multimodalart's picture
multimodalart HF Staff
Bernini-Diffusers-v2 r2v demo
fed6c68 verified
Raw
History Blame Contribute Delete
4.23 kB
# Copyright 2025 Bytedance Ltd. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
OpSlot: a lightweight dispatch point used in generated modeling code.
An ``OpSlot`` is placed at the module level of a generated modeling file.
At model-build time, ``_bind_veomni_ops`` resolves each slot to a concrete
kernel (or ``None`` for eager) via the global ``KERNEL_REGISTRY``.
Inside the model's ``forward`` methods, the pattern is a simple 2-line guard::
if veomni_moe_experts_forward.use_non_eager_impl:
return veomni_moe_experts_forward(self, hidden_states, ...)
# original HF code below, unchanged
"""
from __future__ import annotations
from typing import Any, Callable
from ..utils import logging
from .kernel_registry import KERNEL_REGISTRY
logger = logging.get_logger(__name__)
class OpSlot:
"""A named dispatch slot that can be bound to a kernel implementation."""
def __init__(self, op_name: str, variant: str):
self.op_name = op_name
self.variant = variant
self._kernel: Callable | None = None
self._impl_name: str | None = None # None ⇔ bind() has not been called
def bind(self, impl_name: str) -> None:
"""Resolve *impl_name* via the global registry and bind the result.
OpSlot instances are module-level globals, so two models sharing the
same modeling module share the same slot. Rebinding to a different
``impl_name`` silently overrides the first binding for *both*
instances — we warn so eager-vs-fused evaluation setups spot the
collision early.
"""
if self._impl_name is not None and self._impl_name != impl_name:
logger.warning_rank0(
f"OpSlot('{self.op_name}', '{self.variant}') was already bound to "
f"'{self._impl_name}'; rebinding to '{impl_name}'. Any other model "
"instance sharing this module will pick up the new binding."
)
self._kernel = KERNEL_REGISTRY.resolve(self.op_name, self.variant, impl_name)
self._impl_name = impl_name
@property
def use_non_eager_impl(self) -> bool:
"""``True`` when a non-eager kernel is bound.
Named for the guard pattern at call sites: ``if slot.use_non_eager_impl:
use replacement else fall through to eager HF code``. ``False`` covers
both "bound to eager" (``KERNEL_REGISTRY.resolve`` returned ``None``)
and "never bound".
"""
return self._kernel is not None
def bound_kernel(self) -> Callable | None:
"""Return the resolved kernel callable, or ``None`` if eager / unbound.
Use this when a model needs to *cache* the resolved implementation on
an instance attribute (e.g. ``self.causal_conv1d_fn = slot.bound_kernel()``).
Storing the OpSlot itself would couple the instance to the module-global
slot, so a later ``bind()`` from a second model in the same process would
silently override the first model's kernel.
"""
return self._kernel
def __call__(self, *args: Any, **kwargs: Any) -> Any:
if self._kernel is None:
raise RuntimeError(
f"OpSlot('{self.op_name}', '{self.variant}') has no kernel bound. "
"Call .bind() first or check .use_non_eager_impl before calling."
)
return self._kernel(*args, **kwargs)
def __repr__(self) -> str:
if self._impl_name is None:
state = "unbound"
elif self._kernel is None:
state = "eager"
else:
state = f"kernel={self._kernel}"
return f"OpSlot(op_name={self.op_name!r}, variant={self.variant!r}, {state})"