Pipeline
Attention / state retention semantics
-> Retention IR
-> Lifetime / head stripes
-> LayoutProgram
-> PhysicalPlan
-> immutable KVView
-> SemanticDead ∧ ExecutionComplete
-> backend commit
-> reuseGenerated layouts
| Semantics | Layout |
|---|---|
| Full | AppendOnly |
| Sliding | Periodic |
| Sink + Sliding | Pinned + Periodic |
| Same Chunk | ResettableArena |
| Per-head windows | Lifetime-normalized stripes |
Safety
- Logical identity is separate from physical slot and generation.
- Address programs are the only logical addressing authority.
- Reuse requires semantic death, reader completion, and backend commit.
- Unsupported semantics and engine contracts fail closed.
Records
01 / DESIGN
End-to-end plan
Validation stages, ownership boundary, value gates, and remaining work.
Open record →02 / REAL MODEL
gpt-oss-20b validation
Released checkpoint capacity, admission, overhead, and reclamation certificates.
Open record →03 / APPLICABILITY
Three-model applicability
Qwen fallback, Mistral bounded execution, and GPT-OSS Hybrid plans.
Open record →04 / NORMAL FORM
Lifetime Normalization
Per-head multi-scale windows, max-window baseline, and exact retention amplification.
Open record →Boundary.
SGLang ownership is qualified only for SWA ChunkCache with radix, speculative decoding, overlap, disaggregation, and CUDA Graph disabled.
Full owning and VMM qualification