xiaolibai-sys
164896ed78
- nvfp4: _quantization_metadata -> comfy_quant synthesis, SlotEntry extra params (block_scale), dual-scale LoRA fold, orig_dtype fixup - LoRA: plain-mode block LoRAs no longer silently dropped; folded on GPU after weights land - loader: parallel positioned reads, pipelined group load, no per-tensor clone, header-only dtype probes - block swap: read_block intra-block parallel + disk_workers node option; async D2H write-back on transfer stream; lazy evict_all skips wasted write-back - arch: split models/loader.py out of wan_model.py; single utils/keys.py; LoRA folding consolidated in utils/lora.py; global utils/log.py ([BerniniR][Tag] format, BERNINI_LOG_LEVEL)
69 lines
2.3 KiB
Python
69 lines
2.3 KiB
Python
"""Attention backend configuration node for Bernini-R."""
|
|
|
|
from __future__ import annotations
|
|
|
|
|
|
from ..utils.types import BerniniAttention
|
|
from .backends import (
|
|
BACKEND_NAMES,
|
|
available_backends,
|
|
best_available,
|
|
)
|
|
|
|
from ..utils.log import get_logger as _get_logger
|
|
|
|
logger = _get_logger("Attn")
|
|
class BerniniR_AttentionConfig:
|
|
"""Select attention backend with automatic fallback.
|
|
|
|
Configures which attention implementation to use during sampling.
|
|
``"auto"`` selects the best available backend (SageAttention 3 > 2 >
|
|
FlashAttention > xformers > PyTorch SDPA).
|
|
|
|
Connect to ``BerniniR_ModelLoader`` via its ``attn_backend_args`` input
|
|
to apply the backend at model load time.
|
|
|
|
Inputs:
|
|
backend (COMBO): Attention backend name.
|
|
``"auto"`` — auto-select best available.
|
|
Others only shown if the library is installed.
|
|
force_backend (BOOLEAN): If True, error on unavailable backend
|
|
instead of silently falling back.
|
|
|
|
Output:
|
|
BERNINI_ATTN: Configuration dict consumed by BerniniR_ModelLoader.
|
|
"""
|
|
|
|
@classmethod
|
|
def INPUT_TYPES(s):
|
|
return {
|
|
"required": {
|
|
"backend": (BACKEND_NAMES, {"default": "auto", "tooltip": "Attention implementation. 'auto' = best available. Chain: Sage3 → Sage2/1 → FlashAttn → xformers → SDPA"}),
|
|
"force_backend": ("BOOLEAN", {"default": False, "tooltip": "If True, error on unavailable backend. If False, silently fall back through the chain"}),
|
|
},
|
|
}
|
|
|
|
RETURN_TYPES = ("BERNINI_ATTN",)
|
|
RETURN_NAMES = ("attention_config",)
|
|
FUNCTION = "configure"
|
|
CATEGORY = "Bernini-R/Sampling"
|
|
DESCRIPTION = (
|
|
"Attention backend selector. 'auto' = best available. "
|
|
"Chain: SageAttn3 → SageAttn2 → FlashAttn → xformers → SDPA."
|
|
)
|
|
|
|
def configure(self, backend: str = "auto", force_backend: bool = False):
|
|
available = available_backends()
|
|
best = best_available()
|
|
config = BerniniAttention(
|
|
backend=backend,
|
|
force_backend=force_backend,
|
|
available=tuple(available),
|
|
best=best,
|
|
)
|
|
logger.info(
|
|
f"Attention config: backend={backend}, "
|
|
f"force={force_backend}, available={available}, best={best}"
|
|
)
|
|
return (config,)
|