Repository navigation

[FEAT] Allow local preemption during speculative decoding - #52

Open

0z5a wants to merge 12 commits into

ThinkFlowLab:mainfrom

0z5a:rlt-spec-sched-only

Commits on Sep 24, 2026
- Support speculative round interleaving and request migration
  0z5a
  committed
  88b9f0d
- Record RTX 5090 scheduling and migration E2E results
  0z5a
  committed
  9d8fb39
- Verify graph replay across an interleaved speculative round
  0z5a
  committed
  31d46b8
- Validate graph replay on combined speculative scheduler
  0z5a
  committed
  464f711
- Cover graph replay through preemption and migration
  0z5a
  committed
  d02b4ff
- Keep speculative drafts on device and record RTX 5090 E2E
  0z5a
  committed
  892d82f
- Skip saturated speculative admission scans and benchmark queued E2E
  0z5a
  committed
  704b306
- Measure saturated speculative admission cost
  0z5a
  committed
  6b9ef19
- Record RTX 5090 scheduler and Graph E2E comparisons
  0z5a
  committed
  b5b0b78
Commits on Sep 28, 2026
- Drop benchmark artifacts and align scheduler stack
  0z5a
  committed
  d44e260
Commits on Oct 1, 2026
- Merge main and resolve upstream conflicts
  0z5a
  committed
  35cf1a7
Commits on Oct 9, 2026
- Resolve PR52 review and keep local speculative preemption
  0z5a
  committed
  2597e7b