fix: segmenter tail latent registered before gallocr alloc (DBCache crash)

[KakaruHayate]

Problem

With DBCache enabled (--cache-threshold), the 3-stage segmenter split path
built \outs.latent\ after \ggml_gallocr_alloc_graph, leaving the latent
nodes unallocated. Computing them is UB — on the medium model (latent at
layer 6, fn_blocks=1) the CLI crashed with ACCESS_VIOLATION
(exit 0xC0000005) at the first D3PM step. CI only runs --version, so this
was never caught.

Fix

Register both outputs (x_run + latent) and build the graph before allocating;
alloc only once both are in the graph.

Verified (local CPU build, 30s 44.1k mono, Q8_0, nsteps=4)

• Before: crash 0xC0000005
• After: exit 0, timer 34.3s vs 40.7s cache-off segmenter time
  (DBCache gives ~16% end-to-end, ~20% segmenter speedup at nsteps=4)
• Notes identical (83 == 83) with --seed 42 — same output as cache off.

Summary by CodeRabbit

• Bug Fixes
  • Fixed processing for configurations that use optional latent outputs, preventing missing graph elements during computation.
  • Improved segmenter initialization to ensure all outputs are prepared before processing begins.

[KakaruHayate]

@coderabbitai full review

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: c4a2444d-24f6-4b34-8ae5-a50afc7030df

📥 Commits

Reviewing files that changed from the base of the PR and between b4b5863 and e5c0e80.

📒 Files selected for processing (1)

• src/model.cpp

Included review availability: Your plan includes up to 1 review per rolling hour; 0 remain after this review.

---

📝 Walkthrough

Walkthrough

The segmenter tail now registers and expands the main output and optional latent output before backend graph allocation. It then checks backend support, creates the allocator, and allocates the complete graph.

Changes

Segmenter graph allocation

Layer / File(s)	Summary
Allocate the complete output graph src/model.cpp	The segmenter tail replaces s.finalize with explicit output registration and expansion. It prepares outs.x_run and optional outs.latent before backend validation and graph allocation.

Estimated code review effort: 2 (Simple) | ~10 minutes

Merge Risk: ⚪ Minimal · up to e5c0e

This localized change registers the segmenter outputs before graph allocation and reports successful verification with matching notes; no actionable merge-blocking risk remains beyond normal checks and review.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly identifies the segmenter latent-registration fix and the DBCache crash it prevents.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feature/eff-fix-cache-crash

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

✅ Action performed

Full review finished.