fix(families): repair Community GPU input contracts

[yifeif-nv]

Background

Community GPU failures exposed inherited family contracts that fail independently of the PR being tested: ConvBERT/DeBERTa bundles omit the serialized tokenizer, DeepSeek-V2 queries Hub metadata during offline testing, and Gemma's size-only budget unit allocates a very large dense array.

Exit Criteria

• Official slow-format tokenizer assets produce bundles consumable by the native tokenizer.
• A staged DeepSeek-V2 checkpoint is consumed without an online Hub request.
• Gemma's original budget assertions execute without a hundred-GiB array allocation.
• Community explicitly reports the two large Gemma cases outside its scope while preserving their existing premerge declarations and criteria.

Implementation

Each repair remains family-owned and is split into its own commit:

• Serialize ConvBERT and DeBERTa fast tokenizers directly into the bundle, without modifying read-only checkpoint caches or duplicating sections.
• Request local-only DeepSeek-V2 checkpoint resolution.
• Represent Gemma's zero-valued size fixture with a zero-stride array preserving shape, size, dtype, and all existing assertions.
• Set community_gpu: false for Gemma 3 12B/27B only, retaining premerge: true. The generic manifest guard recognizes and validates this boolean metadata.

Change categories

• Model or runtime behavior
• Public API
• ABI
• Bundle or artifact format
• Dependencies
• Documentation only
• CI or developer tooling

Validation

Commands and Results

Published head 53a92f5c6ebc15ba9320f7d6a6afb56c0fd47952, based on main 9b083a7fdac56f9d0f14084e61e2aa9dc0dc8816:

• python -m pytest families/convbert/tests families/deberta/tests families/deepseek_v2/tests families/gemma/tests -q: 94 passed, 15 existing unselected GPU E2Es skipped in a fresh CPU-only repository container.

• Closed-manifest and premerge guards: 2 passed.

• Real native WordPiece/BPE consumers read fixed bundles generated from pinned official tokenizer assets: both passed. Weight/engine generation was stubbed in this targeted asset-contract control; it is not model GPU parity.

• Original DeepSeek/Gemma assertions and numerical criteria were preserved; Ruff and git diff --check passed.

• Stable CPU run 37753257071: 1,534 Python tests passed, 2 skipped; 246/246 CTests passed. Completed Dev GPU run 37756059191 tested the same head through merge 8900d3b8086eb8eb675b46a7cfebb1df5c99e010 with CI 6e9cc60beafd1b38d0492144c6aa471837a33135: 15/16 selected cases passed on AWS L4. ConvBERT, DeBERTa, all four selected Gemma cases, and the five baseline families passed. The run remains failed because DeepSeek-V2 reached engine construction and TensorRT rejected addMoE on L4; this is no longer an offline checkpoint-resolution failure. The single VM reached functional readiness and deletion was confirmed by both cleanup paths.

• Unchanged local Blackwell control: E2ERunner(CiContext(repo, env))._run(("deepseek_v2",), ("deepseek-v2-tiny",)): 1 selected E2E passed, 0 failed, 0 selected skipped; 13 family unit tests and 1 CTest passed on GB300. The full checkpoint was pinned to katuni4ka/tiny-random-deepseek-v3@ba144b0d3331a5892aa588d82722d382be2b6e6b; original FP16/TP1/NED <=0.25 criteria were unchanged. This is the existing text-parity contract, not a claim of exact token equality.

Hardware, Environment, and Revisions

Published head 53a92f5c on main 9b083a7f. Local checks used an isolated Linux ARM64 CPU-only container with Python 3.12, TensorRT 11.1.0.106 and Transformers 5.2.0. Native tokenizer controls used pinned official ConvBERT and DeBERTa assets; they did not execute GPU model inference. Completed Dev qualification used Linux x86_64, AWS g6.4xlarge / L4, and TensorRT 11.1.0.106. The unchanged DeepSeek control used Linux ARM64, GB300 (SM10.3), Torch 2.12.0+cu130, Transformers 5.2.0, Hub 1.33.0, NumPy 2.4.6, and tokenizers 0.22.2.

Not Run / Remaining Gaps

The AWS Dev run is not green: DeepSeek-V2 native MoE is outside the documented TensorRT 11.1 MoE capability boundary. Its unchanged tiny case passed locally on GB300, but a qualified Community SM10.x/11.x execution route is still pending. Fifteen unselected local CPU-container GPU cases were skipped; targeted tokenizer asset controls stubbed engine and weight generation. Gemma 12B/27B remain explicitly outside Community coverage and are not qualified by these runs. Other checkpoints, TP modes, Internal/Nightly qualification, and performance are not covered by this evidence.

Contributor Self-Review

• I have completed a self-review of this change.

Notes For Future Readers

Community routing requires the generic Dev runner support in #1610. It does not remove the two large cases from Internal/Nightly qualification, and deferral is not a passing result for those cases. Review the family serialization/offline fixes before the fixture and routing changes. This PR does not claim that every failure seen in the broad #1053 run was introduced by that PR or is resolved here.

Risk level

• Low
• Medium
• High

The change affects GPU validation or artifact contracts and requires the recorded target-platform qualification; existing numerical criteria and explicit failure gates are retained.

[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

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

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