ab3790faca
fix MHA singleton mask broadcast across backends ( #6869 )
...
Keep persistent KV caches on the blob allocator, broadcast single-channel attention masks across heads on CPU and Vulkan paths, and add long multi-token cache regressions.
2026-10-10 19:47:13 +08:00
41f031e045
reduce redundant tests that do not contribute to coverage ( #7042 )
...
* adjust tests
* opt shader coverage op
* update gpu_coverage.cpp atomicor coalesce
2026-10-05 08:57:20 +08:00
9f9d4ec815
reduce gpu test cases ( #7017 )
2026-10-01 23:33:05 +08:00
60bfa40044
sdpa flash attention x86 optimization, llm models enable gqa, dispatch for avx2 bfloat2float ( #6923 )
...
* enable gqa
* update perf test
2026-09-01 17:35:42 +08:00
7f162381ab
kvcache allocator ( #6901 )
...
* decode benchmark
* reject kvcache batch
* enable mha vulkan gemm out_elempack 1
2026-08-17 15:44:30 +08:00
13b6d5318c
fix naive mha out-of-bound read ( #6812 )
2026-07-08 13:55:11 +08:00
adee6a0a53
multiheadattention kvcache ( #6333 )
...
* skip input layer forward
* unified kvcache layout
* mha kvcache test
2025-10-15 14:19:57 +08:00
風的彷徨