[chaog992]

What this changes

A halt sets closing, so a prompt pass still running stops with error.Cancelled. The lane core's prompt passes end through LaneHost.cancel (zig/src/core/lane_host.zig:473), which fails the request with halt_reason, as its doc says ("what is left finishes failed with it", lane_host.zig:46). A lone driver's prompt pass instead ended as .cancelled with no message (zig/src/core/lane_host_lone.zig:66). It now ends through LaneHost.cancel too: without a halt, a cancel or a stop still ends cancelled; during a halt, the request fails with the reason; other errors are unchanged.

Where it shows: at the drain deadline (TF_DRAIN_S, 20 s by default) or on a second stop signal, the server halts the engine with "server restarting: retry shortly" (zig/src/server/serve.zig:253). Nemotron with its MTP head runs a lone driver and wires that halt, on Metal (zig/src/native/metal.zig:267) and on CUDA (zig/src/native/cuda.zig:128, not with --record); the other hosts with a lone driver wire no halt and are unchanged. A request whose prompt pass was in that driver got HTTP 200 with finish_reason: "cancelled" and empty content (streamed: a last "cancelled" chunk), and /v1/messages and /v1/responses read it as a finished turn (end_turn, completed). It now gets what drain_test pins for the requests a halt ends (zig/src/server/drain_test.zig:288): a 500 with "server restarting: retry shortly", or, streamed, a server_error event. A request still queued at the halt can reach the lone driver too, once the round in flight frees its lane; it now fails with the reason as well.

Receipt

• Environment: base d1018b3 (main, 1.1.1 merged), fix d119bdb. Zig 0.17.0, Linux x86_64 under WSL2, no GPU used, no model.
• Exactness: n/a: no token changes; only how a halted lone prompt pass ends.
• Decode speed: n/a: an error path.
• Prompt speed: n/a: the change runs only after a lone driver's prompt pass has failed.
• Tests run (each exits 0 unless noted):
  • New test in zig/src/core/lane_host_lone_test.zig, on the fake target: the lone driver holds its prompt pass on an event until the halt is requested, so the pass then finds the host closing; the request must end failed with the halt's reason, with no tokens, no lane left and its lane released once. It waits on events, not on time: 20 of 20 runs pass in Debug and in ReleaseSafe.
  • The lane-host tests, zig test -lc -O <mode> --dep lanes --dep sys -Mroot=zig/src/core/engine_api.zig --dep sys -Mlanes=zig/src/core/lanes/lanes.zig -Msys=zig/src/platform/sys.zig: 218/218 in Debug, ReleaseSafe and ReleaseFast (217/217 on the base). zig build test-lane-api-host: 218/218 (217/217); zig build test-lane-halt-host: 14/14 (13/13).
  • zig build test test-kernel-contract test-golden fuzz-cortheon: 226/230 steps, 3035/3051 tests, 14 skipped, 2 failed, exit 1 (3033/3049 on the base, with the same two failures; the suite runs the lane-host tests in Debug and in ReleaseFast, so the new test counts twice). The two are digest_cache_test's and cuda_cost_identity's same-size rewrite tests, whose rewrite usually lands inside one file-time tick on this WSL2 machine (glm/cuda: load and serve nvidia/GLM-5.3-Flash-FP4-Local as stored #1050's receipt reports the second on plain v1.1.1 too). Server goldens 306/319 and 26/26 equal, 0 unexpected, as on the base.
  • Plain zig build: 24/24 steps. tools/zig/lean_check.py: the same 444 problems as the base, exit 1 on both, none in the touched files; tools/zig/lean_new.py against the base: 0 new. zig fmt --check build.zig zig and git diff --check pass.
  • Outside this PR: a probe shows the queued case ending cancelled with no message on the base and failed with the reason with the fix, and a randomized lane-host harness that halts at random points finds requests ending cancelled that nothing cancelled on the base, and none with the fix.
• Failing receipt: the new test on the base's lane_host_lone.zig fails in Debug, ReleaseSafe and ReleaseFast: expected .failed, found .cancelled (zig/src/core/lane_host_lone_test.zig:171).
• Not run, and why: no Mac and no GPU or model, so no Nemotron server drained on Metal or CUDA; the HTTP effect above comes from reading the server code and drain_test. The lane-host tests compile for aarch64-macos with -mcpu apple_m1 --test-no-exec, a compile check only.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off. (Unchanged: the change only decides how a failed lone prompt pass ends.)
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (n/a: no tokens change.)
• No precision traded for speed. If the bits change, the description says which sums change and why. (No arithmetic changes.)
• A new or changed GPU kernel is registered with the kernel contract (docs/kernel-contract.md); each backend's SUMMARY line is in the receipt. (n/a: no kernel changes.)
• Prompt processing is no slower than the last release. (n/a: an error path after a failed prompt pass.)
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (It needs no model or GPU and never skips.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.

[ashhart]

Thank you. This is on main as 010850d, with your authorship: a halt during the lone driver's prompt pass now ends the request through LaneHost.cancel, so it fails with the halt's reason ("server restarting: retry shortly") instead of finishing cancelled with an empty reply, and your test pins the single lane release. It ships with the next release. Closing as landed.

- Synthema, Ash's AI Assistant.