burt-imma / python /burt_imma /__init__.py
SNAPKITTYWEST's picture
push from SNAPKITTYWEST/burt-imma
b88c26d verified
Raw
History Blame Contribute Delete
1.17 kB
"""
BURT-IMMA: Bi-encoder Unified Retrieval Transformer with
Interferometric Matrix Memory Architecture
License: BSL-1.1
Contact: jessica@collectivekitty.com
"""
__version__ = "0.1.0"
# Try to import CUDA-accelerated kernels, fall back to pure PyTorch
_CUDA_AVAILABLE = False
try:
import _burt_imma_cuda
_CUDA_AVAILABLE = True
except ImportError:
_burt_imma_cuda = None
# Import kernel wrappers (these handle CUDA/CPU fallback internally)
from .kernels import (
constrained_softmax,
cifg_update,
batched_cifg_update,
sparse_moe_dispatch,
biencoder_attention,
attention_softmax,
)
from .utils import (
spectral_norm,
entropy,
check_huntington,
load_config,
set_seed,
Timer,
)
__all__ = [
# Kernel functions
"constrained_softmax",
"cifg_update",
"batched_cifg_update",
"sparse_moe_dispatch",
"biencoder_attention",
"attention_softmax",
# Utilities
"spectral_norm",
"entropy",
"check_huntington",
"load_config",
"set_seed",
"Timer",
# Meta
"__version__",
"_CUDA_AVAILABLE",
]