feat(vllm): capture generation artifacts (5/6)

[biswapanda]

Purpose

Add a bounded vLLM producer session that captures requested generation data, validates capacity before inference, encodes generation_artifact_v1, and delivers the resulting opaque object through either storage target from the lower layers.

What changes

• Adds VllmGenerationArtifactSession with explicit admission, per-delta capture, finalization, and release.
• Supports the vLLM content names moe_routes and selected_logprobs; unknown content names fail fast.
• Requires format generation_artifact_v1, codec zstd, and one generated choice.
• Requires Dynamo --enable-rl and vLLM --enable-return-routed-experts when moe_routes is requested.
• Resolves router IDs and expert counts from model configuration and rejects missing or conflicting model metadata.
• Supports nonzero routed_experts_prompt_start and validates the returned route tensor against the effective sequence range.
• Retains prompt/completion token deltas and selected logprobs within admitted token bounds, while retaining only the latest full routed-expert tensor.
• Produces a ready receipt with format, sorted contents, byte count, SHA-256, and opaque object ID.

Admission and resource bounds

Admission happens before generation and estimates the worst-case canonical payload from prompt length, max_tokens, model router count, experts per token, manifest overhead, and target byte limit. It rejects the request before inference if the decoded or conservatively bounded encoded object cannot fit.

• Decoded bytes default to a hard 64 MiB ceiling and can be lowered with DYN_GENERATION_ARTIFACT_MAX_DECODED_BYTES.
• Encoding/upload concurrency defaults to 2 per event loop and is controlled by DYN_GENERATION_ARTIFACT_PIPELINE_CONCURRENCY.
• Prompt, completion, logprob, route, router-count, manifest, and target limits are revalidated as data arrives and again before encoding.
• release() clears captured tokens, routes, counters, and admission state.

Cancellation semantics

Encoding runs off the event loop and delivery runs behind a bounded semaphore. If the caller is cancelled before an operation starts, the queued task is cancelled. If cancellation arrives after encoding/upload starts, the operation continues in the background and retains its semaphore slot until completion, avoiding an in-process ambiguous half-write and unbounded replacement work. A worker-process crash still requires external object-store reconciliation and is not solved in this PR.

Architecture boundary

This is the vLLM schema/capture layer. The common storage writer still receives only bytes plus a target and remains unaware of routed experts or training experiences. Handler/frontend wiring is intentionally deferred to #14978.

Reviewer guide

• components/src/dynamo/vllm/generation_artifact.py: admission, capture, model-layout validation, finalization, and cancellation-safe delivery.
• components/src/dynamo/vllm/tests/test_generation_artifact_capture.py: bounds, alignment, cancellation, concurrency, and receipt behavior.
• components/src/dynamo/common/generation_artifact_storage.py: side-effect-free target validation used during admission.

Validation

• Format and capture suites: 62 passed.
• Storage suite: 48 passed.
• Local MinIO E2E: 2 passed.
• Independent correctness/security review completed with no remaining actionable findings at the time of publication.

PR stack

Review and merge bottom-up. Each PR is based on the preceding stack branch, so its diff contains only that layer. The original umbrella #14385 remains open and unchanged as a reference.

Order	PR	Layer
1	#14977	Protocol and opaque receipt projection
2	#14979	vLLM binary format
3	#14980	Presigned HTTP PUT delivery
4	#14976	Managed fsspec storage
5	#14975	Bounded vLLM capture
6	#14978	vLLM and frontend integration

[dynamo-review-agent]

The retained route-tensor limit is based on the source tensor's assumed int64 representation, while admission budgets routes using their compact wire dtype. For models with many routers and small expert IDs, a valid int64 vLLM tensor can therefore retain up to roughly 8x the configured decoded-byte limit until finalization, despite the request being admitted under that limit. Canonicalize the tensor before retaining it and enforce the admitted bound on that canonical representation.

🤖 AI Fix

Convert routed_experts to its validated canonical integer dtype before storing it and compare the canonical byte count against the admission budget.