5 Commits

Author SHA1 Message Date
MNNSyncBot d79be7f94d [QNN:Bugfix] Fix vision model errors and abnormal quant params; add Qwen3.5 support
GitOrigin-RevId: 36c983b957bf09de44400548710233355be286c8
2026-05-08 13:35:23 +08:00
YinHanke 5889d95ab6 [LLM:Bugfix] Fix Qwen3-Embedding QNN export pipeline (#4216) 2026-03-10 17:34:50 +08:00
jxt1234 080bb3182b [QNN:Feature] Support QNN Attention for LLM with max_history_token and fix QNN op bugs (#4174) 2026-03-09 14:07:37 +08:00
zhaode.wzd f31132f9d7 [Sync] Update MNN to 3.3.1 and Sync Internal Improvements
- Version: Upgrade to 3.3.1.
- LLM: Refactor with Transformer-like config/model/tokenizer; optimize Tokenizer load speed; enhance KV Cache (mmap, disk storage, prefix cache); fix LoRA and VL model bugs.
- CPU/SME: Support SME/SME2 instructions and mixed NEON/SME multi-threading; optimize RVV intrinsics and FP32 depth-wise kernels.
- Metal/OpenCL: Support Metal KV Cache mmap and GPU family identification; refactor OpenCL Loop operators and fix offset bugs.
- QNN/Android: Optimize QNN build/conversion; update Reranker API; fix Android packaging and SO loading issues.
- General: Fix compilation for Arm32/Win/MinGW; update gitignore and expose public symbols; revert StrideSliceWrite behavior.
2025-12-19 09:45:24 +08:00
xiaying 1ca62a02f7 MNN:Sync: Sync Internal 3.3.0 2025-10-31 11:29:14 +08:00