[AINode] Replace DiT-derived Sundial conditioning head

[CRZbulabula]

Description

This PR resolves the Sundial provenance issue identified in the review of #18611. The prior conditioning head was explicitly identified as adopted from DiT, whose CC-BY-NC-4.0 license is not suitable for Apache source distributions.

The DiT-derived residual and output-layer implementation is replaced with an independently written PyTorch implementation based on Sundial's required tensor and checkpoint contract. The replacement:

• keeps the existing conditional residual and output-projection behavior;
• preserves all published checkpoint parameter names and tensor shapes;
• preserves zero initialization and strict state-dict loading; and
• removes the DiT-derived implementation and attribution from the source tree.

No user-facing behavior change is intended. Class names and computation are expressed independently, while nested module names remain stable solely for checkpoint compatibility.

Focused tests cover the checkpoint key/shape contract, strict state-dict loading, output shapes, gradient propagation, zero-initialized output, and sampling shape. A random-weight comparison against the previous implementation also produced identical output with a maximum absolute delta of 0.0.

Verification

• PYTHONPATH=. python3 -m unittest discover -s tests -p "test_*.py" -v (7 tests passed)
• black --check . with Black 25.1.0
• isort --check-only --profile black . with isort 6.0.1
• python3 -m compileall -q iotdb-core/ainode/iotdb iotdb-core/ainode/tests
• mvn -q -P with-ainode apache-rat:check -pl iotdb-core/ainode -DskipTests

---

This PR has:

• been self-reviewed.
• added unit tests to cover the replacement implementation.
• added comments/docstrings where the compatibility intent is not obvious.

---

Key changed/added classes (or packages if there are too many classes) in this PR

• iotdb.ainode.core.model.sundial.flow_loss
• tests.test_sundial_flow_loss

[Copilot]

🟢 Approval recommended

The replacement preserves the existing computation and checkpoint contract with focused automated coverage.

Pull request overview

Replaces DiT-derived Sundial conditioning components with independently implemented, checkpoint-compatible PyTorch modules.

Changes:

• Reimplements conditional residual and output projection layers.
• Preserves parameter names, initialization, shapes, and behavior.
• Adds focused compatibility and behavior tests.

File summaries

File	Description
iotdb-core/ainode/iotdb/ainode/core/model/sundial/flow_loss.py	Replaces DiT-derived conditioning layers.
iotdb-core/ainode/tests/test_sundial_flow_loss.py	Tests checkpoint compatibility, gradients, initialization, and sampling.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Balanced

---

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