You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Audit every current Python-wrapper responsibility and assign each a destination — unified backend, sidecar, or upstream engine — and enumerate the upstream API gaps each engine must close before the sidecar can claim feature parity.
Motivation
Parent #10835 gates sidecar graduation on parity gaps being identified and closed for the target workload. This DEP is that inventory and gate.
Proposal
To be detailed. Scope: per-engine (vLLM / SGLang / TRT-LLM) mapping tables and upstream API-gap lists, sourced from the Dynamo Sidecar Python Wrapper Functionality Report (see #10835 references). Representative parity areas (the report is the complete source): disaggregation handoff, KV events, load/capacity, logprobs / guided decoding / structured output, multimodal / embeddings, LoRA / weight-update / memory controls, diffusion, snapshot/checkpoint, and typed migration errors.
@tanmayv25 I'd like to help drive this DEP. I can start with the vLLM feature-parity inventory: audit the current Python-wrapper responsibilities, map each one to unified backend / sidecar / upstream engine, and identify the remaining upstream API gaps against the #10838 contract.
If that scope sounds good, I can post the initial vLLM mapping table here first and then extend the same structure to the other engines.
Activity
0z5a commentedon Sep 18, 2026
@tanmayv25 I'd like to help drive this DEP. I can start with the vLLM feature-parity inventory: audit the current Python-wrapper responsibilities, map each one to unified backend / sidecar / upstream engine, and identify the remaining upstream API gaps against the #10838 contract.
If that scope sounds good, I can post the initial vLLM mapping table here first and then extend the same structure to the other engines.