Repository navigation
Releases
Releases · neucine/affon
Affon 0.4.2
Affon 0.4.2
Downloads
- affon-v0.4.2-darwin-arm64: https://downloads.affon.ai/releases/v0.4.2/affon-v0.4.2-darwin-arm64.tar.gz?v=v0.4.2
SHA256:95e34e1e39bbcf52af14976d20bdcb2afe62d897affd9ae14719ccfdf2f30c90 - affon-v0.4.2-linux-x64: https://downloads.affon.ai/releases/v0.4.2/affon-v0.4.2-linux-x64.tar.gz?v=v0.4.2
SHA256:c069b406ff2a3a7f62f4a62690a16af1ba2e363083f5f7aeb96b120c863c5661
Changelog
- Merge pull request #16 from neucine/release/v0.4.2 (1818e39)
- Prepare Affon v0.4.2 release (a31143f)
- Merge pull request #15 from neucine/affon-fast-secret-scan (a8a3f22)
- Run short secret scan after fast checks (697ff13)
- Merge pull request #14 from neucine/affon-unified-ci (e1a34b2)
- Retire GitHub Actions and Windows checks (41e688d)
- Expose Cargo to Affon runner jobs (3514e88)
- Find pinned Bun in macOS Buildkite jobs (37041b3)
- Isolate Affon Buildkite release dependencies per job (e4a7fad)
- Unify Affon CI and two-host release promotion (51e017b)
- Record Node tensor ownership experiment (0b35a87)
- Add OLMoE model and checkpoint loader (3e3ef15)
- Clarify Affon setup and public guides (e7e9269)
- Point observability gap audit to daydream (f834209)
- Improve Affon training and inference and add Chinese README (e94742e)
- Merge pull request #12 from neucine/buildkite-release (4241e8c)
- Wait for release tag input (cd99f05)
- Add Buildkite release pipeline (b2fe310)
- Use hosted macOS ARM64 runner label (2a3dd7c)
- Run Affon release on hosted runners (8aa883e)
Full diff: v0.4.1...v0.4.2
Affon 0.4.1
Affon 0.4.1
Downloads
- affon-v0.4.1-darwin-arm64: https://downloads.affon.ai/releases/v0.4.1/affon-v0.4.1-darwin-arm64.tar.gz?v=v0.4.1
- SHA256:
2be33e75eec5b3a8270ca100b7024b72fc51c57c841cf217f03a912eab8cf2ee
Changelog
- Release Affon 0.4.1 (5639d3f)
- Upload PowerShell installer to distribution (501b5ef)
- Fix Windows Rust linkage with LLVM target (64c3ad4)
- Rename core API and add Windows release support (f3c784b)
- Add shared editor config (06b5609)
- Sync Affon site in local release flow (1e69d79)
Full diff: v0.4.0...v0.4.1
Validation note
On a GitHub-hosted M1 runner, the Metal f32 trigonometry values and strided views test failed its numeric tolerance check (359 of 360 runtime tests passed). The full release check passed on the maintainer's macOS arm64 machine, and the published macOS archive passed local and remote smoke tests. The M1 result remains under investigation.
Affon 0.4.0
Affon 0.4.0
Downloads
- affon-v0.4.0-darwin-arm64: https://downloads.affon.ai/releases/v0.4.0/affon-v0.4.0-darwin-arm64.tar.gz?v=v0.4.0
- SHA256:
319776e9f5ef571593caeb0527bf4a2bb34fd7dc502d9e7410afe7c52eda359e
Changelog
- Release Affon 0.4.0 (0030a9e)
- Run operation e2e tests on available devices (5e64b8f)
- Limit public Affon tests to standalone policy checks (bcd7ccc)
- Keep public Affon checks independent of private repositories (c8018c0)
- Run public PR checks on hosted runners (b782102)
- Expose Cyna dot graph operation in Affon adapter (51daf69)
- Check resident Cyna memory fusion through Affon (b57ecce)
- Check resident Cyna vector and batch matmul through Affon (3d6bb7a)
- Restore CI tests with compatible dependency pins (d8b4f85)
- Canonicalize release flow and use local Actions runners (882add0)
- Move retained parity and benchmark evidence out of affon (6d9688f)
- Check resident quantized graph output in Affon adapter (34ffb6c)
- Submit quantized matmul through Cyna graph adapter (ca284c5)
- Check resident loss graph through Affon (82aef60)
- Run metadata-aware resident graphs from Affon (b227552)
- Check resident softmax through Affon adapter (19cca20)
- Check resident floating unary graphs through Affon (22f899d)
- Compare resident statistics through Affon adapter (69530e2)
- Check resident dense extrema through Affon adapter (1824d08)
- Exercise resident f32 extrema through adapter (41691d1)
- Exercise resident CPU unary graph through adapter (500b441)
- Exercise resident bf16 graph adapter (284167e)
- Exercise resident f64 axis reductions (aef41e5)
- Exercise resident u32 compute core graph (773d4df)
- Select resident i64 axis sums in linked graph (041c756)
- Exercise resident i64 compute core adapter (ff95b6a)
- Exercise resident f64 compute core adapter (f3ff816)
- Link Affon adapter to Cyna resident CPU graph (9558392)
- Verify sorted flag in Cyna top k graph (0f39ab0)
- Verify f64 multiply add through Cyna adapter (587a41d)
- Verify f64 fused graph through Cyna adapter (053cddf)
- Run Cyna normalization and loss graph from Affon (0b41ea4)
- Send scalar metadata through Affon Cyna adapter (3769f19)
- Compose Cyna cat stack and shape reduction from Affon (4e08ffa)
- Adapt Affon top k Cyna graph outputs (bb9d207)
- Run indexed Cyna graphs from Affon adapter (267bcba)
- Send slice ranges through Affon Cyna adapter (6735d19)
- Send layout metadata through Affon Cyna adapter (8957742)
- Adapt Affon Cyna axis swap results (e89ae0d)
- Adapt Affon structural Cyna graph results (8f4660f)
- Adapt Affon mixed dtype cast results (56618bb)
- Check Affon conditional graph memory result (98eeea1)
- Select Cyna conditional graph results in Affon adapter (562493c)
- Verify Cyna comparison graph outputs in Affon (512cb4a)
- Own heterogeneous Cyna graph results in Affon (c9f6643)
- Allocate owned mixed Cyna graph results in Affon (3a0f890)
- Run mixed dtype Cyna byte graphs from Affon (d4fc026)
- Infer typed Cyna graph outputs in Affon adapter (59cd99b)
- Verify Cyna dense statistics and integer means in Affon (d4c671d)
- Verify i64 index reductions through Cyna adapter (be04ddf)
- Use Cyna dense extrema reductions in Affon adapter (fc03c43)
- Add owned Cyna dense graph results to Affon adapter (3de8e70)
- Expose Cyna graph shape inference to Affon (605ffe1)
- Submit bf16 Cyna graphs from Affon adapter (0215111)
- Submit u32 Cyna graphs from Affon adapter (94d30b8)
- Submit f64 Cyna graphs from Affon adapter (81b0497)
- Submit i64 Cyna graphs from Affon adapter (501a0a7)
- Call Cyna i64 matmul from Affon adapter (4153d40)
- Select Cyna graph optimization goal from Affon (6c0ca8f)
- Submit scalar Cyna graphs from Affon (b0ab054)
- Run Cyna softmax graphs from Affon adapter (d7a3e5e)
- Use Cyna unary graph operations from Affon (cfc1201)
- Add typed Affon adapter for Cyna dense graphs (00371c7)
- Add Affon Cyna reduction adapter (a1f529a)
- Add Affon quantized Cyna compute adapter (62f9b27)
- Add Affon Cyna compute core consumer probe (acf2fc9)
- Record validated CUDA parity dispatch improvements (4b575ab)
- Capture CUDA parity and performance evidence (ad34159)
- Emit Program compute timing and capture telemetry (cb13098)
- Speed up CUDA nested tensor readback (567b133)
- Own public affon.ai site content in the product repository (6a917b8)
- Exclude generated performance report from language statistics (c9894f3)
- Capture Affon parity and Metal performance work (f1350d9)
- Reduce Apertus logit selection and softmax work (fe961af)
- Profile matched Apertus kernel and host scheduling (fd5de06)
- Match Apertus functional KV cache in PyTorch peer (079b1ee)
- Align Apertus PyTorch inference benchmark protocol (e8d1bed)
- Broadcast grouped Apertus KV heads during attention (fde632d)
- Record Apertus decode fusion probes and regressions (0000c5e)
- Benchmark Apertus CUDA ops and full decode against PyTorch (84ec8a2)
- Retain CUDA RoPE concat parity and Apertus trace (7568457)
- Retain CUDA f32 index bounds regression evidence (66b843c)
- Record CUDA index parity and Apertus token rate (3f36e37)
- Retain Apertus CUDA reduction parity and token rate (29809e6)
- Keep inference benchmark programs with harness (48db9d5)
- Consolidate inference benchmark harness under test (883e2f1)
- Use singular hf inference benchmark directory (b4feb35)
- Use singular benchmark directory (b924902)
- Group focused quantized benchmark scripts (ad718a8)
- Move benchmark facility under test (772efc4)
- Refresh CUDA bf16 arg-all parity evidence (9a519d7)
- Refresh CUDA short all-axis reduction parity evidence (607e1bd)
- Refresh CUDA softmax parity after byte API change (d2d0546)
- Add tensor byte readout and warn on large arrays (2294222)
- Refresh CUDA concatenation parity evidence (ee2fddf)
- Refresh CUDA long row softmax parity evidence (26d7cd3)
- Refresh CUDA bf16 indexed loss parity after parallel kernel (748a70e)
- Refresh CUDA indexed loss parity after parallel kernel (3bd9b06)
- Refresh CUDA f32 axis reduction parity evidence (79ec41c)
- Refresh CUDA all-axis reduction parity evidence (632b2b7)
- Refresh f32 strided reshape parity timing (97ff268)
- Refresh top three Metal parity gaps (67eb8f2)
- Refresh Metal norm parity after 64 element cutoff (53c6028)
- Refresh Metal norm parity and BF16 cutoff evidence (ab5673d)
- Refresh Metal f32 top-k parity evidence (94188d0)
- Refresh Metal top-k and stale parity evidence (1fa5212)
- Refresh long row log softmax parity evidence (8f4c077)
- Refresh batched one-hot and bf16 normalization parity evidence (b7b184d)
- Refresh large Metal f32 reduction parity evidence (dc33d63)
- Refresh Metal variance and std parity evidence (6751662)
- Refresh batched Metal top-k parity after parallel kernel (383f89d)
- Refresh Metal f32 arg reduction parity evidence (ed515ee)
- Refresh Metal scalar reduce to shape parity evidence (8cab716)
- Refresh Metal reduce to shape parity evidence (9b7614b)
- Refresh Metal bf16 normalization parity evidence (61b6c42)
- Refresh Metal top-k parity evidence after f32 optimization (d1a1d9b)
- Improve program inspection and integrate inference updates (758c443)
- test: complete Affon e2e behavior coverage (b7f4e48)
- Add focused parity runs and independent regression comparisons (ec776c2)
- Add permanent parity facility and retained GPU coverage evidence (eae71ed)
- Add Apertus model definition (c76d700)
- Expose bf16 u32 and composable Conv1D APIs (3e9fcb1)
- Improve local Hugging Face inference (56112eb)
- Expose inspection and optimization telemetry APIs (270c97e)
- make decoder model directly callable (57d0747)
- standardize callable training composition (2475df0)
- Make losses specialized callables (dc6b770)
- Unify callable program composition (6d38c03)
- Remove candidate compute harness (6e4eafc)
- Remove legacy APIs and align Program workflows (ef28fda)
- Expand program training and ops APIs (5884f74)
- feat: introduce declarative program API (5cd2ce8)
- Cut AFFON over to program compute core (cd1f620)
- Profile SmolLM2 with native telemetry and batch Metal inference (b5299dd)
- Add native SmolLM2 inference and incremental text generation (b238a8d)
- Keep implementation docs and archive working evidence outside source control (a577474)
- Move Whisper execution into its model family (8be7378)
- Group Hugging Face processors and Hub snapshot handling (164ec26)
- Organize model families and separate Hugging Face checkpoint adapters (ea7bb2b)
- Record Whisper bottleneck investigation and softmax batching results (8e2d7af)
- Enable ONNX graph batching and record compute performance coverage (1766e79)
- Design bounded Metal graph execution scopes (6572674)
- Add multi-modal Hugging Face inference and native model playground (f388944)
- Merge remote-tracking branch 'origin/main' (2e5de48)
- Update notebook examples and execution outputs (89a55d9)
- Merge pull request #4 from neucine/dependabot/github_actions/github-actions-6512b1d693 (c1d3f14)
- Include Hao standard module declarations in Affon type package (ca7c1cf)
- Associate both TypeScript notebook suffixes with notebook editor (6900170)
- Migrate TypeScript examples to standard ts.ipynb notebooks (fef018e)
- Rename notebook examples to tsnb (135de80)
- Expose Hao Jupyter kernel from Affon CLI (17e7053)
- Print startup device information (baac091)
- Auto-detect default compute device (c361528)
- Bump softprops/action-gh-release in the github-actions group (5268e2b)
Full diff: v0.3.0...v0.4.0
Affon 0.3.0
Affon 0.3.0
Downloads
- Linux x64 archive: https://github.com/neucine/affon/releases/download/v0.3.0/affon-v0.3.0-linux-x64.tar.gz
- Linux x64 standalone binary: https://github.com/neucine/affon/releases/download/v0.3.0/affon-v0.3.0-linux-x64
- Archive SHA256:
a7b18d59c9f7afc03a05dbabdc8b0ecdf9dd7985b4f31336d11737242f86f6dd - Binary SHA256:
86d96c95cbca8dccaf956571026ee729409c72b376abb045cb3897df53971523 - macOS artifacts will be uploaded separately.
Highlights
- Added the Linux CUDA backend with runtime-selected NVIDIA devices, NVRTC kernels, and cuBLAS matmul.
- Added CUDA tensor, autograd, optimizer, graph-execution, loss, indexing, and decoder-LM training coverage.
- Added CUDA memory telemetry and device-neutral decoder training metrics.
- Separated compiled training from eager validation so validation observes current parameters.
- Added platform-aware Linux CUDA and macOS Metal release tests.
Changelog
- Prepare v0.3.0 CUDA release (707845a)
- Separate compiled training from validation forward (4da05f7)
- Expand Metal matmul benchmark coverage (8d5c2af)
- Add Metal matmul benchmark harness (09e67f2)
- Add CUDA training overrides and neutral pool metrics (7c8ce39)
- Integrate CUDA training, graph execution, and decoder support (9fc68cf)
- Attribute embedded Hao runtime memory (052e890)
- Check compute memory regions at shutdown (2dedacc)
- Refine decoder LM training workflow (937bbc6)
See the repository CHANGELOG.md for the full release notes.
Affon 0.2.0
Affon 0.2.0
Downloads
- affon-v0.2.0-darwin-arm64: https://downloads.affon.ai/releases/v0.2.0/affon-v0.2.0-darwin-arm64.tar.gz?v=v0.2.0
- SHA256:
5dcba3281c6d51bc43bda644a8f43aef011cfdfee10fbac7dacd3ff68989af43
Changelog
- Prepare 0.2.0 release (9d01521)
- Enrich public docs and expand test coverage (2595e19)
- Use external compute package (179b045)
- Decouple TS capture semantics from compute core (6e9e7f4)
- Organize compute shared types and sema (1ced4ae)
- Add bottleneck detail to compute trace spans (7f8ab94)
- Normalize compute telemetry metric names (2ea11c2)
- AFFON-0: let Hao own runtime configuration (e6a0ddd)
- AFFON-0: Fix CI and release packaging (0b9f2fb)
- Prepare Affon Next main branch (ffd0d1d)
- AFFON-: remove Metal CPU staging (6df07ad)
- Align compute telemetry and bindings (d83af3b)
- AFFON-: complete compiled where lowering (30d3c20)
- AFFON-30: Complete compute, NN, and dataset extraction (50f995a)
- Extract compute runtime and close parity gaps (2c10624)
- AFFON-29: Continue compute binding parity (0166c19)
- AFFON-28: Fold telemetry types into compute (a2b079c)
- AFFON-27: Pin telemetry error status API (1196645)
- AFFON-26: Polish compute instrumentation (308c978)
- AFFON-25: Instrument planning phases (2a34e2d)
- AFFON-24: Pin shared telemetry interface (50ed02b)
- AFFON-23: Use shared telemetry interface (1df6d45)
- AFFON-22: Bundle source dependencies in releases (db52eea)
- AFFON-21: Pin shared dependency revisions (9f532c3)
- AFFON-20: Wire shared zig-libs telemetry (f72d4fb)
- AFFON-19: Hide backend from public compute API (3125df8)
- AFFON-18: Export standalone compute module (8dc40a8)
- AFFON-17: Complete specialized fusion modules (04dbf2c)
- AFFON-16: Add logsumexp loss fusion (822df01)
- AFFON-15: Reclaim dead graph values (e5d1914)
- AFFON-14: Fuse attention score graphs (67c3cd8)
- AFFON-13: Fuse add layer norm regions (1c68d03)
- AFFON-12: Separate graph fusion execution (4ad3151)
- AFFON-11: Execute GELU matmul epilogues (b5e7f48)
- AFFON-10: Execute matmul epilogue regions (4b1c54f)
- AFFON-9: Execute graph fusion chains (7495809)
- AFFON-8: Keep autograd outside compute core (d52cbb7)
- AFFON-7: Add autograd tape state (e892c67)
- AFFON-6: Add core graph execution (55afd0b)
- AFFON-5: Organize compute model layers (2b18a30)
- AFFON-4: Move eager compute execution (66dc790)
- Move compute ops and CPU kernels to Affon (00be38d)
- Move compute tensor core to Affon (4c6ce39)
- Link Hao through one runtime artifact (3ec173b)
- Initialize Affon package on Hao (54c8675)