Skip to content

A GPU task manager for Windows, Linux, and WSL: per-process VRAM, GPU activity, and LLM model/KV-cache insights (Windows AMD and NVIDIA).

Python

fix: evaluate vLLM topology flag values before skipping API discovery

Reshmi Aravinda5c623cc207b
diff --git a/src/gputaskman/procinfo.py b/src/gputaskman/procinfo.py
index b0d0dc4..63603cc 100644
--- a/src/gputaskman/procinfo.py
+++ b/src/gputaskman/procinfo.py
@@ -156,10 +156,30 @@ class ProcDetail:
@property
def vllm_ambiguous(self) -> bool:
# Distributed/data-parallel layouts need an ownership graph, not just two ancestors.
- return any(arg.split("=", 1)[0].startswith("--data-parallel") or
- arg.split("=", 1)[0] in ("--distributed-executor-backend", "--nnodes", "--node-rank",
- "--headless", "--api-server-count")
- for arg in self.cmdline)
+ # Explicit single-node values are allowed. A --config path belongs to another process and is
+ # never opened, so any layout it could hide stays ambiguous. Aliases: vllm/engine/arg_utils.py.
+ single = {"--data-parallel-size": "1", "-dp": "1", "--api-server-count": "1", "--nnodes": "1",
+ "-n": "1", "--distributed-executor-backend": "mp"}
+ watched = ("--config", "--headless", "--node-rank", *(f for f in single if f.startswith("--")))
+ cmd = self.cmdline
+ for i, arg in enumerate(cmd):
+ flag, separator, value = arg.partition("=")
+ if flag.startswith("--"):
+ flag = flag.replace("_", "-") # vLLM's parser accepts --data_parallel_size
+ if len(flag) > 2 and any(name != flag and name.startswith(flag) for name in watched):
+ return True # argparse accepts unambiguous prefixes such as --api-server
+ if (flag in ("--config", "--headless", "--node-rank")
+ or (flag.startswith("--data-parallel") and flag != "--data-parallel-size")
+ or (flag.startswith("-dp") and flag != "-dp")
+ or (flag.startswith("-n") and flag != "-n")
+ or flag.startswith("-r")):
+ return True
+ if flag in single:
+ if not separator:
+ value = cmd[i + 1] if i + 1 < len(cmd) else ""
+ if value != single[flag]:
+ return True
+ return False
def _lib_name(base: str) -> str | None:
diff --git a/tests/test_vllm.py b/tests/test_vllm.py
index 7fb6588..b10af5c 100644
--- a/tests/test_vllm.py
+++ b/tests/test_vllm.py
@@ -91,6 +91,49 @@ def chain(details, parents):
class VllmTests(unittest.TestCase):
+ def test_vllm_topology_flags_are_value_aware(self):
+ # vllm/engine/arg_utils.py aliases (-dp, -n, -r, -tp, -pp); vllm/utils/argparse_utils.py
+ # FlexibleArgumentParser maps underscores to dashes and keeps argparse prefix matching.
+ allowed = [
+ [],
+ ["--data-parallel-size", "1"], ["--data-parallel-size=1"], ["-dp", "1"], ["-dp=1"],
+ ["--api-server-count", "1"], ["--api-server-count=1"],
+ ["--distributed-executor-backend", "mp"], ["--distributed-executor-backend=mp"],
+ ["--nnodes", "1"], ["--nnodes=1"], ["-n", "1"], ["-n=1"],
+ ["--data_parallel_size", "1"], ["--api_server_count=1"], ["--distributed_executor_backend", "mp"],
+ ["--tensor-parallel-size", "2"], ["-tp", "2"],
+ # PP is intentionally unchanged pending maintainer guidance.
+ ["--pipeline-parallel-size", "2"], ["-pp", "2"],
+ ["--data-parallel-size=1", "--api-server-count=1", "--distributed-executor-backend=mp"],
+ ]
+ rejected = [
+ ["--data-parallel-size", "2"], ["--data-parallel-size=2"], ["-dp", "2"], ["-dp=2"],
+ ["-dp2"], ["-dpn", "0"], ["-dpl", "2"],
+ ["--data-parallel-rank", "0"], ["--data-parallel-size-local", "1"],
+ ["--data-parallel-external-lb"],
+ ["--api-server-count", "2"], ["--api-server-count=0"],
+ ["--distributed-executor-backend", "ray"],
+ ["--distributed-executor-backend=external_launcher"],
+ ["--nnodes", "2"], ["--nnodes=2"], ["-n", "2"], ["-n2"],
+ ["--node-rank", "0"], ["--node-rank=0"], ["-r", "0"], ["-r0"],
+ ["--headless"], ["--config", "launch.yaml"], ["--config=launch.yaml"],
+ ["--data_parallel_size", "2"], ["--api_server_count=2"], ["--distributed_executor_backend", "ray"],
+ ["--node_rank", "0"], ["--data_parallel_rank=0"],
+ ["--api-server", "2"], ["--nnode", "2"], ["--distributed-executor", "ray"], ["--head"],
+ ["--conf", "launch.yaml"], ["--node-r", "0"],
+ ["--data-parallel-size"], ["-dp"], ["--api-server-count"],
+ ["--distributed-executor-backend"], ["--nnodes"],
+ ["--data-parallel-size", "--api-server-count", "1"],
+ ["--data-parallel-size=1", "--config", "launch.yaml"],
+ ["-dp", "2", "--api-server-count", "1"],
+ ]
+ for args in allowed:
+ with self.subTest(allowed=args):
+ self.assertFalse(ProcDetail(10, "python", cmdline=LAUNCH + args).vllm_ambiguous)
+ for args in rejected:
+ with self.subTest(rejected=args):
+ self.assertTrue(ProcDetail(10, "python", cmdline=LAUNCH + args).vllm_ambiguous)
+
def test_signatures_are_hints_even_without_mapped_models_or_titles(self):
cases = [
("vllm", LAUNCH, "server"),