[CI][ROCm] Expand weight loading test coverage on AMD and cap its KV cache (#54037)

Signed-off-by: Stefan Koncarevic <Stefan.Koncarevic@amd.com>
This commit is contained in:
stefankoncarevic
2026-09-01 02:57:22 -05:00
committed by GitHub
parent 63988f3c2d
commit b65af5e339
3 changed files with 21 additions and 3 deletions
@@ -7,6 +7,7 @@ import pytest
import torch
from vllm.platforms import current_platform
from vllm.utils.mem_constants import GiB_bytes
MAX_MODEL_LEN = 1024
MODEL_NAME = os.environ.get(
@@ -41,6 +42,9 @@ def test_weight_loading(vllm_runner):
quantization=None if QUANTIZATION == "None" else QUANTIZATION,
max_model_len=MAX_MODEL_LEN,
tensor_parallel_size=2,
# Generating 20 tokens needs a tiny KV cache, while sizing and
# allocating a full-size one dominates this test on large devices.
kv_cache_memory_bytes=2 * GiB_bytes,
) as model:
output = model.generate_greedy("Hello world!", max_tokens=20)
print(output)