Giorgio Oppo
37f66ba3c9
Scope the legacy half/unsplit projection policy and parent-chunk attention
dispatch to SSD streaming graphs. Resident prompts again use main's FP32
small-batch F16/Q8 projections, F32 split-K reductions and per-range attention
at the dense/sparse boundary. This corrects a policy regression that changed
GDN state and subsequent greedy tokens. Decode and SSD arithmetic are retained.
Add a bounded regression test with a frozen 0aaea5a dispatch oracle, independent
attention references, negative controls, guard regions and immutable inputs
and weights. Exercise retained/unretained command buffers and fast/safe Metal
math. Require SSD mode for the existing legacy-unsplit generation reference.
Archive source provenance, replay helpers, a frozen HC shader oracle, the
resident fixture generator and full-vocabulary/greedy A/B captures. Exclude
models, executables, object files and generated source checkouts. Pending
experimental production kernel changes remain outside this commit.
Validation on Apple M1 Max, 32 GiB:
- Build ds4, ds4-server and regression tests from exactly the staged sources.
- Four Metal validation/math/command-buffer configurations pass; the old
resident policy fails the regression test as expected.
- Resident repeated-weight Q4 fixture: all 248320 logits are bit-identical to
main for 29, 575 and 5942-token prompts; both 64-step greedy dumps match.
- Full Q4 SSD: 29 and 575-token prompt dumps are byte-identical before/after.
- HC shader reference: 71793 comparisons, no differences.
- One-shot/session generation and SSD legacy-reference tests match all logits
across the prompt frontier and eight continuation tokens.
- Memory planning and whitespace checks pass.
The resident fixture uses repeated real layer weights and zero PLE. Results
are arithmetic regressions, not quality or performance claims; the tester's
full Q2/M4 Max configuration still needs external confirmation.