Conversation
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
|
/tag-and-rerun-ci |
test_intel_xpu_backend.py failure (separate from this PR)The OCR test now passes. The XPU backend benchmark fails with UR_RESULT_ERROR_OUT_OF_DEVICE_MEMORY while loading meta-llama/Llama-3.2-1B after test_deepseek_ocr.py has already used ~6GB on the GPU. That looks like XPU memory not being fully reclaimed between test files, not a bug in the image URL change. Fixing it probably needs CI or harness changes (ordering, delay, or explicit cache cleanup), not this diff. So I think after this PR merged,I'll try to solve backend.py issues. I think it is not quite easy solve like OCR test. Do you think it is good approach? |
|
@airMeng @mingfeima @Kangyan-Zhou I reordered the XPU suite so lighter tests run first. If that fixes CI, it’s the simplest fix. Long term, if we still want strict alphabetical order (as the comment suggests), we’ll need another approach—e.g. isolating runs so each test starts from a clean device memory state. # Add Intel XPU tests
-# NOTE: please sort the test cases alphabetically by the test file name
+# NOTE: Intentionally NOT alphabetical. Lighter benchmarks run first because
+# heavy models (e.g. DeepSeek-OCR ~6GB) can leave XPU device memory unreclaimed,
+# causing OOM for subsequent tests on memory-constrained devices.See: run_suite.py in PR |
|
/rerun-failed-ci |
@airMeng |
Motivation
The
test_deepseek_ocr.pytest in the XPU CI suite (per-commit-xpu) is consistently failing with aNon-base64 digit founderror. The server's image loader cannot resolve the relative file path../../examples/assets/example_image.pngpassed asimage_data, causing it to fall through to base64 decoding and fail.Related Link: https://github.com/sgl-project/sglang/actions/runs/23862150941/job/69572094739?pr=20501
Modifications
../../examples/assets/example_image.pngwithDEFAULT_IMAGE_URL(a GitHub raw URL) intest/srt/xpu/test_deepseek_ocr.py, consistent with how other VLM tests handle image input.DEFAULT_IMAGE_URLto the imports fromsglang.test.test_utils.Accuracy Tests
N/A — This change only affects the test file, not model outputs.
Speed Tests and Profiling
N/A — No impact on inference speed.
Checklist