…he reachable ceiling instead of force-appending it (#16256)

Signed-off-by: Dongfeng Yu <dongfengy@nvidia.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

…16398)

Signed-off-by: Yuxian Qiu <142763828+yuxianq@users.noreply.github.com>

…UDA-graph stale-pointer IMA) (#16307)

Signed-off-by: Iman Tabrizian <10105175+tabrizian@users.noreply.github.com>

…-OSS pairing (#16378)

Signed-off-by: tianruih <tianruih@nvidia.com>

…/output lengths (#16538)

Signed-off-by: yufeiwu-nv <230315618+yufeiwu-nv@users.noreply.github.com>

* Why?

While investigating NVBug 6336747, we found a cluster epilog could
remove PyTorch shared-memory files belonging to another same-account
job on the node.

The resulting worker abort exposed a separate API bug: when no terminal
response arrived, we could could still wait on the result indefinitely
because it ignored the timeout.

The cluster cleanup issue is being fixed independently. Honoring the
existing timeout contract gives callers a reliable bound when an
executor stops producing responses.

* What?

Apply timeout as a total budget across all response steps, measured
with a monotonic clock. Pass the remaining budget to both queue
implementations and raise `TimeoutError` when it expires. Preserve
nonblocking `timeout=0 `behavior for responses already queued, and add
focused coverage for these paths.

Signed-off-by: William Zhang <133824995+2ez4bz@users.noreply.github.com>

…in_adp and is_all_greedy_sample`, keep… (#16440)

Signed-off-by: ZhaoyangWang <zhaoyangw@nvidia.com>

Signed-off-by: yiqingy <yiqingy@tensorrt-llm-infra-debug-vm-01.nvidia.com>
Co-authored-by: yiqingy <yiqingy@tensorrt-llm-infra-debug-vm-01.nvidia.com>