mrq/bitsandbytes-rocm
1
0
Fork 1
Code Issues 1 Pull Requests Packages Projects Releases Wiki Activity

Revert the layout if weights were reordered

Browse Source
This commit is contained in:
Max Ryabinin 2023-02-25 06:02:06 +01:00
parent cd4d904a4c
commit cc608c04c2
1 changed files with 42 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

55
bitsandbytes/nn/modules.py
View File

@ -9,6 +9,8 @@ import torch.nn.functional as F
from torch import Tensor, device, dtype, nn from torch import Tensor, device, dtype, nn
import bitsandbytes as bnb import bitsandbytes as bnb
import bitsandbytes.functional
from bitsandbytes.autograd._functions import get_inverse_transform_indices, undo_layout
from bitsandbytes.optim import GlobalOptimManager from bitsandbytes.optim import GlobalOptimManager
T = TypeVar("T", bound="torch.nn.Module") T = TypeVar("T", bound="torch.nn.Module")
@ -210,7 +212,7 @@ class Int8Params(torch.nn.Parameter):
class Linear8bitLt(nn.Linear): class Linear8bitLt(nn.Linear):
def __init__(self, input_features, output_features, bias=True, has_fp16_weights=True, def __init__(self, input_features, output_features, bias=True, has_fp16_weights=True,
memory_efficient_backward=False, threshold=0.0, index=None): memory_efficient_backward=False, threshold=0.0, index=None):
super().__init__(input_features, output_features, bias) super().__init__(input_features, output_features, bias)
assert not memory_efficient_backward, "memory_efficient_backward is no longer required and the argument is deprecated in 0.37.0 and will be removed in 0.39.0" assert not memory_efficient_backward, "memory_efficient_backward is no longer required and the argument is deprecated in 0.37.0 and will be removed in 0.39.0"
self.state = bnb.MatmulLtState() self.state = bnb.MatmulLtState()
@ -225,21 +227,48 @@ class Linear8bitLt(nn.Linear):
self.weight = Int8Params(self.weight.data, has_fp16_weights=has_fp16_weights, requires_grad=has_fp16_weights) self.weight = Int8Params(self.weight.data, has_fp16_weights=has_fp16_weights, requires_grad=has_fp16_weights)
def _save_to_state_dict(self, destination, prefix, keep_vars): def _save_to_state_dict(self, destination, prefix, keep_vars):
super()._save_to_state_dict(destination, prefix, keep_vars) if not self.state.has_fp16_weights and self.state.CB is None and self.state.CxB is not None:
# reorder weight layout back from ampere/turing to row
reorder_layout = True
weight_clone = self.weight.data.clone()
else:
reorder_layout = False
# we only need to save SCB as extra data, because CB for quantized weights is already stored in weight.data try:
weight_name = "SCB" if reorder_layout:
if self.state.tile_indices is None:
order, tile_size = self.state.formatB, self.state.get_tile_size()
transform = lambda x: \
bitsandbytes.functional.transform(x.to(self.weight.data.device), from_order="row",
to_order=order)[0].to(x.device)
with torch.no_grad():
self.state.tile_indices = get_inverse_transform_indices(transform, tile_size).to(
self.state.CxB.device)
# case 1: .cuda was called, SCB is in self.weight CB = (
param_from_weight = getattr(self.weight, weight_name) undo_layout(self.state.CxB, self.state.tile_indices)
# case 2: self.init_8bit_state was called, SCB is in self.state )
param_from_state = getattr(self.state, weight_name)
key_name = prefix + f"{weight_name}" self.weight.data = CB
if param_from_weight is not None:
destination[key_name] = param_from_weight if keep_vars else param_from_weight.detach() super()._save_to_state_dict(destination, prefix, keep_vars)
elif not self.state.has_fp16_weights and param_from_state is not None:
destination[key_name] = param_from_state if keep_vars else param_from_state.detach() # we only need to save SCB as extra data, because CB for quantized weights is already stored in weight.data
weight_name = "SCB"
# case 1: .cuda was called, SCB is in self.weight
param_from_weight = getattr(self.weight, weight_name)
# case 2: self.init_8bit_state was called, SCB is in self.state
param_from_state = getattr(self.state, weight_name)
key_name = prefix + f"{weight_name}"
if param_from_weight is not None:
destination[key_name] = param_from_weight if keep_vars else param_from_weight.detach()
elif not self.state.has_fp16_weights and param_from_state is not None:
destination[key_name] = param_from_state if keep_vars else param_from_state.detach()
finally:
if reorder_layout:
self.weight.data = weight_clone
def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,
missing_keys, unexpected_keys, error_msgs): missing_keys, unexpected_keys, error_msgs):