iHore/vLLM-4x-RTX-3090
8
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
main
vLLM-4x-RTX-3090/patches/0068-guard-the-bounds-guard-telemetry-code-default-is-war.patch
T
MOVZX 777e910dd0 Initial commit 
Signed-off-by: MOVZX <movzx@yahoo.com>
2026-10-15 01:01:52 +07:00

227 lines
10 KiB
Diff
Raw Permalink Blame History

From 0f95d8a93025684b4a36b6e57ccd908b589e812e Mon Sep 17 00:00:00 2001
From: halt95 <halt95@users.noreply.github.com>
Date: Thu, 1 Oct 2026 05:05:23 -0400
Subject: [PATCH 068/103] guard: the bounds-guard telemetry code default is
warn
VLLM_E2_GUARD_MODE now defaults to warn in code, matching what the shipped launchers set; v2 through v2.2.0
defaulted to off. The comment above it, the MODES docstring, the vllm/envs.py default and its annotation follow.
tests/v1/worker/gpu/test_guard_default_mode.py checks that code, comment and envs.py agree, from the sources, so
it runs without CUDA, and fails on the earlier source.
---
.../v1/worker/gpu/test_guard_default_mode.py | 86 +++++++++++++++++++
.../test_guard_telemetry_defaults_drain.py | 10 +--
vllm/envs.py | 4 +-
vllm/v1/worker/guard_telemetry.py | 22 ++---
4 files changed, 105 insertions(+), 17 deletions(-)
create mode 100644 tests/v1/worker/gpu/test_guard_default_mode.py
diff --git a/tests/v1/worker/gpu/test_guard_default_mode.py b/tests/v1/worker/gpu/test_guard_default_mode.py
new file mode 100644
index 0000000..4dc4516
--- /dev/null
+++ b/tests/v1/worker/gpu/test_guard_default_mode.py
@@ -0,0 +1,86 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""The guard's CODE default is warn (since v2.5), and the code, its comment and
+vllm/envs.py agree.
+
+Source-level pins: they parse the files instead of importing them, so they run on a machine with no
+CUDA, no zmq and no working vllm import. The behavioural check at the end imports the
+module only when that works.
+"""
+
+import ast
+import importlib
+import os
+import re
+import sys
+
+import pytest
+
+_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), *([".."] * 4)))
+_GT = os.path.join(_ROOT, "vllm", "v1", "worker", "guard_telemetry.py")
+_ENVS = os.path.join(_ROOT, "vllm", "envs.py")
+
+
+def _read(path):
+ with open(path, encoding="utf-8") as fh:
+ return fh.read()
+
+
+def _getenv_defaults(src, name):
+ """every default literal passed to os.environ.get / os.getenv for `name`"""
+ out = []
+ for node in ast.walk(ast.parse(src)):
+ if not isinstance(node, ast.Call) or len(node.args) < 2:
+ continue
+ fn = ast.unparse(node.func)
+ if fn not in ("os.environ.get", "os.getenv"):
+ continue
+ a0, a1 = node.args[0], node.args[1]
+ if isinstance(a0, ast.Constant) and a0.value == name and isinstance(a1, ast.Constant):
+ out.append(a1.value)
+ return out
+
+
+def test_code_default_is_warn():
+ d = _getenv_defaults(_read(_GT), "VLLM_E2_GUARD_MODE")
+ assert d == ["warn"], d
+
+
+def test_envs_default_is_warn_everywhere():
+ src = _read(_ENVS)
+ assert _getenv_defaults(src, "VLLM_E2_GUARD_MODE") == ["warn"]
+ assert re.search(r'VLLM_E2_GUARD_MODE: str = "warn"', src)
+ assert 'VLLM_E2_GUARD_MODE: str = "off"' not in src
+
+
+def test_comment_and_docstring_agree_with_the_default():
+ src = _read(_GT)
+ line = next(ln for ln in src.splitlines() if ln.startswith("_MODE = os.environ.get("))
+ i = src.splitlines().index(line)
+ above = "\n".join(src.splitlines()[max(0, i - 8) : i])
+ assert "Release defaults" in above and "warn, 0.05 s" in above
+ assert "v2.5" in above
+ # the MODES block must name warn as the default and must not call off the default
+ assert re.search(r"warn\s+\(DEFAULT since v2\.5", src)
+ assert not re.search(r"off\s+\(DEFAULT", src)
+ # the interval default stays 0.05 (the comment says so)
+ assert _getenv_defaults(src, "VLLM_E2_GUARD_INTERVAL") == ["0.05"]
+
+
+def test_failing_control_the_old_default_would_be_caught():
+ """the pin must fail on the earlier source (default off)"""
+ old = _read(_GT).replace('"VLLM_E2_GUARD_MODE", "warn"', '"VLLM_E2_GUARD_MODE", "off"')
+ assert old != _read(_GT)
+ assert _getenv_defaults(old, "VLLM_E2_GUARD_MODE") == ["off"]
+ assert _getenv_defaults(old, "VLLM_E2_GUARD_MODE") != ["warn"]
+
+
+def test_module_default_with_env_absent(monkeypatch):
+ monkeypatch.delenv("VLLM_E2_GUARD_MODE", raising=False)
+ monkeypatch.delenv("VLLM_E2_GUARD_INTERVAL", raising=False)
+ sys.modules.pop("vllm.v1.worker.guard_telemetry", None)
+ try:
+ gt = importlib.import_module("vllm.v1.worker.guard_telemetry")
+ except Exception as e: # noqa: BLE001 (no CUDA / zmq on this host: the source pins above stand in)
+ pytest.skip(f"vllm.v1.worker.guard_telemetry not importable here: {e!r}")
+ assert gt._MODE == "warn"
diff --git a/tests/v1/worker/gpu/test_guard_telemetry_defaults_drain.py b/tests/v1/worker/gpu/test_guard_telemetry_defaults_drain.py
index 606292e..32a99c1 100644
--- a/tests/v1/worker/gpu/test_guard_telemetry_defaults_drain.py
+++ b/tests/v1/worker/gpu/test_guard_telemetry_defaults_drain.py
@@ -2,8 +2,8 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Bounds-guard instrumentation: release defaults, reader loss and the final drain.
-1. Release defaults are off / 0.05 with both variables absent (earlier
- builds defaulted to warn), and a
+1. Release defaults are warn / 0.05 with both variables absent (v2 through
+ v2.2.0 defaulted to off), and a
non-finite or non-positive interval is rejected at import.
2. A reader that loses a device arms the stop (the fatal-site contract is
otherwise silently void).
@@ -32,7 +32,7 @@ def _fresh(monkeypatch, **env):
def test_release_defaults_with_env_absent(monkeypatch):
gt = _fresh(monkeypatch)
- assert gt._MODE == "off" # v2: default off
+ assert gt._MODE == "warn" # code default warn since v2.5
assert pytest.approx(0.05) == gt._INTERVAL
@@ -45,7 +45,7 @@ def test_bad_interval_rejected_at_import(monkeypatch, bad):
def test_reader_loss_arms_the_stop(monkeypatch):
gt = _fresh(
monkeypatch, VLLM_E2_GUARD_MODE="warn"
- ) # the reader exists only when opted in (default off)
+ ) # explicit, so the test does not depend on the default
device = torch.device("cuda:0")
gt.prealloc(device)
gt.reset_for_test()
@@ -65,7 +65,7 @@ def test_reader_loss_arms_the_stop(monkeypatch):
def test_drain_reports_last_step_hit_and_raises_for_fatal_site(monkeypatch):
gt = _fresh(
monkeypatch, VLLM_E2_GUARD_MODE="warn"
- ) # warn mode (opt-in): a plain site must NOT raise
+ ) # warn mode: a plain site must NOT raise
device = torch.device("cuda:0")
gt.prealloc(device)
gt.reset_for_test()
diff --git a/vllm/envs.py b/vllm/envs.py
index f13f442..4356854 100755
--- a/vllm/envs.py
+++ b/vllm/envs.py
@@ -399,7 +399,7 @@ if TYPE_CHECKING:
# Guard telemetry: mode raise|warn|off and the reader poll interval in
# seconds. counter() returns a real tensor in every mode, so the Triton
# arguments and the captured graph do not depend on either.
- VLLM_E2_GUARD_MODE: str = "off"
+ VLLM_E2_GUARD_MODE: str = "warn"
VLLM_E2_GUARD_INTERVAL: str = "0.05"
# Debug allocator telemetry (poll seconds + output directory) and the
# scheduler allocation log path.
@@ -2426,7 +2426,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
).strip(),
"VLLM_QSA_KV_SCALES": lambda: os.getenv("VLLM_QSA_KV_SCALES", "").strip(),
"VLLM_E2_GUARD_MODE": lambda: (
- os.getenv("VLLM_E2_GUARD_MODE", "off").strip().lower()
+ os.getenv("VLLM_E2_GUARD_MODE", "warn").strip().lower()
),
"VLLM_E2_GUARD_INTERVAL": lambda: os.getenv("VLLM_E2_GUARD_INTERVAL", "0.05"),
"VLLM_E2_MEMSTATS": lambda: os.getenv("VLLM_E2_MEMSTATS", "0"),
diff --git a/vllm/v1/worker/guard_telemetry.py b/vllm/v1/worker/guard_telemetry.py
index 63c450f..b02a7c3 100644
--- a/vllm/v1/worker/guard_telemetry.py
+++ b/vllm/v1/worker/guard_telemetry.py
@@ -40,13 +40,13 @@ MODES -- VLLM_E2_GUARD_MODE:
serving product than a loud stop. Read check_pending below before
using it -- the raise is late and it is local to one rank.
Targeted repro boots.
- warn log only -- EXCEPT for the sites in _ALWAYS_FATAL_SITES, which arm
- the raise whatever the mode says, because their masked value has
- already been written somewhere durable. USE THIS FOR LONG STRESS
- BOOTS.
- off (DEFAULT since v2: the reader thread
- and the QSA clip reader cost ~0.3 ms per decode step on TP2xPP2; opt in
- with VLLM_E2_GUARD_MODE=warn for diagnosis)
+ warn (DEFAULT since v2.5) log only -- EXCEPT for the sites in
+ _ALWAYS_FATAL_SITES, which arm the raise whatever the mode says,
+ because their masked value has already been written somewhere durable.
+ USE THIS FOR LONG STRESS BOOTS.
+ off (the default in v2 through v2.2.0: the reader thread and the QSA clip
+ reader cost ~0.3 ms per decode step on TP2xPP2; set
+ VLLM_E2_GUARD_MODE=off to skip the reader)
no reader thread. The kernels still mask and still count; nothing is
read -- INCLUDING the sites in _ALWAYS_FATAL_SITES. Use only when you
have decided you do not want the evidence.
@@ -218,10 +218,12 @@ SITE_NAMES = {
}
-# Release defaults: off, 0.05 s.
+# Release defaults: warn, 0.05 s.
# Earlier revisions shipped raise / 0.5 while every qualification boot set both
# variables explicitly, so the defaults were never the thing being tested.
-_MODE = os.environ.get("VLLM_E2_GUARD_MODE", "off").strip().lower()
+# v2 through v2.2.0 shipped the code default off; v2.5 puts
+# it back to warn so the code default, this comment and the shipped launchers agree.
+_MODE = os.environ.get("VLLM_E2_GUARD_MODE", "warn").strip().lower()
if _MODE not in ("raise", "warn", "off"):
# fail closed: a set-but-broken gate must not silently disarm the instrument
raise ValueError(f"VLLM_E2_GUARD_MODE must be raise|warn|off, got {_MODE!r}")
@@ -491,7 +493,7 @@ def start_reader() -> None:
host, and a device-to-host copy from a second thread while a capture is
active is exactly the call class cudaStreamCaptureModeGlobal -- the default
for torch.cuda.graph -- rejects. The QSA clip-counter reader survives it in
- practice, but that one is maintainer-gated and this one can be enabled by any user, so
+ practice, but that one is opt-in and this one is on by default, so
the precedent is evidence, not proof. Starting after capture deletes the
question rather than arguing it.
--
2.43.0
Reference in New Issue View Git Blame Copy Permalink