A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[CI/Build][Bugfix] Fix Qwen VL tests on CPU (#23818)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Li, Jiang committed
Aug 28, 2025 at 19:57 UTC
67cee40da035b7478483c76dfbe0bfc321c3822f
2 files changed
+14
-14
.buildkite/scripts/hardware_ci/run-cpu-test.sh
+10
-10
@@ -49,23 +49,23 @@ function cpu_tests() {
49
# Run kernel tests
50
docker exec cpu-test-"$NUMA_NODE" bash -c "
51
set -e
52
- pytest -v -s tests/kernels/test_onednn.py"
52
+ pytest -x -v -s tests/kernels/test_onednn.py"
53
54
# Run basic model test
55
docker exec cpu-test-"$NUMA_NODE" bash -c "
56
set -e
57
# Note: disable until supports V1
58
- # pytest -v -s tests/kernels/attention/test_cache.py -m cpu_model
59
- # pytest -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
58
+ # pytest -x -v -s tests/kernels/attention/test_cache.py -m cpu_model
59
+ # pytest -x -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
60
61
# Note: disable Bart until supports V1
62
- pytest -v -s tests/models/language/generation -m cpu_model \
62
+ pytest -x -v -s tests/models/language/generation -m cpu_model \
63
--ignore=tests/models/language/generation/test_bart.py
64
- VLLM_CPU_SGL_KERNEL=1 pytest -v -s tests/models/language/generation -m cpu_model \
64
+ VLLM_CPU_SGL_KERNEL=1 pytest -x -v -s tests/models/language/generation -m cpu_model \
65
--ignore=tests/models/language/generation/test_bart.py
66
67
- pytest -v -s tests/models/language/pooling -m cpu_model
68
- pytest -v -s tests/models/multimodal/generation \
67
+ pytest -x -v -s tests/models/language/pooling -m cpu_model
68
+ pytest -x -v -s tests/models/multimodal/generation \
69
--ignore=tests/models/multimodal/generation/test_mllama.py \
70
--ignore=tests/models/multimodal/generation/test_pixtral.py \
71
-m cpu_model"
@@ -73,20 +73,20 @@ function cpu_tests() {
73
# Run compressed-tensor test
74
docker exec cpu-test-"$NUMA_NODE" bash -c "
75
set -e
76
- pytest -s -v \
76
+ pytest -x -s -v \
77
tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs[False-10-32-neuralmagic/Llama-3.2-1B-quantized.w8a8]"
78
79
# Note: disable it until supports V1
80
# Run AWQ test
81
# docker exec cpu-test-"$NUMA_NODE" bash -c "
82
# set -e
83
- # VLLM_USE_V1=0 pytest -s -v \
83
+ # VLLM_USE_V1=0 pytest -x -s -v \
84
# tests/quantization/test_ipex_quant.py"
85
86
# Run multi-lora tests
87
docker exec cpu-test-"$NUMA_NODE" bash -c "
88
set -e
89
- pytest -s -v \
89
+ pytest -x -s -v \
90
tests/lora/test_qwen2vl.py"
91
92
# online serving
vllm/model_executor/models/utils.py
+4
-4
@@ -507,10 +507,10 @@ def merge_multimodal_embeddings(
507
This updates ``inputs_embeds`` in place.
508
"""
509
if isinstance(placeholder_token_id, list):
510
- placeholder_token_id = torch.tensor(placeholder_token_id,
511
- pin_memory=True).to(
512
- device=input_ids.device,
513
- non_blocking=True)
510
+ placeholder_token_id = torch.tensor(
511
+ placeholder_token_id,
512
+ pin_memory=is_pin_memory_available()).to(device=input_ids.device,
513
+ non_blocking=True)
514
return _merge_multimodal_embeddings(
515
inputs_embeds,
516
torch.isin(input_ids, placeholder_token_id),