[Edge] A deploy YAML must mean the same thing whichever flag loaded it

`vllm_omni/deploy/edge/qwen3_tts.yaml` carries `orchestrator:
parallel_stage_init: true`, but `_apply_profile_orchestrator_defaults` ran
only on the `deploy_profile` path. Pointing `--deploy-config` at the same file
loaded its stages and silently dropped the block, so the edge deployment
started sequentially:

    --deploy-profile edge                       30.76 s
    --deploy-config .../edge/qwen3_tts.yaml     63.76 s   (same file)

with nothing in the log to say why. The block now applies however the config
was reached. Explicit arguments still win, and the allow-list stays narrow --
a test checks an unrecognised key cannot be injected through it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
tzhouamandClaude Opus 5 committed 2026-09-12 21:29:10 +08:00
1 parent a51461c131
commit a7e6c412ab
2 files changed
+61

No files matched your search

+52
View File
@@ -98,3 +98,55 @@ def test_edge_profile_carries_parallel_stage_init():
assert args["parallel_stage_init"] is True
# unknown block keys are ignored
assert "init_timeout" not in args
@pytest.mark.core_model
@pytest.mark.cpu
def test_orchestrator_block_applies_however_the_config_was_loaded(tmp_path):
"""A deploy YAML must mean the same thing whichever flag loaded it.
Applying the `orchestrator:` block only on the profile path made
`--deploy-config deploy/edge/qwen3_tts.yaml` load the stages but drop
`parallel_stage_init: true`, so an edge deployment started sequentially
(63.8 s against 30.5 s on Qwen3-TTS) with nothing in the log to say why.
"""
from vllm_omni.entrypoints import utils as entry_utils
path = tmp_path / "profile.yaml"
path.write_text(
"pipeline: qwen3_tts\n"
"orchestrator:\n"
" parallel_stage_init: true\n"
" stage_init_timeout: 900\n",
encoding="utf-8",
)
out = entry_utils.apply_deploy_profile("m", {"deploy_config": str(path)})
assert out["parallel_stage_init"] is True
assert out["stage_init_timeout"] == 900
@pytest.mark.core_model
@pytest.mark.cpu
def test_explicit_arguments_still_beat_the_orchestrator_block(tmp_path):
from vllm_omni.entrypoints import utils as entry_utils
path = tmp_path / "profile.yaml"
path.write_text("orchestrator:\n stage_init_timeout: 900\n", encoding="utf-8")
out = entry_utils.apply_deploy_profile(
"m", {"deploy_config": str(path), "stage_init_timeout": 120})
assert out["stage_init_timeout"] == 120
@pytest.mark.core_model
@pytest.mark.cpu
def test_unknown_orchestrator_keys_are_ignored(tmp_path):
"""The block is a narrow allow-list, not a way to inject arbitrary args."""
from vllm_omni.entrypoints import utils as entry_utils
path = tmp_path / "profile.yaml"
path.write_text(
"orchestrator:\n parallel_stage_init: true\n rm_rf_slash: true\n",
encoding="utf-8")
out = entry_utils.apply_deploy_profile("m", {"deploy_config": str(path)})
assert out["parallel_stage_init"] is True
assert "rm_rf_slash" not in out
+9
View File
@@ -325,6 +325,15 @@ def apply_deploy_profile(model: str, args_dict: dict[str, Any]) -> dict[str, Any
"""Translate ``deploy_profile`` into ``deploy_config`` in-place (used by CLI and engine)."""
profile = args_dict.pop("deploy_profile", None)
if not profile:
# A deploy YAML carries the same `orchestrator:` block whichever flag
# loaded it. Applying it only on the profile path made the file mean
# different things depending on how it was passed, silently: pointing
# --deploy-config at deploy/edge/qwen3_tts.yaml loaded its stages but
# dropped `parallel_stage_init: true`, so an edge deployment started
# sequentially -- 63.8 s instead of 30.5 s on Qwen3-TTS -- with nothing
# in the log to say why.
if args_dict.get("deploy_config"):
_apply_profile_orchestrator_defaults(args_dict["deploy_config"], args_dict)
return args_dict
if args_dict.get("deploy_config"):
raise ValueError("`deploy_profile` and `deploy_config` are mutually exclusive; pass only one.")