[CICD] : iluvatar adapt Iluvatar BI-V150 to the vLLM 0.24 line - #526
Conversation
| # existing Qwen3.6 cases. | ||
| container_volumes: | ||
| - /dev:/dev | ||
| - /tmp:/tmp |
There was a problem hiding this comment.
why mount /tmp from the host?
There was a problem hiding this comment.
I revalidated the container without the host /tmp bind mount. Device detection, tensor smoke, and plugin setup/import all passed, so the mount has been removed.
| - /dev:/dev | ||
| - /tmp:/tmp | ||
| - /lib/modules:/lib/modules | ||
| - /sys:/sys |
There was a problem hiding this comment.
why we mount /sys from the host?
There was a problem hiding this comment.
The redundant mount has been removed.
| --shm-size=64g | ||
| --env GEMS_VENDOR=iluvatar | ||
| --env VLLM_PLUGINS=fl | ||
| --env CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 |
There was a problem hiding this comment.
Is this necessary? can it be removed?
There was a problem hiding this comment.
This runner exposes 16 BI-V150 devices, while the current CI cases and TP8 validation are intended to use 8 devices. The variable limits the job to devices 0-7 and prevents it from consuming the full runner. I believe it should be kept unless device isolation is already enforced at the runner level.
| tests: | ||
| e2e: | ||
| inference: | ||
| qwen3_6: ["27b_tp4_eager"] |
There was a problem hiding this comment.
we need to test graph mode.
| inference: | ||
| qwen3_6: ["27b_tp4_eager"] | ||
| serving: | ||
| qwen3_6: ["35b_a3b_tp4_eager"] |
There was a problem hiding this comment.
we need to test graph mode.
| # CoreX does not support torch.cuda graph capture. The two capture | ||
| # cases abort the shared pytest process and invalidate later results. | ||
| exclude: | ||
| - compilation/test_graph_capture.py |
There was a problem hiding this comment.
need to confirm the conclusion.
There was a problem hiding this comment.
why need to exclude those test cases?
|
The iluvatar workflow didn't run. You may need to modify workflows/ci.yml to enable test on iluvatar. |
4eb7bbe to
6a5944e
Compare
| # CoreX 4.5 returns ixErrorNotSupported from CUDA graph capture_begin. | ||
| # These modules require graph support during import or execution. | ||
| exclude: | ||
| - worker/test_model_runner.py |
There was a problem hiding this comment.
why we need to exclude those test cases?
|
/rerun-failed-ci |
|
🚀 Re-triggered failed CI jobs for this PR (requested by @hooray03). |
6a5944e to
5cdb292
Compare
2 participants
Summary
Adapt Iluvatar BI-V150 CI support to the vLLM 0.24 line.
Changes
device_overrides.No operator or Python implementation files are changed.