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 6, 07:00 UTC
[ROCm][CI] Mark gemma3 as large GPU test to avoid OOM on MI250 (#37610)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Mar 20, 2026 at 23:57 UTC
0d50fa1db616c83ba3c67f506385fe373f585084
2 files changed
+19
-15
.buildkite/test-amd.yaml
+15
-14
@@ -39,8 +39,7 @@
39
#####################################################################################################################################
40
# #
41
# IMPORTANT: #
42
-# * Currently AMD CI has MI300 agents, MI325 agents, and MI355 agents. Of those, AMD is using mostly MI325 and MI355. AMD team #
43
-# is actively working on enabling more MI300 machines. All upcoming feature improvements are tracked in: #
42
+# * Currently AMD CI has MI250 agents, MI325 agents, and MI355 agents. All upcoming feature improvements are tracked in: #
43
# https://github.com/vllm-project/vllm/issues/34994 #
44
# #
45
#-----------------------------------------------------------------------------------------------------------------------------------#
@@ -49,13 +48,15 @@
48
# * [Pytorch Nightly Dependency Override Check]: if this test fails, it means the nightly torch version is not compatible with #
49
# some of the dependencies. Please check the error message and add the package to #
50
# whitelist in `/vllm/tools/pre_commit/generate_nightly_torch_test.py`. #
52
-# * [Entrypoints Integration Test (LLM)]: #
51
+# * [Entrypoints Integration (LLM)]: #
52
# - {`pytest -v -s entrypoints/llm/test_generate.py`}: It needs a clean process #
53
# - {`pytest -v -s entrypoints/offline_mode`}: Needs to avoid interference with other tests #
55
-# * [V1 Test e2e + engine]: The test uses 4 GPUs, but we schedule it on 8-GPU machines for stability. See discussion here: #
56
-# https://github.com/vllm-project/vllm/pull/31040 #
57
-# * [V1 others]: #
58
-# - Split the tests to avoid interference #
54
+# * [Engine / Engine (1 GPU) / e2e Scheduling / e2e Core / V1 e2e / Spec Decode / V1 Sample + Logits / V1 Core + KV + Metrics]: #
55
+# - Previously a single "V1 Test e2e + engine" step, now split across multiple groups. #
56
+# - V1 e2e (2/4 GPUs) uses 4 GPUs but is scheduled on 8-GPU machines for stability. See: #
57
+# https://github.com/vllm-project/vllm/pull/31040 #
58
+# * [V1 Sample + Logits / V1 Core + KV + Metrics / V1 others (CPU)]: #
59
+# - Previously a single "V1 others" step, now split to avoid interference. #
60
# - Integration test for streaming correctness (requires special branch for __harness__ lib). #
61
# * [V1 others (CPU)]: Split the tests to avoid interference #
62
# * [PyTorch Compilation Unit Tests]: Run unit tests defined directly under `compile/`, not including subdirectories, which #
@@ -83,9 +84,9 @@
84
# run plamo2 model in vLLM. #
85
# * [Language Models Test (Extended Generation)]: Install fast path packages for testing against transformers (mamba, conv1d) #
86
# and to run plamo2 model in vLLM. #
86
-# * [Multi-Modal Models (Standard)]: #
87
+# * [Multi-Modal Models (Standard) 1-4]: #
88
# - Do NOT remove `VLLM_WORKER_MULTIPROC_METHOD=spawn` setting as ROCm requires this for certain models to function. #
88
-# * [Transformers Nightly Models Test]: Whisper needs `VLLM_WORKER_MULTIPROC_METHOD=spawn` to avoid deadlock. #
89
+# * [Transformers Nightly Models]: Whisper needs `VLLM_WORKER_MULTIPROC_METHOD=spawn` to avoid deadlock. #
90
# * [Plugin Tests (2 GPUs)]: #
91
# - {`pytest -v -s entrypoints/openai/test_oot_registration.py`}: It needs a clean process #
92
# - {`pytest -v -s models/test_oot_registration.py`}: It needs a clean process #
@@ -94,11 +95,11 @@
95
# - There is some Tensor Parallelism related processing logic in LoRA that requires multi-GPU testing for validation. #
96
# - {`pytest -v -s -x lora/test_gptoss_tp.py`}: Disabled for now because MXFP4 backend on non-cuda platform doesn't support #
97
# LoRA yet. #
97
-# * [Distributed Tests (GPU_TAG)]: Don't test llama model here, it seems hf implementation is buggy. See: #
98
-# https://github.com/vllm-project/vllm/pull/5689 #
99
-# * [Distributed Tests (GPU_TAG)]: Some old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293 in #
100
-# favor of new tests in fusions_e2e. We avoid replicating the new jobs in #
101
-# this file as it's deprecated. #
98
+# * [Distributed Tests (NxGPUs)(HW-TAG)]: Don't test llama model here, it seems hf implementation is buggy. See: #
99
+# https://github.com/vllm-project/vllm/pull/5689 #
100
+# * [Distributed Tests (NxGPUs)(HW-TAG)]: Some old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293 #
101
+# in favor of new tests in fusions_e2e. We avoid replicating the new jobs in #
102
+# this file as it's deprecated. #
103
# #
104
#####################################################################################################################################
105
tests/models/multimodal/generation/test_common.py
+4
-1
@@ -220,7 +220,10 @@ VLM_TEST_SETTINGS = {
220
vllm_runner_kwargs={
221
"model_impl": "transformers",
222
},
223
- marks=[pytest.mark.core_model],
223
+ marks=[
224
+ pytest.mark.core_model,
225
+ *([large_gpu_mark(min_gb=80)] if current_platform.is_rocm() else []),
226
+ ],
227
),
228
"idefics3-transformers": VLMTestInfo(
229
models=["HuggingFaceTB/SmolVLM-256M-Instruct"],