooovvv) in #1524
• fix(sdk): avoid Windows Hexagon buffer cap by RemiliaForever (@RemiliaForever) in #1530
• feat(sdk): support caller-provided input embeddings for LLM generation by Mengsheng Wu (@mengshengwu) in #1482
• fix(sdk): run geniex-bench timing runs to n_gen tokens, ignoring EOS by Mengsheng Wu (@mengshengwu) in #1537
• feat(sdk): upgrade QAIRT plugin to 2.50 by RemiliaForever (@RemiliaForever) in #1535
• fix(sdk): resume partial range downloads safely by RemiliaForever (@RemiliaForever) in #1540
• docs(cli): add Windows driver update guide by RemiliaForever (@RemiliaForever) in #1541
• feat(sdk): geniex-bench self-abort on divergence by Perry Cheng (@zhycheng614) in #1517
• docs(cli): document VLM NPU workaround with a dedicated vision session by Mengsheng Wu (@mengshengwu) in #1546
• fix(sdk): attach media to geniex-bench VLM accuracy prompt by Mengsheng Wu (@mengshengwu) in #1543
• test(qdc): keep the device pytest session alive after a native abort by Mengsheng Wu (@mengshengwu) in #1539
• chore(sdk): bump llama.cpp submodule to e467e5b5e by Perry Cheng (@zhycheng614) in #1544
• fix(server): separate reasoning by default by RemiliaForever (@RemiliaForever) in #1545
• fix(sdk): map Android SM8750/SM8850 to for-Galaxy chipset aliases by Mengsheng Wu (@mengshengwu) in #1549
• feat(cli): add JEV structured decisions by Khoi VN (@vnk8071) in #1496
• ci(sdk): switch QDC artifact upload/log download to v2 API by Mengsheng Wu (@mengshengwu) in #1551
• fix(sdk): make llama.cpp OpenCL work on Adreno 663 (IQ-9075) by Mengsheng Wu (@mengshengwu) in #1538
• feat(sdk)!: add structured logits decisions by RemiliaForever (@RemiliaForever) in #1552
• fix(sdk): default HTP1 session so mmproj runs on its own HTP session by Perry Cheng (@zhycheng614) in #1548

New Contributors

• Gerard Louis Andres Recinto (@gerardrecinto) made their first contribution in #1518
• Khoi VN (@vnk8071) made their first contribution in #1496

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