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 15, 09:45 UTC
Adding "Basic Models Test" and "Multi-Modal Models Test (Extended) 3" in AMD Pipeline (#18106)
Signed-off-by: Alexei V. Ivanov <alexei.ivanov@amd.com> Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Alexei-V-Ivanov-AMD committed
May 15, 2025 at 10:49 UTC
566ec04c3d3fb8e8b0876f04cd3be036d1f8d3ac
4 files changed
+25
-3
.buildkite/scripts/hardware_ci/run-amd-test.sh
+8
@@ -82,6 +82,14 @@ if [[ $commands == *"pytest -v -s basic_correctness/test_basic_correctness.py"*
82
commands=${commands//"pytest -v -s basic_correctness/test_basic_correctness.py"/"VLLM_USE_TRITON_FLASH_ATTN=0 pytest -v -s basic_correctness/test_basic_correctness.py"}
83
fi
84
85
+if [[ $commands == *"pytest -v -s models/test_registry.py"* ]]; then
86
+ commands=${commands//"pytest -v -s models/test_registry.py"/"pytest -v -s models/test_registry.py -k 'not BambaForCausalLM and not GritLM and not Mamba2ForCausalLM and not Zamba2ForCausalLM'"}
87
+fi
88
+
89
+if [[ $commands == *"VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'"* ]]; then
90
+ commands=${commands//"VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'"/"VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2 and not BambaForCausalLM and not Gemma2ForCausalLM and not Grok1ModelForCausalLM and not Zamba2ForCausalLM and not Gemma2Model and not GritLM'"}
91
+fi
92
+
93
if [[ $commands == *"pytest -v -s compile/test_basic_correctness.py"* ]]; then
94
commands=${commands//"pytest -v -s compile/test_basic_correctness.py"/"VLLM_USE_TRITON_FLASH_ATTN=0 pytest -v -s compile/test_basic_correctness.py"}
95
fi
.buildkite/test-pipeline.yaml
+3
-3
@@ -455,7 +455,7 @@ steps:
455
##### models test #####
456
457
- label: Basic Models Test # 24min
458
- mirror_hardwares: [amdexperimental]
458
+ mirror_hardwares: [amdexperimental, amdproduction]
459
torch_nightly: true
460
source_file_dependencies:
461
- vllm/
@@ -527,7 +527,7 @@ steps:
527
- pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
528
529
- label: Multi-Modal Models Test (Extended) 3
530
- mirror_hardwares: [amdexperimental]
530
+ mirror_hardwares: [amdexperimental, amdproduction]
531
optional: true
532
source_file_dependencies:
533
- vllm/
@@ -537,7 +537,7 @@ steps:
537
- pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
538
539
- label: Quantized Models Test
540
- mirror_hardwares: [amdexperimental]
540
+ mirror_hardwares: [amdexperimental, amdproduction]
541
source_file_dependencies:
542
- vllm/model_executor/layers/quantization
543
- tests/models/quantization
requirements/rocm-test.txt
+6
@@ -22,4 +22,10 @@ decord==0.6.0
22
#sentence-transformers # required by entrypoints/openai/test_score.py
23
sentence-transformers==3.4.1
24
25
+# Basic Models Test
26
+matplotlib==3.10.3
27
+
28
+# Multi-Modal Models Test (Extended) 3
29
+blobfile==3.0.0
30
+
31
tests/models/test_transformers.py
+8
@@ -2,6 +2,8 @@
2
"""Test the functionality of the Transformers backend."""
3
import pytest
4
5
+from vllm.platforms import current_platform
6
+
7
from ..conftest import HfRunner, VllmRunner
8
from ..utils import multi_gpu_test
9
from .utils import check_logprobs_close
@@ -33,6 +35,9 @@ def check_implementation(
35
)
36
37
38
+@pytest.mark.skipif(
39
+ current_platform.is_rocm(),
40
+ reason="Llama-3.2-1B-Instruct, Ilama-3.2-1B produce memory access fault.")
41
@pytest.mark.parametrize(
42
"model,model_impl",
43
[
@@ -64,6 +69,9 @@ def test_distributed(
69
"meta-llama/Llama-3.2-1B-Instruct", **kwargs)
70
71
72
+@pytest.mark.skipif(
73
+ current_platform.is_rocm(),
74
+ reason="bitsandbytes quantization is currently not supported in rocm.")
75
@pytest.mark.parametrize("model, quantization_kwargs", [
76
(
77
"meta-llama/Llama-3.2-1B-Instruct",