fix(vllm): serialize routed_experts as base64 with start offset

[biswapanda]

Overview:

Make MoE routed-experts capture portable for RL log-prob recomputation: base64-encoded with a start offset so consumers can align the routing to the completion.

Details:

_serialize_routed_experts now:

• encodes the expert-id bytes as base64 (was base85), matching the encoding consumers already decode;
• emits a start field (= SamplingParams.routed_experts_prompt_start, the leading prompt rows vLLM trims) so consumers can slice to the completion;
• keeps shape and dtype so the consumer reads the right element width instead of assuming a fixed type.

The token-generation path forwards start from the request's sampling params (0 when unset).

Where should the reviewer start?

components/src/dynamo/vllm/handlers.py — _serialize_routed_experts and its call site in the token-generation path.

Related Issues

🚫 This PR is NOT linked to an issue:

• Confirmed — no related issue

Summary by CodeRabbit

Release Notes

• Bug Fixes

  • Updated routed-experts serialization encoding for improved compatibility
  • Enhanced routed expert prompt window alignment with start position tracking

• Tests

  • Added validation for routed expert prompt positioning

[coderabbitai]

Walkthrough

Routed-experts serialization is updated to use base64 encoding instead of base85 and extended with an optional start offset parameter derived from sampling_params.routed_experts_prompt_start. The token generation path passes this offset, and tests validate both the encoding change and offset behavior.

Changes

Routed-experts serialization with start offset

Layer / File(s)	Summary
Routed-experts serialization contract and implementation components/src/dynamo/vllm/handlers.py	_serialize_routed_experts signature updated to accept optional start: int = 0 parameter. Serialization implementation switches from base85 to base64 encoding and emits start field in returned metadata dict alongside data, shape, and dtype.
Token generation integration components/src/dynamo/vllm/handlers.py	Token generation call site updated to extract start offset from sampling_params.routed_experts_prompt_start (defaulting to 0 when absent) and pass it to _serialize_routed_experts.
Test updates and new coverage components/src/dynamo/vllm/tests/test_vllm_worker_handler.py	Existing routed-experts test assertion updated to decode base64 instead of base85. New async test verifies routed_experts.start echoes SamplingParams.routed_experts_prompt_start with conditional skip for older vLLM versions lacking the field.

🎯 2 (Simple) | ⏱️ ~10 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 16.67% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately summarizes the main changes: switching routed_experts serialization from base85 to base64 encoding and adding a start offset parameter.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description check	✅ Passed	The pull request description is well-structured and complete, covering all required template sections with meaningful content.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

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

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@components/src/dynamo/vllm/tests/test_vllm_worker_handler.py`:
- Around line 367-370: The test currently blanket-catches all exceptions when
assigning sampling_params.routed_experts_prompt_start which can hide unrelated
regressions; change the broad except Exception to only catch the specific
unsupported-assignment errors (e.g., except (AttributeError, TypeError):) so
only missing/unsupported attribute assignments trigger pytest.skip and any other
exceptions are allowed to surface and fail the test; locate the assignment to
sampling_params.routed_experts_prompt_start in the test_vllm_worker_handler.py
test and restrict the catch accordingly.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: f0747d31-d964-4aaa-a22b-dd232f2f8163

📥 Commits

Reviewing files that changed from the base of the PR and between a1ea8a0 and 47acc25.

📒 Files selected for processing (2)

• components/src/dynamo/vllm/handlers.py
• components/src/dynamo/vllm/tests/test_vllm_worker_handler.py