faster-rwkv

Commits

• 644b53b: Update termux script (#13) (daquexian) #13
• a5673dd: update README.md about termux script (daquexian)
• a68ffc3: don't build tests when cross compiling (daquexian)
• d855d2f: midi music support (daquexian)
• ea6270e: update midi (daquexian)
• 753de73: many updates (#15) (daquexian) #15
• a4f3b50: update sampler and add tests (daquexian)
• 82f0f28: fix binary op shape inference, add wip onnx code (daquexian)
• e9c3384: rename x-meta to export-x (daquexian)
• 0205cef: fix ci (daquexian)
• 513371a: support init state (daquexian)
• ca033d0: add preprocess_init_prompt.py (daquexian)
• 9f4768b: fix the wrong help msg of midi_music.cpp (daquexian)
• f96e958: set top_p to 0.3 in chat.cpp (daquexian)
• eafafe8: support set seed of sampler and gives the same result on platforms (daquexian)
• 15440d8: fix java method typo (daquexian)
• 3503a10: add the missing setSeed java method (daquexian)
• 5003705: arm int4 (#18) (daquexian) #18
• ccc096a: check fp16 is supported when model is int4 (daquexian)
• ae4e819: refine error msg, reduce ncnn embed size (daquexian)
• d63b3eb: update path in ci (daquexian)
• 7833153: update download_binaries_and_models_termux.sh (daquexian)
• 6d2affb: update download_binaries_and_models_termux.sh and readme (daquexian)
• b895c25: update download_binaries_and_models_termux.sh and readme (daquexian)
• 2e79729: remove '-g' flag in android toolchain file in ci build (daquexian)

Commits

• 5003705: arm int4 (#18) (daquexian) #18

Commits

• 6a51c6a: use ncnn gemm for 2d matmul, faster (WIP: call multiple gemms to do batch matmul) (daquexian)
• 6d9503d: update download links (daquexian)
• ac32b5e: CI (#7) (daquexian) #7
• 1e4f175: android aar build and test (#8) (daquexian) #8
• 233177c: build and upload pure export_ncnn (#9) (daquexian) #9
• de21df0: refine export_ncnn.cpp (daquexian)
• 589357b: update README.md (daquexian)
• 43a1dfc: reorg readme (daquexian)
• 2746fd7: auto release (daquexian)
• 57e5d2d: add the concept of graph-grained backend (#10) (daquexian) #10
• 60fe59a: update README.md (daquexian)
• 448e5d2: only create latest release when pushing to master (#11) (daquexian) #11
• 990f683: fix error on legacy models (daquexian)
• 6599098: remove unused EXPECT_EQ (daquexian)
• cc3e65f: ncnn a32w8 (#12) (daquexian) #12
• 8041ed1: run ci on every push and pr (daquexian)

Commits

• 60fe59a: update README.md (daquexian)

Support loading ncnn models from Android assets

Example:

model = new Model("asset:RWKV-5-ABC-82M-v1-20230901-ctx1024-ncnn", "ncnn fp32", getAssets());

v5 models support (great thanks to @AsakusaRinne )
music models support
temperature, top_k and top_p support
some bugfixes by @nihui and @AsakusaRinne

more models are published at https://huggingface.co/daquexian/fr-models/tree/main

NOTE: the aar file is actually built from commit 055be4, newer than tag v0.0.2

Release v0.0.1