[Kurcide]

The native DeepSeek starter now runs on a TensorFold 1.2.0 bundle using the pinned LibertAIDAI EXL3 base. The profile payload is unchanged. The wizard validates the operator and base-model contract, stages the large profile without loading it into memory, and preserves speculative decoding.

The runtime separates stock and profile caches, keeps profile session snapshots in RAM, and validates follower profiles. Bundles include the matching engine source, auxiliary kernel source, and license notices. The compatibility docs distinguish this path from the official-FP8 vLLM path and record TensorFold's existing JSON-schema limitation.

Validation:

• 780 GPU comparisons across 60 native-operator cases, including mixed rows and changing graph masks.
• 15 live protocol checks, separate stock/profile RAM reuse, active DSpark, two cancelled streams, and clean shutdown of all four ranks.
• Real four-rank wizard preparation and a 1.09 GB profile transfer with bounded memory.
• 779 core tests and 69 focused setup, profile, bundle and catalog checks passed locally; the complete suite also runs in CI.

Detailed results are in benchmarks/results/deepseek-tensorfold-native-qualification-20261101.json. These checks do not establish a refusal rate, a throughput benchmark, or full-model equivalence between different engines.

Submitted on behalf of Aevonix Inc. Third-party runtime sources retain their supplied licenses and notices.

• I agree to the FlashAlign contribution terms version 1.0 for this contribution and have authority to grant the rights described there.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-01T03:53:47.102137Z	2307689	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.