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, 15:00 UTC
[ROCM][CI] Fix AMD Examples Test Group (#30276)
Signed-off-by: Yida Wu <yida.wu@amd.com> Signed-off-by: Yida <yida.wu@amd.com>
Concurrensee committed
Dec 11, 2025 at 17:03 UTC
2cc5affc388d3d134bacc14f042405ead925531b
3 files changed
+17
-2
.buildkite/test-amd.yaml
+1
-2
@@ -435,7 +435,7 @@ steps:
435
436
- label: Examples Test # 30min
437
timeout_in_minutes: 45
438
- mirror_hardwares: [amdexperimental]
438
+ mirror_hardwares: [amdexperimental, amdproduction]
439
agent_pool: mi325_1
440
# grade: Blocking
441
working_dir: "/vllm-workspace/examples"
@@ -455,7 +455,6 @@ steps:
455
# for multi-modal models
456
- python3 offline_inference/audio_language.py --seed 0
457
- python3 offline_inference/vision_language.py --seed 0
458
- - python3 offline_inference/vision_language_pooling.py --seed 0
458
- python3 offline_inference/vision_language_multi_image.py --seed 0
459
- python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
460
# for pooling models
examples/offline_inference/basic/embed.py
+8
@@ -4,6 +4,9 @@
4
from argparse import Namespace
5
6
from vllm import LLM, EngineArgs
7
+from vllm.attention.backends.registry import AttentionBackendEnum
8
+from vllm.config import AttentionConfig
9
+from vllm.platforms import current_platform
10
from vllm.utils.argparse_utils import FlexibleArgumentParser
11
12
@@ -20,6 +23,11 @@ def parse_args():
23
24
25
def main(args: Namespace):
26
+ if current_platform.is_rocm():
27
+ args.attention_config = AttentionConfig(
28
+ backend=AttentionBackendEnum.FLEX_ATTENTION
29
+ )
30
+
31
# Sample prompts.
32
prompts = [
33
"Hello, my name is",
examples/offline_inference/basic/score.py
+8
@@ -4,6 +4,9 @@
4
from argparse import Namespace
5
6
from vllm import LLM, EngineArgs
7
+from vllm.attention.backends.registry import AttentionBackendEnum
8
+from vllm.config import AttentionConfig
9
+from vllm.platforms import current_platform
10
from vllm.utils.argparse_utils import FlexibleArgumentParser
11
12
@@ -20,6 +23,11 @@ def parse_args():
23
24
25
def main(args: Namespace):
26
+ if current_platform.is_rocm():
27
+ args.attention_config = AttentionConfig(
28
+ backend=AttentionBackendEnum.FLEX_ATTENTION
29
+ )
30
+
31
# Sample prompts.
32
text_1 = "What is the capital of France?"
33
texts_2 = [