Handle spinloop ext load failure gracefully (#43659)

Signed-off-by: Patrick Schlangen <pschlan@amd.com>
Co-authored-by: Shengqi Chen <harry-chen@outlook.com>
This commit is contained in:
pschlan-amd
2026-06-03 16:09:52 +00:00
committed by GitHub
co-authored by Shengqi Chen
parent 51e0c579b0
commit 0a5cbf633e
2 changed files with 16 additions and 7 deletions
+2
View File
@@ -112,6 +112,8 @@ endif()
# #
# spinloop extension (pure CXX; must stay above the non-CUDA device branch so # spinloop extension (pure CXX; must stay above the non-CUDA device branch so
# CPU builds define the target before the early return) # CPU builds define the target before the early return)
# This extension requires SABI 3.11 since it relies on Py_buffer support. Loading
# failure is handled gracefully on vLLM side for lower Python versions.
# #
set(VLLM_SPINLOOP_EXT_SRC "csrc/spinloop.cpp") set(VLLM_SPINLOOP_EXT_SRC "csrc/spinloop.cpp")
set(SPINLOOP_COMPILE_FLAGS "") set(SPINLOOP_COMPILE_FLAGS "")
@@ -38,9 +38,19 @@ from vllm.utils.network_utils import (
is_valid_ipv6_address, is_valid_ipv6_address,
) )
if envs.VLLM_USE_SPINLOOP_EXT: logger = init_logger(__name__)
from vllm.spinloop import spinloop
SPINLOOP_EXT_ENABLED = False
if envs.VLLM_USE_SPINLOOP_EXT:
try:
from vllm.spinloop import spinloop
SPINLOOP_EXT_ENABLED = True
except ImportError:
logger.warning(
"spinloop extension could not be loaded, disabling VLLM_USE_SPINLOOP_EXT!"
)
SPINLOOP_TIMEOUT_SECONDS = 0.1 SPINLOOP_TIMEOUT_SECONDS = 0.1
if TYPE_CHECKING: if TYPE_CHECKING:
@@ -82,9 +92,6 @@ def to_bytes_big(value: int, size: int) -> bytes:
return value.to_bytes(size, byteorder="big") return value.to_bytes(size, byteorder="big")
logger = init_logger(__name__)
LONG_WAIT_TIME_LOG_MSG = ( LONG_WAIT_TIME_LOG_MSG = (
"No available shared memory broadcast block found " "No available shared memory broadcast block found "
"in %d seconds. This typically happens " "in %d seconds. This typically happens "
@@ -552,7 +559,7 @@ class MessageQueue:
written_flag = metadata_buffer[0] written_flag = metadata_buffer[0]
return not (written_flag and read_count != self.buffer.n_reader) return not (written_flag and read_count != self.buffer.n_reader)
if envs.VLLM_USE_SPINLOOP_EXT and not check(): if SPINLOOP_EXT_ENABLED and not check():
spinloop(metadata_buffer, check, timeout=SPINLOOP_TIMEOUT_SECONDS) spinloop(metadata_buffer, check, timeout=SPINLOOP_TIMEOUT_SECONDS)
if not check(): if not check():
@@ -673,7 +680,7 @@ class MessageQueue:
written_flag = metadata_buffer[0] written_flag = metadata_buffer[0]
return not (not written_flag or read_flag) return not (not written_flag or read_flag)
if envs.VLLM_USE_SPINLOOP_EXT and not check(): if SPINLOOP_EXT_ENABLED and not check():
spinloop( spinloop(
metadata_buffer[0 : self.local_reader_rank + 1], metadata_buffer[0 : self.local_reader_rank + 1],
check, check,