feat(ci): dispatch a CI run to a GPU generation

[yueming-yuan]

Chain 4/5 — stacked on b200-ci-baseline-identity.

The second axis. 2/5 gave every CUDA test a hardware=[...] capability; this lets a PR say which of those generations the run should execute on, and routes tests accordingly.

PR labels	Effect
(none)	each test runs once, at its home stage
run-on-hopper	Hopper stages only; Blackwell-only tests do not run
run-on-blackwell	everything that supports Blackwell moves there
both	a portable test runs twice, once per generation
run-ci-blackwell-only	the Blackwell-exclusive tests, across all domains

So run-ci-megatron + run-on-blackwell = every megatron test that can run on Blackwell, on Blackwell, and nowhere else.

Safety

absorb is the valve: only an explicit run-on-* lets a test leave its home stage. Plain PRs, nightly, weekly, release and called workflows all resolve to absorb=False, where the home-stage invariant makes routing unreachable — so their selection is byte-identical to before this commit. A test pins that property across every stage, arch and dispatch combination.

Design notes

Routing is derived, not tabulated. The destination is the smallest stage on the target generation with enough GPUs. A hand-written home → destination map would encode a fleet shape that does not exist yet; this rule gives today's degenerate answer (one Blackwell stage takes everything) and a 1:1 mirror once a second Blackwell host is partitioned, from the same three lines.

No new flag, no workflow edit. A stage belongs to exactly one generation, so --suite already names it. An --arch flag would be a second copy of that fact, free to drift out of step.

One predicate. stage_selection shares run_suite's runs_in_stage; without that a run-on-blackwell PR would still allocate every Hopper runner and then find nothing to run on them.

hardware.py grows num_gpus / runs_on and becomes the single source of truth — CI_SUITES and CUDA_SUITE_RUNS_ON now derive from it.

Comment gateway

LABEL_PATTERN accepts run-on-* so /run-on-blackwell works. This does not widen what a comment may add: allowed_labels stays an exact default-deny allowlist, and the pattern only rejects malformed entries in it.

🤖 Generated with Claude Code

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}