Files
vLLM-4x-RTX-3090/patches/0070-qsa-keep-the-clip-counter-start-line-s-log-token.patch
MOVZX 777e910dd0 Initial commit
Signed-off-by: MOVZX <movzx@yahoo.com>
2026-10-15 01:01:52 +07:00

47 lines
2.5 KiB
Diff

From cde56de93bd067484ced3193b07fa0f93e94578d Mon Sep 17 00:00:00 2001
From: halt95 <halt95@users.noreply.github.com>
Date: Thu, 1 Oct 2026 05:39:19 -0400
Subject: [PATCH 070/103] qsa: keep the clip-counter start line's log token
The post-capture start line keeps the token "QSA clip counter armed (read interval Ns)" that log checks look for;
the previous change had reworded it. A source test pins the token.
---
tests/models/qwen4_exp/test_qsa_clip_reader_lifecycle.py | 3 +++
vllm/models/qwen4_exp/nvidia/qsa.py | 5 ++++-
2 files changed, 7 insertions(+), 1 deletion(-)
diff --git a/tests/models/qwen4_exp/test_qsa_clip_reader_lifecycle.py b/tests/models/qwen4_exp/test_qsa_clip_reader_lifecycle.py
index a81feb7..7de02bc 100644
--- a/tests/models/qwen4_exp/test_qsa_clip_reader_lifecycle.py
+++ b/tests/models/qwen4_exp/test_qsa_clip_reader_lifecycle.py
@@ -161,6 +161,9 @@ def test_source_pins():
fn = next(n for n in ast.walk(tree) if isinstance(n, ast.FunctionDef) and n.name == "_qsa_clip_count")
assert "threading.Thread" not in ast.unparse(fn), "the clip reader must not be started from the forward"
assert "register_post_capture_hook(_qsa_clip_start_reader)" in qsa
+ # the serving smoke greps r"QSA clip counter armed \(read interval [0-9.]+s\)" (an earlier wording of this line
+ # failed it); the post-capture start line must keep that token.
+ assert "QSA clip counter armed (read interval %.0fs)" in qsa
gtree = ast.parse(_src(_GT_PATH))
sr = next(n for n in ast.walk(gtree) if isinstance(n, ast.FunctionDef) and n.name == "start_reader")
stmts = [s for s in sr.body if not (isinstance(s, ast.Expr) and isinstance(s.value, ast.Constant))]
diff --git a/vllm/models/qwen4_exp/nvidia/qsa.py b/vllm/models/qwen4_exp/nvidia/qsa.py
index 467f7ad..a10f0a7 100644
--- a/vllm/models/qwen4_exp/nvidia/qsa.py
+++ b/vllm/models/qwen4_exp/nvidia/qsa.py
@@ -241,8 +241,11 @@ def _qsa_clip_start_reader() -> None:
name="qsa-clip-reader",
daemon=True,
).start()
+ # The leading "QSA clip counter armed (read interval Ns)" is the token the serving smoke greps
+ # (v2.2.0 logged it from the first forward); keep it verbatim when editing this line.
logger.info(
- "QSA clip counter reader started after capture (read interval %.0fs, %d layer(s) registered)",
+ "QSA clip counter armed (read interval %.0fs); reader started after capture "
+ "(%d layer(s) registered)",
_qsa_clip_interval,
len(_qsa_clip_counters),
)
--
2.43.0