Repository navigation

fix(trainer-rank): raise routed-row estimates from checkpoint observations - #981

Draft

bradhilton wants to merge 29 commits into

dalinar/cp-layout-recompute-floorfrom

dalinar/routed-share-observation

Draft

fix(trainer-rank): raise routed-row estimates from checkpoint observations#981
bradhilton wants to merge 29 commits into
dalinar/cp-layout-recompute-floorfrom
dalinar/routed-share-observation

Commits on Sep 26, 2026
- Raise EP routed-row pricing to each checkpoint's observed routing
  
  bradhilton
  and
  claude
  committed
  f16b126
- Agree on checkpoint targets; observe no-grad waves; seed snapshots
  
  bradhilton
  and
  claude
  committed
  6cca90b
- Also agree on the route epoch counter at loads; cover snapshot targets
  
  bradhilton
  and
  claude
  committed
  e34327d
- Read the route epoch counter with a default for bare trainers
  
  bradhilton
  and
  claude
  committed
  572be04
- Collect HybridEP managers into a typed list
  
  bradhilton
  and
  claude
  committed
  43db31f
- Merge the rebased per-rank CP2 layouts into routed-share observation
  
  bradhilton
  and
  claude
  committed
  b4ce830
- Merge the per-rank CP2 layout review fixes into routed-share observation
  
  bradhilton
  and
  claude
  committed
  4b87e8f
Commits on Sep 27, 2026
- Merge the layout-path TE test into routed-share observation
  
  bradhilton
  and
  claude
  committed
  3999370
Commits on Sep 28, 2026
- Merge the layout-boundary adapter-gradient port (#978 53feda89b)
  
  bradhilton
  and
  claude
  committed
  b978418
- Merge the #978 combine-floor test (5163520d8)
  
  bradhilton
  and
  claude
  committed
  5b6a947
- Merge the #978 per-rank adapter pairing (75a3912a7)
  
  bradhilton
  and
  claude
  committed
  66e945a
- Merge #978's sequential gradient groups (f47a06f27)
  
  bradhilton
  and
  claude
  committed
  59da6f8
- Merge #978's pending-gradient tests (29c6a82ad)
  
  bradhilton
  and
  claude
  committed
  1645975
- Merge #978's closed-form group order and pending reuse (500553bb4)
  
  bradhilton
  and
  claude
  committed
  1768ab1
- Merge #978's two-slot split test (6eb86f7b4)
  
  bradhilton
  and
  claude
  committed
  e309eaa
- Merge #978's two-slot lower-bound assertion (bd26e7e9a)
  
  bradhilton
  and
  claude
  committed
  dd4adfb
- Merge #978's layer-mismatch test
  
  bradhilton
  and
  claude
  committed
  ca126c5
- Merge #978's staged head and decoder pricing
  
  bradhilton
  and
  claude
  committed
  ea9ea84
- Merge #978's head-stage row state
  
  bradhilton
  and
  claude
  committed
  1d79f15
- Merge #978's traced head staging
  
  bradhilton
  and
  claude
  committed
  44b622e
- Merge #978's fused-statistics proof for head staging
  
  bradhilton
  and
  claude
  committed
  c9b12b6
- Merge #978's TE-free profile readings
  
  bradhilton
  and
  claude
  committed
  9dda1d6
- Merge #978's strict staged head statistics and V1t revert
  
  bradhilton
  and
  claude
  committed
  0404e27
- Merge #978's per-kernel fused-statistics proof
  
  bradhilton
  and
  claude
  committed
  a8a3d5c
- Merge #978's strict binding per thread, group and staged admission
  
  bradhilton
  and
  claude
  committed
  e288179
Commits on Sep 29, 2026
- Raise EP routed-row pricing to each checkpoint's observed routing, on the extracted layout
  
  bradhilton
  and
  claude
  committed
  b6bed2f
- Merge #978's round-2 port (645c906e9) into the routed-share port
  
  bradhilton
  and
  claude
  committed
  b76895a
- Merge #978's ungrouped layouts test (bf79e4468)
  
  bradhilton
  and
  claude
  committed
  726317d
- Record #981's pre-extraction history (e28817973) as ported
  
  bradhilton
  and
  claude
  committed
  07bc450

Commits

Commits on Sep 26, 2026

Commits on Sep 27, 2026

Commits on Sep 28, 2026

Commits on Sep 29, 2026