Conversation
The sequential offload conflict check and model offload warning in
`Pipeline.to()` only applied to CUDA and Intel XPU devices. Ascend NPU
is also an accelerator where moving a pipeline to the device conflicts
with offloading — users calling `.to("npu")` should get the same
clear error/warning instead of silently broken behavior.
5 changes, 2 insertions(+), 2 deletions(-) — adds "npu" to the
device_type lists at lines 2691 and 2704 (previously ["cuda", "xpu"]).
|
Hi @li-lizhe, thanks for the PR! It does not appear to link an issue it fixes. If this PR addresses an existing issue, please add a closing keyword (e.g. Please note that PRs without a linked issue are likely to be automatically closed 10 days after this notice. Once the PR links an issue (or gets the |
|
Linked the tracked issue (#14878) in the description, so the reminder above no longer applies. Separately: the workflow runs on this PR are all sitting in |
|
CI has never started on this PR: its workflow runs are stuck in They need a repo-write maintainer to approve them (Checks tab -> "Approve and run"). The description now links #14878, so the auto-close reminder no longer applies. /cc @yiyixuxu @sayakpaul (recent committers to |
This branch has not been deployed
1 participant
The sequential offload conflict check (line 2691) and model offload warning (line 2704) in
Pipeline.to()only applied to CUDA and Intel XPU devices. Ascend NPU is also an accelerator where moving a pipeline to the device conflicts with offloading — users calling.to("npu")should get the same clear error/warning instead of silently broken behavior.Changes:
["cuda", "xpu"]→["cuda", "xpu", "npu"]at two guard points (lines 2691 and 2704)Effect:
"cuda""xpu""npu""cpu""hpu"2 insertions(+), 2 deletions(-) — clean, minimal, device-agnostic.
Verified on Ascend 910B NPU:
device_type='npu'now correctly captured by the guard list.Fixes #14878