deps: AiDotNet.Tensors 0.133.2 (one sigmoid on every CPU vendor)

[ooples]

Why

Several PRs have a shard that fails only on Intel runners: RepViTSAM (#2277) and EfficientSAM (#2272, #2269). Bisected on hosted runners, the cause is Tensors' Intel-only table sigmoid. It is accurate, but numerically different from the Padé kernel AMD runs, and AdamW's start-up transient amplifies the difference. AiDotNet.Tensors #1069, released in 0.133.2, puts every vendor on one kernel and also makes sigmoid(±inf) saturate instead of returning NaN.

Changes

• Bump AiDotNet.Tensors and the three AiDotNet.Native.* packages from 0.133.0 to 0.133.2, together.
• 0.133.1 renamed DeferredArrayMaterializer to HostSync, with the same TryMaterialize / Register / Remove. The four sharded-training call sites and CpuOffloadShardingConfigTests follow. The hunks match fix: fused training matches eager - optimizer formulas, persistent fused steps, full trainable set #2267's rename exactly, so they merge cleanly whichever lands first; only the version line will conflict.

Verified locally

• Library and test project build (net10.0).
• CpuOffloadShardingConfigTests: 11/11 passed.
• UnitTests.Optimizers plus the EfficientSAM and RepViTSAM training invariants: 346/346 passed (AMD).

The full matrix runs here, since this is a package bump. The Intel-runner behaviour shows on any shard that lands on an Intel CPU, where these models previously failed.

Overlap: #2267 and #2254 bump to 0.133.1. Once this merges, they can drop their version change and take 0.133.2 from master.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
aidotnet_website	Ready	Preview	Oct 1, 2026 12:11am UTC
aidotnet-playground-api	Ready	Preview	Oct 1, 2026 12:11am UTC

[coderabbitai]

Warning

Review limit reached

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Next included review available in 50 minutes.

Check out review usage here.

View limit details

Limit details: You’ve used the included review currently available.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: Repository: ooples/AiDotNet/.coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: 805f60f7-dcce-4596-88df-62d3ba945fa0

📥 Commits

Reviewing files that changed from the base of the PR and between 1b72f56 and a96ccb6.

📒 Files selected for processing (4)

• Directory.Packages.props
• src/DistributedTraining/ShardedModelBase.cs
• src/DistributedTraining/ShardedOptimizerBase.cs
• tests/AiDotNet.Tests/UnitTests/DistributedTraining/CpuOffloadShardingConfigTests.cs

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

Autopilot is currently an internal CodeRabbit preview.

---

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.}

[github-actions]

Commit messages auto-fixed

One or more commit messages did not follow Conventional Commits, so they were rewritten to comply (subject case, header length ≤ 100, valid type). Each commit and its diff were preserved — no squashing.

The branch was force-pushed with the corrected messages. If you have local work on this branch, run git pull --rebase (or reset to the remote) before pushing again.

[vercel]

Deployment failed for project aidotnet_website with the following error:

Resource is limited - try again in 24 hours (more than 100, code: "api-deployments-free-per-day").

Learn More: https://vercel.com/franklins-projects-02a0b5a0?upgradeToPro=build-rate-limit

[Copilot]

Copilot review overview

🟢 Approval recommended

The lockstep package bump is valid and all affected API references were consistently migrated.

Review effort: Balanced
Findings: None

What changed in this PR

Updates the tensor stack to 0.133.2 for CPU-independent sigmoid behavior and adopts the renamed synchronization API.

Changes:

• Bumps all tensor/native packages in lockstep.
• Replaces DeferredArrayMaterializer calls with HostSync.

File	Description
Directory.Packages.props	Updates tensor packages to 0.133.2.
src/​DistributedTraining/​ShardedOptimizerBase.cs	Migrates optimizer offloading to HostSync.
src/​DistributedTraining/​ShardedModelBase.cs	Migrates model offloading to HostSync.
tests/​AiDotNet.Tests/​UnitTests/​DistributedTraining/​CpuOffloadShardingConfigTests.cs	Updates synchronization tests for the renamed API.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[ooples]

On hold, two failures:

• GraFPrint compiled-vs-eager parity: 0.133.1 made the previously vacuous parity test honest (Tensors a8a201eb), exposing that compiled and eager BatchNorm compute different single-sample statistics. Evidence is posted on fix(batchnorm): normalize a single-sample conv activation over its spatial positions #2272, which is rewriting that path; this PR rebases on it.
• WordCharEmbeddingLayerTests.TapeGradient_ShouldMatchNumericalGradient: intermittent. It also fails on Tensors 0.133.0 (seen on fix(tests): measured training budget for RepViTSAM, and a nightly cross-ISA agreement lane #2277, AMD runner) and is being fixed separately.

[ooples]

Superseded: master is on AiDotNet.Tensors 0.134.0 (bumped via #2276), which contains everything 0.133.2 brought. Merging master here would leave this PR with no changes.