Repository navigation

vela: add the 0.8B native response finishing and backbone execution on CUDA - #177

Open

MrDongsls wants to merge 47 commits into

InferMatrixLab:mainfrom

MrDongsls:model/vela2-native-executor

Open

vela: add the 0.8B native response finishing and backbone execution on CUDA#177
MrDongsls wants to merge 47 commits into
InferMatrixLab:mainfrom
MrDongsls:model/vela2-native-executor

Commits on Oct 24, 2026
- vela: add the 0.8B native bundle export, manifest gate and wire contract
  Mr.Dong
  committed
  739e5ff
- vela: add the 0.8B native token-level processing
  Mr.Dong
  committed
  9812853
Commits on Oct 25, 2026
- vela: add the 0.8B native token-level processing
  Mr.Dong
  committed
  7391de2
- vela: add the 0.8B native F32 heads, scoring and span decoding
  Mr.Dong
  committed
  dd5064a
- vela: add the 0.8B native F32 heads, scoring and span decoding
  Mr.Dong
  committed
  7d50f57
- vela: add the 0.8B native SystemOne response finishing
  Mr.Dong
  committed
  82dfa0a
- vela: add the 0.8B native backbone execution on CUDA
  Mr.Dong
  committed
  149a9ba
- Merge remote-tracking branch 'upstream/main' into model/vela2-native-processing
  Mr.Dong
  committed
  57994cb
- Merge remote-tracking branch 'upstream/main' into model/vela2-native-heads
  Mr.Dong
  committed
  ba86215
- Merge remote-tracking branch 'upstream/main' into model/vela2-native-executor
  Mr.Dong
  committed
  b6dc583
- Merge model/vela2-hybrid-reference-serving and keep the Vela native rows alongside the Clef/pplx-decider additions
  Mr.Dong
  committed
  1685b5b
- Merge model/vela2-native-processing: keep the canonical shared layer, add the Clef/pplx rows
  Mr.Dong
  committed
  6bedca9
- Merge model/vela2-native-heads: keep the executor layer, fold in the Clef/pplx rows
  Mr.Dong
  committed
  30e0054
Commits on Oct 26, 2026
- vela: restore float_roundtrip and drop the duplicate sha2 dev-dependency
  Mr.Dong
  committed
  df901b2
- vela: make the heads parity test ignored with tighten-only env gates
  Mr.Dong
  committed
  e6715a3
- vela: score structural mismatches as full drift in the heads comparator
  Mr.Dong
  committed
  44012d6
- vela: verify the bundle manifest and hidden width before the heads load
  Mr.Dong
  committed
  1a169bc
- vela: record the chunk-packing rounding constraint in run_all
  Mr.Dong
  committed
  774d589
- vela: make the response drift strict and put the opt-in checks behind --ignored
  Mr.Dong
  committed
  76f233e
- vela: answer non-object bodies with the reference worker's 400 message
  Mr.Dong
  committed
  2de81d5
- vela: make the render-parity test opt-in behind --ignored
  Mr.Dong
  committed
  c1e4edb
- vela: index unit_ids by code points, not byte offsets
  Mr.Dong
  committed
  cca9174
- vela: fail the bundle gate closed on unpinned files and an empty pin table
  Mr.Dong
  committed
  9f13566
- Merge branch 'model/vela2-hybrid-reference-serving' into model/vela2-native-processing
  Mr.Dong
  committed
  ddaecd8
- vela: recount the always-run tests after the bundle-gate cases
  Mr.Dong
  committed
  0f1e4da
- Merge branch 'model/vela2-native-processing' into model/vela2-native-heads
  Mr.Dong
  committed
  c2d704d
- Merge branch 'model/vela2-native-heads' into model/vela2-native-executor
  Mr.Dong
  committed
  064244d
Commits on Oct 31, 2026
- vela: stage the native export atomically
  Mr.Dong
  committed
  b9a60c0
- Merge branch 'model/vela2-native-processing' into model/vela2-native-heads
  Mr.Dong
  committed
  86f6898
- Merge branch 'model/vela2-native-heads' into model/vela2-native-executor
  Mr.Dong
  committed
  bc0d268
- vela: tighten the GPU logit gate and close the bundle walk
  Mr.Dong
  committed
  65766b8
Commits on Nov 1, 2026
- vela: restore the export manifest and stage bf16 inside the bundle
  Mr.Dong
  committed
  eadb785
- vela: anchor the manifest gate and close the unsafe key surface
  Mr.Dong
  committed
  eec1fe6
- Merge branch 'model/vela2-native-processing' into model/vela2-native-heads
  Mr.Dong
  committed
  ea6b0de
- vela: run the heads and calibration loads through the anchored gate
  Mr.Dong
  committed
  0734903
- fix(vela): keep the verified package reusable after token-plan export
  Mr.Dong
  committed
  e7643f3
- Merge remote-tracking branch 'infermatrix/main' into model/vela2-native-processing
  Mr.Dong
  committed
  9576f78
- Merge branch 'model/vela2-native-processing' into model/vela2-native-heads
  Mr.Dong
  committed
  8aa28fb
- fix(vela): end spans at rejected words and gate logits on finiteness
  Mr.Dong
  committed
  ce8c0ca
- Merge model/vela2-native-heads: fold in the anchored loads and keep the executor layer
  Mr.Dong
  committed
  5301e6f
- fix(vela): tighten the GPU logit gate on shape and finiteness
  Mr.Dong
  committed
  bc0fb9c
Commits on Nov 2, 2026
- drop the stray 4B/9B head-input fixtures and unrelated skill file added by the 5301e6f5 merge
  Mr.Dong
  committed
  3385ccb
- gate the token-plan recorder behind package verification
  Mr.Dong
  committed
  7f9ec58
- Merge branch 'model/vela2-native-processing' into model/vela2-native-heads
  Mr.Dong
  committed
  74b0ed8
- gate the head-input recorder behind package verification
  Mr.Dong
  committed
  8377aea
- drop a duplicated doc line from the recorder gate merge
  Mr.Dong
  committed
  af51215
- Merge branch 'model/vela2-native-heads' into model/vela2-native-executor
  Mr.Dong
  committed
  f34c015

Commits

Commits on Oct 24, 2026

Commits on Oct 25, 2026

Commits on Oct 26, 2026

Commits on Oct 31, 2026

Commits on Nov 1, 2026

Commits on Nov 2, 2026