[Deprecation] Deprecate items scheduled for 0.29 (#55353)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
@@ -1,70 +1,8 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Test that verifies no implicit GPU-CPU synchronization occurs during
|
||||
speculative decoding generation under expected conditions.
|
||||
"""
|
||||
|
||||
import multiprocessing
|
||||
import sys
|
||||
import traceback
|
||||
"""Tests for sync-free speculative decoding with async scheduling."""
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def sync_tracker():
|
||||
"""
|
||||
Fixture that patches CommonAttentionMetadata.seq_lens_cpu to detect
|
||||
lazy init syncs. Prints stack traces immediately when syncs occur.
|
||||
"""
|
||||
from vllm.v1.attention.backend import CommonAttentionMetadata
|
||||
|
||||
# Shared counter for cross-process communication (inherited by fork)
|
||||
sync_count = multiprocessing.Value("i", 0)
|
||||
|
||||
# Save original property
|
||||
original_prop = CommonAttentionMetadata.seq_lens_cpu
|
||||
original_fget = original_prop.fget
|
||||
|
||||
# Create tracking wrapper
|
||||
def tracking_seq_lens_cpu(self):
|
||||
if self._seq_lens_cpu is None:
|
||||
# Increment counter
|
||||
with sync_count.get_lock():
|
||||
sync_count.value += 1
|
||||
count = sync_count.value
|
||||
# Print stack trace immediately (shows in subprocess output)
|
||||
print(f"\n{'=' * 60}", file=sys.stderr)
|
||||
print(f"SYNC #{count}: seq_lens_cpu lazy init triggered!", file=sys.stderr)
|
||||
print(f"{'=' * 60}", file=sys.stderr)
|
||||
traceback.print_stack(file=sys.stderr)
|
||||
print(f"{'=' * 60}\n", file=sys.stderr)
|
||||
sys.stderr.flush()
|
||||
return original_fget(self)
|
||||
|
||||
# Apply patch
|
||||
CommonAttentionMetadata.seq_lens_cpu = property(tracking_seq_lens_cpu)
|
||||
|
||||
class SyncTracker:
|
||||
@property
|
||||
def count(self) -> int:
|
||||
return sync_count.value
|
||||
|
||||
def assert_no_sync(self, msg: str = ""):
|
||||
count = sync_count.value
|
||||
assert count == 0, (
|
||||
f"Unexpected GPU-CPU sync: seq_lens_cpu lazy init triggered "
|
||||
f"{count} times. See stack traces above. {msg}"
|
||||
)
|
||||
|
||||
yield SyncTracker()
|
||||
|
||||
# Restore original property
|
||||
CommonAttentionMetadata.seq_lens_cpu = original_prop
|
||||
torch._dynamo.reset()
|
||||
|
||||
|
||||
# Test configurations: (model, spec_model, method, num_spec_tokens, backend_env)
|
||||
SPEC_DECODE_CONFIGS = [
|
||||
@@ -97,18 +35,16 @@ SPEC_DECODE_CONFIGS = [
|
||||
SPEC_DECODE_CONFIGS,
|
||||
)
|
||||
def test_no_sync_with_spec_decode(
|
||||
sync_tracker,
|
||||
model: str,
|
||||
spec_model: str,
|
||||
method: str,
|
||||
num_spec_tokens: int,
|
||||
monkeypatch,
|
||||
vllm_runner,
|
||||
):
|
||||
"""
|
||||
Test that no implicit GPU-CPU sync occurs during speculative decoding
|
||||
generation.
|
||||
"""
|
||||
# Import vLLM AFTER sync_tracker fixture has applied the patch
|
||||
"""Test generation while the worker rejects unintended GPU-CPU syncs."""
|
||||
monkeypatch.setenv("VLLM_GPU_SYNC_CHECK", "error")
|
||||
|
||||
from vllm import SamplingParams
|
||||
from vllm.config import CompilationConfig
|
||||
|
||||
@@ -158,5 +94,3 @@ def test_no_sync_with_spec_decode(
|
||||
assert outputs[0].outputs[0].text, (
|
||||
f"{method} target={model}: generated completion text is empty"
|
||||
)
|
||||
|
||||
sync_tracker.assert_no_sync()
|
||||
|
||||
Reference in New Issue
Block a user