[Rust] multimodal generation back-ported on vllm v0.28.0 - #56421

Overview

Add the minimal vLLM 0.28 Rust frontend support needed for Dynamo-backed Qwen3.5 image generation during RL training.

This is not a duplicate of #55047: that PR targets current main and includes broader multimodal work. This branch is based directly on releases/v0.28.0 and keeps only the v0.28 compatibility path required by the tested Dynamo stack.

Changes

• Preserve native Generate sampling semantics across gRPC, including omitted temperature and explicit zero or false values.
• Advertise separate capabilities for native sampling JSON and preprocessed multimodal features.
• Accept image-only engine-ready multimodal features and forward them to TextRequest.mm_features.
• Validate MessagePack structure, inline tensor layout, feature ranges, prompt bounds, and mixed raw/preprocessed inputs.
• Bind cache identity to validated feature content instead of trusting caller-provided identifiers or hashes.
• Report the engine world size required by Dynamo managed-RL discovery.
• Accept both the older short engine-output tuple and vLLM 0.28's current 18-field tuple while discarding unused speculative-decode and routed-expert metadata without allocation.

The change intentionally excludes LoRA handling, audio/video support, chat model-limit refactors, and unrelated code from #55047.

Validation

• cargo fmt --all -- --check
• cargo check -p vllm-server --tests
• cargo test -p vllm-server grpc::convert::tests:: — 22 passed
• cargo test -p vllm-server preprocessed_image — 7 passed
• cargo test -p vllm-server control_reports_server_and_model_info — 1 passed
• cargo test -p vllm-engine-core-client --lib — 97 passed, including both output tuple lengths
• Cross-language MessagePack fixture produced by Dynamo decodes successfully.
• Five-step Qwen3.5-4B multimodal RL run passed on bis-rl-3 with Prime-RL main plus [Core] [Bugfix] Refactor block manager subsystem for better testability #3492, Dynamo [Installation]: #14398 plus the minimal image-only sidecar layer, and this PR: 5/5 orchestrator steps, 5/5 trainer steps, 2/2 trainable rollouts per step, zero rollout errors/cancellations, six NCCL weight receives, and successful pre/post image probes.

AI assistance

AI assistance was used to analyze the existing PRs, implement the compatibility backport, and draft tests. The submitting developer reviewed the changes and validation results.