Pull requests / #169
#169 generate/verify: make diagnostic dumps work under --spec for native packs
closed · @samuelishida · 0 comentarios · En GitHub
Multi-GPUAMD / HIPDocumentationWindows
Descripción
A native (IQ) pack decodes through verify windows only, so `--dump-logits`, `--dump-layers`, and `--dump-final-r` must emit rows from those windows rather than from the ordinary token loop. What this PR does: - Captures the residual ladder inside the verify-window graph when `--dump-layers` is set. Rows are the fused-layout residual (layer l's write stays pending inside layer l+1's fused read), which is what the engine actually holds. - Writes `--dump-logits` and `--dump-final-r` rows for accepted window rows only; rejected drafts are conditioned on tokens the ordinary path never feeds. - Rewrites the logits header at close with the real row count so readers see the size the file actually holds. - Refuses `--dump-halves` under `--spec` cleanly. - Decouples the expert-tier residency-table staging from the dump-gate so the table is available both to the token graph and to verify windows, while preserving the original gate for `--no-capture` / `--no-token-graph` paths. - Provision-gates `ple_parity` in CMake because its >100 MB fixtures and per-artifact oracle include were never committed; the missing supply chain is documented in `tests/python/README.md`. Also documents the native-pack dump semantics in `docs/AMD_HIP.md`.
En el sitio
Enlaces a install, modelos, releases.