GPUs follow the engine's own CUDA_VISIBLE_DEVICES

Its entries are the cards present, and a local child's allocation is named through them, so under =2,3 card 0 is device 2. Remote workers keep mapping against their own mask.
2026-10-15 09:53:38 +02:00
parent 290b9fa705
commit f2a823218a
5 changed files with 69 additions and 10 deletions
@@ -979,7 +979,7 @@ class FlowController:
            with placer.claim(  # type: ignore[union-attr]
                wanted, device=device, policy=policy, node=node_id, run=run_id
            ) as (target, allocation):
-                env = derive_env(wanted, allocation)
+                env = derive_env(wanted, allocation, local=target.worker is None)
                if target.worker is None:
                    return pool.for_env(env).run(  # type: ignore[union-attr]
                        owner,