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 11, 00:00 UTC
[ROCm][CI] Set VLLM_FLOAT32_MATMUL_PRECISION="tf32" For terratorch Tests In AMD CI (#31242)
Signed-off-by: Micah Williamson <micah.williamson@amd.com>
Micah Williamson committed
Dec 23, 2025 at 21:21 UTC
3ce791ac77ab9d1eafe9c73190fb3e0a687c908a
2 files changed
+12
-4
.buildkite/test-amd.yaml
+11
-3
@@ -162,7 +162,10 @@ steps:
162
- tests/entrypoints/test_chat_utils
163
commands:
164
- export VLLM_WORKER_MULTIPROC_METHOD=spawn
165
- - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/
165
+ - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/test_vision_embeds.py
166
+ # Need tf32 to avoid conflicting precision issue with terratorch on ROCm.
167
+ # TODO: Remove after next torch update
168
+ - VLLM_FLOAT32_MATMUL_PRECISION="tf32" pytest -v -s entrypoints/openai/test_vision_embeds.py
169
- pytest -v -s entrypoints/test_chat_utils.py
170
171
- label: Entrypoints Integration Test (API Server 2)
@@ -979,7 +982,10 @@ steps:
982
- export MIOPEN_DEBUG_CONV_GEMM=0
983
- pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
984
- pip freeze | grep -E 'torch'
982
- - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing
985
+ - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing --ignore models/multimodal/pooling/test_prithvi_mae.py
986
+ # Need tf32 to avoid conflicting precision issue with terratorch on ROCm.
987
+ # TODO: Remove after next torch update
988
+ - VLLM_FLOAT32_MATMUL_PRECISION="tf32" pytest -v -s models/multimodal/pooling/test_prithvi_mae.py -m core_model
989
- cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model # Otherwise, mp_method="spawn" doesn't work
990
991
- label: Multi-Modal Accuracy Eval (Small Models) # 5min
@@ -1341,7 +1347,9 @@ steps:
1347
# end platform plugin tests
1348
# begin io_processor plugins test, all the code in between uses the prithvi_io_processor plugin
1349
- pip install -e ./plugins/prithvi_io_processor_plugin
1344
- - pytest -v -s plugins_tests/test_io_processor_plugins.py
1350
+ # Need tf32 to avoid conflicting precision issue with terratorch on ROCm.
1351
+ # TODO: Remove after next torch update
1352
+ - VLLM_FLOAT32_MATMUL_PRECISION="tf32" pytest -v -s plugins_tests/test_io_processor_plugins.py
1353
- pip uninstall prithvi_io_processor_plugin -y
1354
# end io_processor plugins test
1355
# begin stat_logger plugins test
tests/models/test_terratorch.py
+1
-1
@@ -38,7 +38,7 @@ def test_inference(
38
max_num_seqs=32,
39
default_torch_num_threads=1,
40
) as vllm_model:
41
- vllm_output = vllm_model.llm.encode(prompt)
41
+ vllm_output = vllm_model.llm.encode(prompt, pooling_task="plugin")
42
assert torch.equal(
43
torch.isnan(vllm_output[0].outputs.data).any(), torch.tensor(False)
44
)