v0.7.0

Hexagon HTP

Microsoft-signed HTP catalog (llama.cpp @ 4ff829). No cert import required on Windows on Snapdragon.

What's Changed

• fix(sdk): use resolved device for VLM encoder by RemiliaForever (@RemiliaForever) in #1437
• fix(sdk): patch llama.cpp OpenCL add kernel for Adreno by RemiliaForever (@RemiliaForever) in #1438
• ci: add Breeze grading workflow by RemiliaForever (@RemiliaForever) in #1424
• feat(qairt): enable HTP multicore for QAIRT bundles by MAN$I VERMA 🦀 (@mansiverma897993) in #1313
• feat(sdk): load QAIRT system prompt from metadata by RemiliaForever (@RemiliaForever) in #1443
• feat(sdk): expose the QAIRT runtime path on the C API by Mengsheng Wu (@mengshengwu) in #1441
• feat(sdk): add VLM compute device override by RemiliaForever (@RemiliaForever) in #1446
• fix(server): rebuild full VLM context after a failed generation by RemiliaForever (@RemiliaForever) in #1442
• chore(sdk): bump geniex-qairt to merged gemma4 VLM pipeline fix by David Qian (@Davidqian123) in #1448
• fix(sdk): keep bench output when generation exhausts the context by Perry Cheng (@zhycheng614) in #1451
• fix(server): decode base64 audio input by MAN$I VERMA 🦀 (@mansiverma897993) in #1458
• fix(go): retry chunk download on any error, not just timeout/EOF by Mengsheng Wu (@mengshengwu) in #1459
• fix(sdk): apply context params to VLM encoder by RemiliaForever (@RemiliaForever) in #1460
• feat(qairt): native stop sequences in the geniex-qairt pipeline by MAN$I VERMA 🦀 (@mansiverma897993) in #1342
• fix(sdk): defer geniex-bench top-p to bundle/plugin default when unset by Paul Zhu (@vinovo) in #1461
• fix(cli): print download cancel hint before progress by Mengsheng Wu (@mengshengwu) in #1468
• fix(qairt): surface QNN backend logs through --log/GENIEX_LOG by Mengsheng Wu (@mengshengwu) in #1470
• feat(sdk): add Qwen3.5-4B to geniex-bench model catalog by Mengsheng Wu (@mengshengwu) in #1469
• feat: unify HTP power-mode control across qairt and llama_cpp by Mengsheng Wu (@mengshengwu) in #1447
• fix(sdk): name supported formats when a repo has no loadable model files by zdc (@Hahahaooovvv) in #1467

Full Changelog: v0.6.1...v0.7.0