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.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WgBfkm4rCmtd1T4vHWjEPx
This commit is contained in:
2026-10-15 09:53:38 +02:00
co-authored by Claude Opus 5.5
parent 0d382edcbf
commit 78792a95bb
5 changed files with 69 additions and 10 deletions
+1 -1
View File
@@ -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,