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 13, 17:00 UTC
[CI/Build] Fix `test_defaults_with_usage_context` in AMD CI (#27926)
Signed-off-by: zhewenli <zhewenli@meta.com>
Zhewen Li committed
Nov 5, 2025 at 15:40 UTC
0b8e871e5ef2256d3e62a9a4cb111522c5c93372
1 file changed
+4
-2
tests/v1/engine/test_engine_args.py
+4
-2
@@ -53,10 +53,12 @@ def test_defaults_with_usage_context():
53
vllm_config: VllmConfig = engine_args.create_engine_config(UsageContext.LLM_CLASS)
54
55
from vllm.platforms import current_platform
56
+ from vllm.utils.mem_constants import GiB_bytes
57
58
+ device_memory = current_platform.get_device_total_memory()
59
device_name = current_platform.get_device_name().lower()
58
- if "h100" in device_name or "h200" in device_name:
59
- # For H100 and H200, we use larger default values.
60
+ if device_memory >= 70 * GiB_bytes and "a100" not in device_name:
61
+ # For GPUs like H100, H200, and MI300x with >= 70GB memory
62
default_llm_tokens = 16384
63
default_server_tokens = 8192
64
default_max_num_seqs = 1024