Adding "Basic Models Test" and "Multi-Modal Models Test (Extended) 3" in AMD Pipeline (vllm-project#18106)

Signed-off-by: Alexei V. Ivanov <alexei.ivanov@amd.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>

Author: Alexei-V-Ivanov-AMD

.buildkite/scripts/hardware_ci/run-amd-test.sh

@@ -82,6 +82,14 @@ if [[ $commands == *"pytest -v -s basic_correctness/test_basic_correctness.py"*
   commands=${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"}
 fi
 
+if [[ $commands == *"pytest -v -s models/test_registry.py"* ]]; then
+  commands=${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'"}
+fi
+
+if [[ $commands == *"VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'"* ]]; then
+  commands=${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'"}
+fi
+
 if [[ $commands == *"pytest -v -s compile/test_basic_correctness.py"* ]]; then
   commands=${commands//"pytest -v -s compile/test_basic_correctness.py"/"VLLM_USE_TRITON_FLASH_ATTN=0 pytest -v -s compile/test_basic_correctness.py"}
 fi

.buildkite/test-pipeline.yaml

@@ -455,7 +455,7 @@ steps:
 #####  models test  #####
 
 - label: Basic Models Test # 24min
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
@@ -527,7 +527,7 @@ steps:
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
 
 - label: Multi-Modal Models Test (Extended) 3
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   optional: true
   source_file_dependencies:
   - vllm/
@@ -537,7 +537,7 @@ steps:
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
 
 - label: Quantized Models Test
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - vllm/model_executor/layers/quantization
   - tests/models/quantization

requirements/rocm-test.txt

@@ -22,4 +22,10 @@ decord==0.6.0
 #sentence-transformers # required by entrypoints/openai/test_score.py
 sentence-transformers==3.4.1
 
+# Basic Models Test
+matplotlib==3.10.3
+
+# Multi-Modal Models Test (Extended) 3
+blobfile==3.0.0
+
 

tests/models/test_transformers.py

@@ -2,6 +2,8 @@
 """Test the functionality of the Transformers backend."""
 import pytest
 
+from vllm.platforms import current_platform
+
 from ..conftest import HfRunner, VllmRunner
 from ..utils import multi_gpu_test
 from .utils import check_logprobs_close
@@ -33,6 +35,9 @@ def check_implementation(
     )
 
 
+@pytest.mark.skipif(
+    current_platform.is_rocm(),
+    reason="Llama-3.2-1B-Instruct, Ilama-3.2-1B produce memory access fault.")
 @pytest.mark.parametrize(
     "model,model_impl",
     [
@@ -64,6 +69,9 @@ def test_distributed(
                          "meta-llama/Llama-3.2-1B-Instruct", **kwargs)
 
 
+@pytest.mark.skipif(
+    current_platform.is_rocm(),
+    reason="bitsandbytes quantization is currently not supported in rocm.")
 @pytest.mark.parametrize("model, quantization_kwargs", [
     (
         "meta-llama/Llama-3.2-1B-Instruct",