city96-ComfyUI-GGUF/ops.py

# (c) City96 || Apache-2.0 (apache.org/licenses/LICENSE-2.0)
import gguf
import torch

import comfy.ops
from .dequant import dequantize_tensor

class GGMLTensor(torch.Tensor):
    """
    Main tensor-like class for storing quantized weights
    """
    def __init__(self, *args, tensor_type, tensor_shape, patches=[], **kwargs):
        super().__init__()
        self.tensor_type = tensor_type
        self.tensor_shape = tensor_shape
        self.patches = patches

    def __new__(cls, *args, tensor_type, tensor_shape, patches=[], **kwargs):
        return super().__new__(cls, *args, **kwargs)

    def to(self, *args, **kwargs):
        new = super().to(*args, **kwargs)
        new.tensor_type = getattr(self, "tensor_type", None)
        new.tensor_shape = getattr(self, "tensor_shape", new.data.shape)
        new.patches = getattr(self, "patches", []).copy()
        return new

    def clone(self, *args, **kwargs):
        return self

    def detach(self, *args, **kwargs):
        return self

    def copy_(self, *args, **kwargs):
        # fixes .weight.copy_ in comfy/clip_model/CLIPTextModel
        try:
            return super().copy_(*args, **kwargs)
        except Exception as e:
            print(f"ignoring 'copy_' on tensor: {e}")

    def __deepcopy__(self, *args, **kwargs):
        # Intel Arc fix, ref#50
        new = super().__deepcopy__(*args, **kwargs)
        new.tensor_type = getattr(self, "tensor_type", None)
        new.tensor_shape = getattr(self, "tensor_shape", new.data.shape)
        new.patches = getattr(self, "patches", []).copy()
        return new

    @property
    def shape(self):
        if not hasattr(self, "tensor_shape"):
            self.tensor_shape = self.size()
        return self.tensor_shape

class GGMLLayer(torch.nn.Module):
    """
    This (should) be responsible for de-quantizing on the fly
    """
    comfy_cast_weights = True
    dequant_dtype = None
    patch_dtype = None

    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
        for k,v in state_dict.items():
            if k[len(prefix):] == "weight":
                self.weight = torch.nn.Parameter(v, requires_grad=False)
            elif k[len(prefix):] == "bias" and v is not None:
                self.bias = torch.nn.Parameter(v, requires_grad=False)
            else:
                missing_keys.append(k)

    def _save_to_state_dict(self, destination, prefix, keep_vars):
        # This is a fake state dict for vram estimation
        weight = torch.zeros_like(self.weight, device=torch.device("meta"))
        destination[prefix + "weight"] = weight
        if self.bias is not None:
            bias = torch.zeros_like(self.bias, device=torch.device("meta"))
            destination[prefix + "bias"] = bias
        return

        # This would return the actual state dict
        destination[prefix + "weight"] = self.get_weight(self.weight)
        if bias is not None:
            destination[prefix + "bias"] = self.get_weight(self.bias)

    def get_weight(self, tensor, dtype):
        if tensor is None:
            return

        # consolidate and load patches to GPU in async
        patch_list = []
        device = tensor.device
        for function, patches, key in getattr(tensor, "patches", []):
            patch_list += move_patch_to_cuda(patches, device)

        # dequantize tensor while patches load
        weight = dequantize_tensor(tensor, dtype, self.dequant_dtype)

        # apply patches
        if patch_list:
            if self.patch_dtype is None:
                weight = function(patch_list, weight, key)
            else:
                # for testing, may degrade image quality
                patch_dtype = dtype if self.patch_dtype == "target" else self.patch_dtype
                weight = function(patch_list, weight, key, patch_dtype)
        return weight

    def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
        if input is not None:
            if dtype is None:
                dtype = input.dtype
            if bias_dtype is None:
                bias_dtype = dtype
            if device is None:
                device = input.device

        bias = None
        non_blocking = comfy.model_management.device_supports_non_blocking(device)
        if s.bias is not None:
            bias = s.get_weight(s.bias.to(device), dtype)
            bias = comfy.ops.cast_to(bias, bias_dtype, device, non_blocking=non_blocking, copy=False)

        weight = s.get_weight(s.weight.to(device), dtype)
        weight = comfy.ops.cast_to(weight, dtype, device, non_blocking=non_blocking, copy=False)
        return weight, bias

class GGMLOps(comfy.ops.manual_cast):
    """
    Dequantize weights on the fly before doing the compute
    """
    class Linear(GGMLLayer, comfy.ops.manual_cast.Linear):
        def forward_comfy_cast_weights(self, input):
            weight, bias = self.cast_bias_weight(input)
            return torch.nn.functional.linear(input, weight, bias)

    class Conv2d(GGMLLayer, comfy.ops.manual_cast.Conv2d):
        def forward_comfy_cast_weights(self, input):
            weight, bias = self.cast_bias_weight(input)
            return self._conv_forward(input, weight, bias)

def move_patch_to_cuda(item, device):
    if isinstance(item, torch.Tensor):
        return item.to(device, non_blocking=True)
    elif isinstance(item, tuple):
        return tuple(move_patch_to_cuda(x, device) for x in item)
    elif isinstance(item, list):
        return [move_patch_to_cuda(x, device) for x in item]
    else:
        return item