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: guard multi-stage pipeline parallel topology
Scott Hanselman91c04d354da1
diff --git a/src/gputaskman/procinfo.py b/src/gputaskman/procinfo.pyindex 63603cc..8365cc4 100644--- a/src/gputaskman/procinfo.py+++ b/src/gputaskman/procinfo.py@@ -159,7 +159,8 @@ class ProcDetail:# 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"}+ "-n": "1", "--distributed-executor-backend": "mp", "--pipeline-parallel-size": "1",+ "-pp": "1"}watched = ("--config", "--headless", "--node-rank", *(f for f in single if f.startswith("--")))cmd = self.cmdlinefor i, arg in enumerate(cmd):@@ -171,6 +172,7 @@ class ProcDetail: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("-pp") and flag != "-pp")or (flag.startswith("-n") and flag != "-n")or flag.startswith("-r")):return Truediff --git a/tests/test_vllm.py b/tests/test_vllm.pyindex b10af5c..4a91ec2 100644--- a/tests/test_vllm.py+++ b/tests/test_vllm.py@@ -102,8 +102,8 @@ class VllmTests(unittest.TestCase):["--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"],+ ["--pipeline-parallel-size", "1"], ["--pipeline-parallel-size=1"], ["-pp", "1"], ["-pp=1"],+ ["--pipeline_parallel_size=1"],["--data-parallel-size=1", "--api-server-count=1", "--distributed-executor-backend=mp"],]rejected = [@@ -121,8 +121,11 @@ class VllmTests(unittest.TestCase):["--node_rank", "0"], ["--data_parallel_rank=0"],["--api-server", "2"], ["--nnode", "2"], ["--distributed-executor", "ray"], ["--head"],["--conf", "launch.yaml"], ["--node-r", "0"],+ ["--pipeline-parallel-size", "2"], ["--pipeline-parallel-size=2"],+ ["-pp", "2"], ["-pp=2"], ["-pp2"],+ ["--pipeline-parallel", "1"], ["--pipeline-parallel=1"],["--data-parallel-size"], ["-dp"], ["--api-server-count"],- ["--distributed-executor-backend"], ["--nnodes"],+ ["--distributed-executor-backend"], ["--nnodes"], ["--pipeline-parallel-size"],["--data-parallel-size", "--api-server-count", "1"],["--data-parallel-size=1", "--config", "launch.yaml"],["-dp", "2", "--api-server-count", "1"],