v0.5.0

Hexagon HTP

Microsoft-signed HTP catalog (llama.cpp @ 873e5d). No cert import required on Windows on Snapdragon.

What's Changed

• ci(bench): add gemma-4 GPU cells and a --compute filter to QDC bench by Mengsheng Wu (@mengshengwu) in #1335
• fix(server): report speculative-decoding stats in OpenAI usage by Mengsheng Wu (@mengshengwu) in #1336
• ci(bench): add gemma-4-26B-A4B-it QAT GGUF for QDC bench by Mengsheng Wu (@mengshengwu) in #1347
• fix(bench): run VLM llama_cpp cells as text LLM in ctx/pp/tg sweep by Mengsheng Wu (@mengshengwu) in #1343
• fix(sdk): default llama.cpp swa_full to false by Mengsheng Wu (@mengshengwu) in #1346
• feat(server): parse Gemma 4 tool-call format by RemiliaForever (@RemiliaForever) in #1345
• feat(sdk): add MTP speculative-decoding coverage to bench matrix by Mengsheng Wu (@mengshengwu) in #1348
• feat(sdk): return PROMPT_TOO_LONG when the prompt itself overflows by RemiliaForever (@RemiliaForever) in #1349
• feat(sdk): measure media encoder time separately from prefill by RemiliaForever (@RemiliaForever) in #1354
• fix(cli): drop unsupported modalities before VLM generate by RemiliaForever (@RemiliaForever) in #1356
• fix(cli): run update installer silently by MAN$I VERMA (@mansiverma897993) in #1352
• test(sdk): add CPU + qairt unit tests for chat template, KV shift, mtmd by Mengsheng Wu (@mengshengwu) in #1355
• fix(bench): raise QDC job timeout and drop hybrid from the matrix by Mengsheng Wu (@mengshengwu) in #1358
• test(sdk): add cross-backend logits parity check by Mengsheng Wu (@mengshengwu) in #1357
• test(qdc): skip test_mtp_multi_turn on QCS9075M by Mengsheng Wu (@mengshengwu) in #1359
• feat(sdk): inject tools into the QAIRT VLM chat template by RemiliaForever (@RemiliaForever) in #1360
• feat(sdk): accept an explicit device list as a compute unit by Mengsheng Wu (@mengshengwu) in #1361
• fix(sdk): reuse VLM KV via char-level prefix match by RemiliaForever (@RemiliaForever) in #1363
• test(python): disable thinking in llm quality keyword cells by RemiliaForever (@RemiliaForever) in #1365
• fix(resume): re-fetch file when size mismatches legacy no-marker state by Mengsheng Wu (@mengshengwu) in #1364
• feat(sdk): add a CPU-only linux build variant for baseline armv8.0 boards by David Qian (@Davidqian123) in #1340
• feat(sdk): resolve AIHM version from latest.txt instead of pinning by zdc (@Hahahaooovvv) in #1353
• test(sdk): stop leaking tempdirs in model-manager store tests by RemiliaForever (@RemiliaForever) in #1368
• feat(android): add a CPU-only AAR for baseline armv8.0 devices by RemiliaForever (@RemiliaForever) in #1366
• docs(docs): document multi-HTP device list in MTP tutorial and CLI reference by Mengsheng Wu (@mengshengwu) in #1370
• feat(sdk): fetch AIHM index from releases/latest/, pinned as fallback by zdc (@Hahahaooovvv) in #1372
• chore: bump up llama.cpp by Paul Zhu (@vinovo) in #1374

Full Changelog: v0.4.0...v0.5.0