Fix: preserve inactive prefill indexer score rows

[indigo1973]

The golden score buffer is initialized with FP32_NEG_INF for all T
rows, but only the first num_tokens rows are active. Assigning scores
to every row overwrote the sentinel values of inactive rows when
num_tokens < T, causing prefill_score_compare to report false golden
mismatches.

Limit both the source and destination to the active token prefix so
inactive score rows retain FP32_NEG_INF.

Before the fix, start_pos=0 and num_tokens=8 failed score validation
with 240 inactive sentinel mismatches:

• task_20260901_170618_126348616272

Tested on Ascend A5 in the wzh environment with:

• PyPTO: 4054785f520cb093d877038f3063a259d3afc3e3
• runtime/simpler: dbdd041e957420ea15b03e878400dd4de5e9c34c
• PTOAS: 0.60
• newer PTO-ISA: ad509ef4aaa078e0db3f89f9e760b27578099ee7

Test results:

• start_pos=0, num_tokens=8: PASS task_20260901_172411_133338029446
• start_pos=0, num_tokens=128: PASS task_20260901_172455_134010821348

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Team

Run ID: 462de610-96a5-457b-8df7-b5dd96010e01

📥 Commits

Reviewing files that changed from the base of the PR and between b7eafa4 and 1690bc2.

📒 Files selected for processing (1)

• models/deepseek_v4_pro/prefill_indexer.py

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

---

📝 Walkthrough

Walkthrough

The prefill indexer now populates golden scores only for active tokens. Inactive score rows retain their initialized sentinel values.

Changes

Active Token Score Handling

Layer / File(s)	Summary
Restrict score writes to active tokens models/deepseek_v4_pro/prefill_indexer.py	score_full writes now target only the active num_tokens prefix. Inactive rows retain their sentinel values.

Estimated code review effort: 1 (Trivial) | ~2 minutes

Merge Risk: ⚪ Minimal · up to 1690b

This localized change limits score writes to active token rows so inactive prefill score sentinels are preserved. No actionable merge-blocking risk remains beyond normal checks and review.

Poem

I’m a rabbit guarding scores in a row
Active tokens get a careful glow
Sleeping rows keep sentinels in place
No extra writes disturb their space
The indexer hops with measured grace

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 1 functions across 1 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the main change: preserving inactive prefill indexer score rows.
Description check	✅ Passed	The description directly explains the inactive-row sentinel mismatch, the active-prefix fix, and the validation results.
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.

• Fix all pre-merge checks with AI

---

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.}

[lwDavid]

Reviewed.