[UPSTREAM] Document PRs on upstream-sync

[elevran]

Changes done on upstream-sync and should move to upstream

• Pod Role: [UPSTREAM-SYNC] Define and set Pod Role for prefill/decode filtering #127 (replaced by alternative: use upstream Pod labels in filter)
• Mutate headers: [UPSTREAM-SYNC] Added support for mutating headers #129 (needed for P/D) (replaced by changes in upstream to move headers to Req.Headers)
• Post response handling: TBD (needed for SessionAffinity, PrefixAware and KVCacheAware)
• Set request id end to end: TBD (needed to correlate across calls for plugins that implement multiple interfaces)
• Allow schedulers to be called with an existing Scheduling context: [UPSTREAM] Break up scheduling to allow calling with an existing context #134 (no longer needed, mutated headers moved to request)
• Enable OpenAI ChatCompletions API: TBD

Do NOT close this issue - it is used to collect pending upstream work

Upstream changes needed (e.g., v1/models scraping) are NOT collected here

[nirrozenbaum]

@elevran can you describe more in detail what is the purpose of this issue?
is the list represents the capabilities we want to contribute to upstream BEFORE summit?
should we include ongoing PRs in upstream that are pending review? (like /v1/models)

[elevran]

It's a book-keeping / visibility issue. We do not know when/if these are merged upstream and there is no commitment to contribute upstream by any specific date.

The goal is to track all PRs done on the upstream-sync branch and then tick them off as they are accepted upstream.
PR's on the upstream are not tracked here as the goal is to have a list of known changes in our branch vs the latest upstream.
We will continue to sync to from upstream to the branch so it is up to date.

[elevran]

We have moved llm-d-inference-scheduler to directly use upstream without an intermediate fork