Loading...

XML

Word

Printable

Type: Bug
Resolution: Unresolved
Priority: Undefined
Fix Version/s: None
Affects Version/s: 4.20
Component/s: Node / Instaslice-operator
Labels:
- triaged

Activity Type:
Quality / Stability / Reliability
Blocked:
False
Blocked Reason:

Hide

None

Show
None
Story Points:
None
Severity:
None
Regression:
None

Target Backport Versions:
None
Target Version:
None
Release Blocker:
None
Sprint:
OCP Node Sprint 275 (green)
sprint_count:
1

SFDC Cases Counter:
SFDC Cases Open:
SFDC Cases Links:

PX Impact Score:

Release Note Status:
None
Release Note Type:
None
Release Note Text:
None

Escape Reason:
Escape Impact:
Corrective Measures:
SDLC stage when should've been found:

Description of problem:

    When using the latest vllm container image, the model fails to serve due to their backend library (triton) not recognizing CUDA_VISIBLE_DEVICES set to the MIG UUID

Version-Release number of selected component (if applicable):

How reproducible:

Steps to Reproduce:

Pods serving the models using vllm will fail if CUDA_VISIBILE_DEVICES is not represented in the int form due to https://github.com/liammcinroy/vllm/pull/1 
https://github.com/vllm-project/vllm/issues/6551      
https://github.com/vllm-project/vllm/issues/13815

Actual results:

   Expected results:

     vllm to work with DAS Operator

Additional info:

links to

openshift/instaslice-operator#907: OCPBUGS-60686: set CUDA_VISIBLE_DEVICES to device index instead of UUID

Assignee:: Harshal Patil

Reporter:: Harshal Patil

Need Info From:: None

Contributors:: None

QA Contact:: Bhargavi Gudi

Doc Contact:: None

Votes:: 0 Vote for this issue

Watchers:: 6 Start watching this issue

Created:: 2025/08/20 1:44 PM

Updated:: 2025/09/01 10:33 AM