[CI][MiniCPM-o]MiniCPM-o ACC PERF CI (#5575)

Signed-off-by: R2-Y <ruiruyang2@gmail.com>
Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>
This commit is contained in:
ruirui(rein) yangandamy-why-3459 authored and GitHub committed 2026-08-04 18:46:20 +08:00
1 parent 434c220da1
commit a5eb2a9a87
50 files changed
+1756 -102

No files matched your search

+27
View File
@@ -2164,6 +2164,33 @@ class TestBaseConfigInheritance:
class TestPlatformOverrides:
"""Test platform-specific deploy config overrides."""
def test_minicpmo_4_5_cuda_caps_talker_kv_cache(self):
pipeline = resolve_pipeline_config("minicpmo_4_5")
assert isinstance(pipeline, PipelineConfig)
deploy_path = Path(get_deploy_config_path("minicpmo_4_5.yaml"))
cuda = _apply_platform_overrides(load_deploy_config(deploy_path), platform="cuda")
cuda_stages = merge_pipeline_deploy(pipeline, cuda)
# Talker KV is left to automatic sizing (no hard kv_cache_memory_bytes cap).
assert "kv_cache_memory_bytes" not in cuda_stages[1].yaml_engine_args
# The CUDA memory budget must not leak into the existing NPU profile.
npu = _apply_platform_overrides(load_deploy_config(deploy_path), platform="npu")
npu_stages = merge_pipeline_deploy(pipeline, npu)
assert "kv_cache_memory_bytes" not in npu_stages[1].yaml_engine_args
# Multi-GPU replica profiles place Talker on dedicated devices and
# also leave KV sizing automatic.
for filename in (
"minicpmo_4_5_3gpu_stage1_replicas.yaml",
"minicpmo_4_5_4gpu_stage1_replicas.yaml",
):
replica = _apply_platform_overrides(
load_deploy_config(Path(get_deploy_config_path(filename))), platform="cuda"
)
replica_stages = merge_pipeline_deploy(pipeline, replica)
assert "kv_cache_memory_bytes" not in replica_stages[1].yaml_engine_args
def test_npu_overrides(self):
deploy_path = Path(__file__).parent.parent / "vllm_omni" / "deploy" / "qwen3_omni_moe.yaml"
if not deploy_path.exists():