fix: repair family parity and dependency handling

[yifeif-nv]

Background

Several declared single-GPU cases fail on numerical drift, missing tokenizer metadata, reference dependencies, or accumulated temporary bundles. These changes repair those paths while preserving the original numerical thresholds and frame integrity checks.

Exit Criteria

• The affected original cases pass with unchanged inputs and numerical thresholds; frame order, count and completeness checks remain strict.
• Current-main integration and required checks complete before merge.

Implementation

• Accumulate LFM2 convolution products in FP32; use optimization level 0 for ELF's FP32 main engine, retaining level 1 for FP16.
• Preserve Llama's inline chat template, load Qwen MoE's inline template, and normalize Nemotron Speech language-token metadata in a staged copy.
• Declare PersonaPlex's reference dependencies and its pip build-isolation policy.
• Release passed E2E cases' temporary directories before subsequent cases.
• Complete Wan2.2 reference dependencies and accept the native/reference frame filenames while enforcing a continuous, ordered, homogeneous sequence.

Change categories

• Model or runtime behavior
• Dependencies
• CI or developer tooling

Validation

Commands and Results

python -m tools.ci pipeline selective-e2e, using the original manifest selections:

Family	Original cases	Result
LFM2	lfm2-1.2b-greedy and five regression cases	6 passed
Llama	nemotron-nano-4b, tinyllama-1.1b	2 passed on the current-main publishing candidate
Nemotron Speech	nemotron-3.5-asr-streaming-0.6b	Passed
PersonaPlex	personaplex-7b, personaplex-7b-l0	2 passed
Qwen MoE	qwen3-moe-30b-a3b, qwen3-moe-tiny-random	2 passed
ELF	Original DE-EN, OWT and XSum L0 cases	3 passed
Wan2.2	wan22-ti2v-5b, wan22-ti2v-5b-l0	2 passed (official reference and original invariant-only L0)
Flux	flux-2-dev, flux-2-dev-fp8	2 passed with the same controlled 240 GiB scratch limit

python3 -m tools.community_ci source-quality --base 74fe95229178d9e5980de30c024150a3a4680399: passed, including 172 tests on the publishing checkout.

Comparator regression plus existing reference/reporting tests: 36 passed. The same new regression tests against the old comparator reproduce six failures.

Hardware, Environment, and Revisions

Local Linux/aarch64 GPU validation used TensorRT 11.1.0.106 and Python 3.12, with the original checkpoint and precision selections. Publishing candidate: 135998f8936d17a66e6d38f3ce55e7401d0be688. Wan2.2 validation runs on this candidate; the other tested family implementations are byte-identical to their validated versions.

Not Run / Remaining Gaps

A complete final Source/tooling Nightly rerun has not completed; these local results do not establish full Nightly readiness.

Notes For Future Readers

Main already provides Llama's EOS support; this PR adds its template fallback. PersonaPlex needs tooling support for dependency-build.json before its opt-in is enabled.

One ELF qualification comparison measured about 35% higher sampling time and 25% higher decode time. This is a numerical-stability tradeoff, not a performance benchmark. Temporary cleanup retains call-failed directories, but not necessarily directories from later teardown/exit failures; exported evidence remains available.

Risk level

• Medium

Several family outputs and dependency builds change; broader performance and full-suite qualification remain outstanding.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

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