Development Build

Commits

• be717cf: fix rwkv2onnx relative path bug (daquexian)

v0.0.14

Tests

• add kernel benchmark (#25) #25 (Rinne)

Commits

• 0c570a5: fix wrong cuda architecture (daquexian)
• f078aee: add doc about quantization (daquexian)
• 7b50ff7: update doc (daquexian)
• a770c37: update doc (daquexian)
• 24a7649: update doc (daquexian)
• 6a1d331: add dependabot (#16) (nihui) #16
• be19206: improve precision (cuda), misc changes (daquexian)
• 6f65056: update (daquexian)
• e944b5d: [IMPORTANT] fix degraded performance of chat example (daquexian)
• d3ec403: Revert "update" (daquexian)
• b0de833: Revert "improve precision (cuda), misc changes" (daquexian)

v0.0.13

Commits

• 161c86f: fix windows error (daquexian)

v0.0.12

Commits

• cb754c6: fix absolute output path when converting to onnx (daquexian)

v0.0.11

Commits

• 85e6fee: update README (daquexian)
• 46d5f22: fix rwkv2onnx v4 model conversion (daquexian)

v0.0.10

Commits

• 1062de2: fix rwkv2onnx package (daquexian)

v0.0.9

Commits

• d45a0c9: update README for onnx exporting (daquexian)
• 2abf009: refine rwkv2onnx (daquexian)

v0.0.8

Commits

• 2c83339: fix classifier in pyproject.toml (daquexian)

v0.0.7

Features

• add support for v5 seq mode inference on cuda (#19) #19 (Rinne)
• support one mode of v5.1 and v5.2 model on cuda. (#22) #22 (Rinne)
• support v5.2 seq mode on cuda. (#23) #23 (Rinne)

Commits

• 2680cd9: update termux script for new int4 model (daquexian)
• 85544f8: export onnx (daquexian)
• 2fd6f2f: support printing state (daquexian)
• 067b538: fix exporting onnx (daquexian)
• 9e3b33a: onnx backend works (daquexian)
• a41accb: update onnx backends (daquexian)
• 528b451: support onnx fp16 and external data (daquexian)
• 2c3edfc: support exporting v5.1/2 onnx models (daquexian)
• 81d90f4: fallback to one mode if seq mode is not implemented (daquexian)
• 7ae8645: fix onnx fp16 bug (daquexian)
• d3fdc4b: rwkv2onnx pip package (#24) (daquexian) #24

v0.0.6

Commits

• 00c1f98: update download_binaries_and_models_termux.sh for v0.0.5 (daquexian)
• 3058192: always re-download chat in download_binaries_and_models_termux.sh (daquexian)
• b5056e6: refactor tokenizer (daquexian)
• ee62fcb: update ci for the compatibility of old tokenizers (daquexian)
• e888925: fix and update ci (daquexian)
• d2d954c: add missing import in java (daquexian)
• bbb4a00: update termux script (daquexian)
• bfd6347: fix wrong test (daquexian)
• 6b1579e: add more error msg (daquexian)
• 5eda245: fix bug about the last line in config file (daquexian)
• 50b4dc7: add the missing nativeHandle field in java classes (daquexian)
• 373af55: remove wrong check (daquexian)
• 0903b8c: tokenizer supports init from asset (daquexian)
• 72befec: follow benchmark_world.py (daquexian)
• 702f2b1: add test about ResetStates (daquexian)
• a3d5dc0: add python binding (#20) (daquexian) #20
• bbea06a: add a python example (daquexian)
• d9099f9: a16w4 naive impl (daquexian)
• ff236fa: update int4 (daquexian)
• d29465f: fix world tokenizer bug (daquexian)
• 365871e: ppl test: remove the dependency of pytorch, support custom size (daquexian)
• 82f45a2: update cmake for termux native build (daquexian)
• 955bea6: update cmake for termux native build (daquexian)
• 336bf49: force enable rtti when python binding is enabled (daquexian)
• a2e6592: fix cmake error (daquexian)
• 22e59e9: update ppl script (daquexian)
• ebbfbf6: update (daquexian)
• 1f66def: refine new int4 dq (daquexian)
• d76dfe0: fix ci path (daquexian)
• ebe2c71: int8 embedding (daquexian)
• bec6c48: relax abc ppl range in ci (daquexian)
• cab0573: relax threshold in ci (daquexian)
• c1f0e4a: relax threshold in ci (daquexian)
• 036dbd9: relax threshold in ci (daquexian)