85 lines
2.9 KiB
Python
85 lines
2.9 KiB
Python
"""Scoped handling of expected bitsandbytes activation-cast diagnostics."""
|
|
|
|
from contextlib import contextmanager
|
|
from importlib import metadata
|
|
import logging
|
|
import re
|
|
import threading
|
|
import warnings
|
|
|
|
|
|
_CAST_MESSAGES = frozenset(
|
|
f"MatMul8bitLt: inputs will be cast from torch.{dtype} "
|
|
"to float16 during quantization"
|
|
for dtype in ("bfloat16", "float32")
|
|
)
|
|
_BNB_LOGGER = "bitsandbytes.autograd._functions"
|
|
_MINIMUM_BNB_VERSION = (0, 46, 1)
|
|
|
|
|
|
def warn_if_suspicious_8bit_stack(node_name: str) -> None:
|
|
"""Warn, without blocking inference, when bnb predates the project floor."""
|
|
try:
|
|
version = metadata.version("bitsandbytes")
|
|
except Exception:
|
|
return
|
|
parts = tuple(int(piece) for piece in re.findall(r"\d+", version)[:3])
|
|
normalized = parts + (0,) * (3 - len(parts))
|
|
if normalized < _MINIMUM_BNB_VERSION:
|
|
warnings.warn(
|
|
f"CaptionForge's {node_name} node detected bitsandbytes {version}; "
|
|
"this older 8-bit inference stack may cause severe slowdowns or compatibility issues. "
|
|
"CaptionForge will continue without modifying packages.",
|
|
RuntimeWarning,
|
|
stacklevel=2,
|
|
)
|
|
|
|
|
|
class _QuantizedCastLogFilter(logging.Filter):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.thread_id = threading.get_ident()
|
|
|
|
def filter(self, record):
|
|
return not (
|
|
record.name == _BNB_LOGGER
|
|
and record.levelno == logging.WARNING
|
|
and record.thread == self.thread_id
|
|
and record.getMessage() in _CAST_MESSAGES
|
|
)
|
|
|
|
|
|
@contextmanager
|
|
def quantized_inference_warnings(enabled: bool):
|
|
"""Preserve diagnostics except known BF16/FP32 casts during 8-bit inference.
|
|
|
|
No filter is installed at import time or for Default mode. catch_warnings
|
|
restores the caller's filters even when generation raises an exception.
|
|
On Python 3.10-3.12 warning filters are process-wide during this brief scope;
|
|
the exact message, category and module keep the suppression narrowly bounded.
|
|
"""
|
|
if not enabled:
|
|
yield
|
|
return
|
|
|
|
with warnings.catch_warnings():
|
|
warnings.filterwarnings(
|
|
"ignore",
|
|
message=(
|
|
r"\AMatMul8bitLt: inputs will be cast from torch\.(?:bfloat16|float32) "
|
|
r"to float16 during quantization\Z"
|
|
),
|
|
category=UserWarning,
|
|
module=r"\Abitsandbytes\.autograd\._functions\Z",
|
|
)
|
|
# Recent bitsandbytes versions use logging rather than warnings.warn.
|
|
# Attach to the emitting logger so propagation/ComfyUI handlers remain
|
|
# untouched. Limit this route to the current inference thread as well.
|
|
logger = logging.getLogger(_BNB_LOGGER)
|
|
log_filter = _QuantizedCastLogFilter()
|
|
logger.addFilter(log_filter)
|
|
try:
|
|
yield
|
|
finally:
|
|
logger.removeFilter(log_filter)
|