Mark McLoughlin

Results 63 comments of Mark McLoughlin

I drafted some thoughts, and found it super difficult to clarify this in words or ASCII art so let's try this ... does this match your thinking? ![vLLM Interval Metrics...

In the "preempted prefill" case, I had imagined the queued interval to be up until the final `SCHEDULED` event ... nothing useful happened with the request, its waiting to be...

@robertgshaw2-redhat sorry, I added another couple to this PR since they were all originally added by the same PR

> Add conditional logging prefixing based on `VLLM_ENABLE_PREFIX_PROC_LOGS` environment variable. I'd be strongly in favor of using an explicit command-line argument and config setting for this, e.g. `--enable-log-prefix` like `--enable-log-requests`...

See also #28582 adding an env var to disable `decorate_logs()` As per #25700 I think we should add a command line parameter instead of an env var

> Main reason is to avoid adding dynamo (or pieces of dynamo) as a dependency to vLLM. For our logger, we would need to add in a rust based publisher...

> @markmc I've changed the publishing to hook into a similar point as the KV Connector API. It follows the thought process I laid out in the RFC. [#16669 (comment)](https://github.com/vllm-project/vllm/issues/16669#issuecomment-2819245423)...

I played around with this a little and added an example subscriber. Feel free to pull in if you think it's good to include :+1: ``` $ git fetch https://github.com/markmc/vllm.git...

> > FYI a start has been made here [#20504](https://github.com/vllm-project/vllm/pull/20504) > > That PR (which was merged to `main` on [7/9/2025](https://github.com/vllm-project/vllm/pull/20504#event-18495144925)) Let's call this issue done and ... > explicitly...