fix get_offloaded_device

Signed-off-by: Kyle Sayers <[email protected]>

Author: kylesayrs

src/compressed_tensors/transform/factory/base.py

@@ -18,6 +18,7 @@
 
 import torch
 import torch.nn.utils.parametrize as P
+import tqdm
 from compressed_tensors import InternalModule
 from compressed_tensors.registry.registry import RegistryMixin, T
 from compressed_tensors.transform import (
@@ -84,15 +85,21 @@ def create_transform(self, module: Module, args: TransformArgs) -> "TransformBas
         """
         raise NotImplementedError()
 
-    def apply_to_model(self, model: Module):
+    def apply_to_model(self, model: Module, use_tqdm=True):
         """
         Create transforms and apply them to the model
 
         :param model: module to apply transforms to
         """
-        for arg in self.scheme.apply:
-            for _, module in match_named_modules(model, arg.targets, arg.ignore):
-                self._apply_to_module(module, arg)
+        modules_args = [
+            (module, arg)
+            for arg in self.scheme.apply
+            for _, module in match_named_modules(model, arg.targets, arg.ignore)
+        ]
+
+        desc = f"Applying {self.name} transforms"
+        for module, arg in tqdm.tqdm(modules_args, desc=desc, disable=(not use_tqdm)):
+            self._apply_to_module(module, arg)
 
         self._update_tied_weights()
 

src/compressed_tensors/transform/factory/hadamard.py

@@ -22,7 +22,7 @@
     apply_transform_weight,
     get_transform_size,
 )
-from compressed_tensors.utils import get_execution_device
+from compressed_tensors.utils import get_execution_device, get_offloaded_device
 from compressed_tensors.utils.helpers import ParameterizedDefaultDict
 from torch import Tensor, device, dtype
 from torch.nn import Module, Parameter
@@ -54,8 +54,9 @@ def create_transform(self, module: Module, args: TransformArgs):
         assert hasattr(module, "weight")
         size = get_transform_size(module, args.location, self.scheme.head_dim)
         exec_device = get_execution_device(module)
+        device = get_offloaded_device(module)
 
-        factory_kwargs = {"construct_device": exec_device}
+        factory_kwargs = {"device": device, "construct_device": exec_device}
         weight = self.weights.get(size, factory_kwargs=factory_kwargs)
         # TODO: permutations should be keyed by fused modules, not weight
         perm = self.perms[weight] if self.scheme.randomize else None
@@ -64,12 +65,12 @@ def create_transform(self, module: Module, args: TransformArgs):
     def _create_weight(
         self,
         size: int,
+        device: device,
         construct_device: device,
     ) -> Parameter:
-        # construct on execution device, cache shared tensor on cpu
         precision = self.scheme.precision
         data = deterministic_hadamard_matrix(size, precision, construct_device)
-        data = data.to(device="cpu")
+        data = data.to(device=device)
         return Parameter(data, requires_grad=self.scheme.requires_grad)
 
     def _create_permutation(self, weight: Parameter) -> Parameter:
@@ -104,10 +105,9 @@ def forward(self, value: Tensor) -> Tensor:
         if self.args.inverse:
             weight = weight.T
 
-        # onloading is done by accelerate if parent module is offloaded
         return (
             apply_transform_weight(
-                weight.to(dtype=self._precision, device=value.device),
+                weight.to(dtype=self._precision),
                 value.to(self._precision),
                 self.args.location,
                 self.module_type,

src/compressed_tensors/transform/factory/matrix_multiply.py

@@ -22,7 +22,8 @@
     get_transform_size,
 )
 from compressed_tensors.utils.helpers import ParameterizedDefaultDict
-from torch import Tensor
+from compressed_tensors.utils.offload import get_offloaded_device
+from torch import Tensor, device
 from torch.nn import Module, Parameter
 
 
@@ -52,20 +53,22 @@ def create_transform(self, module: Module, args: TransformArgs):
         """
         assert hasattr(module, "weight")
         size = get_transform_size(module, args.location, self.scheme.head_dim)
+        device = get_offloaded_device(module)
 
-        weight = self.weights.get(size)
+        factory_kwargs = {"device": device}
+        weight = self.weights.get(size, factory_kwargs=factory_kwargs)
         if args.inverse:
             weight = self.inverses[weight]
 
         return RandomMatrixTransform(weight, self.scheme, args, type(module))
 
-    def _create_weight(self, size: int) -> Parameter:
+    def _create_weight(self, size: int, device: device) -> Parameter:
         # TODO: construct such that weight is invertible (has non-zero determinant)
         data = torch.rand(
             (size, size),
             generator=self.generator,
             dtype=self.scheme.precision,
-            device=torch.device("cpu"),
+            device=device,
         )
         return Parameter(data, requires_grad=self.scheme.requires_grad)
 
@@ -91,9 +94,8 @@ def __init__(
         self._precision = scheme.precision if args.is_online() else torch.float64
 
     def forward(self, value: Tensor) -> Parameter:
-        # onloading is done by accelerate if parent module is offloaded
         return apply_transform_weight(
-            self.weight.to(dtype=self._precision, device=value.device),
+            self.weight.to(dtype=self._precision),
             value.to(self._precision),
             self.args.location,
             self.module_type,
@@ -102,7 +104,7 @@ def forward(self, value: Tensor) -> Parameter:
     def right_inverse(self, value: Tensor) -> Tensor:
         inverse = high_precision_invert(self.weight)
         return apply_transform_weight(
-            inverse.to(dtype=self._precision, device=value.device),
+            inverse.to(dtype=self._precision),
             value.to(self._precision),
             self.args.location,
             self.module_type,

src/compressed_tensors/transform/factory/random_hadamard.py

@@ -31,10 +31,10 @@ class RandomHadamardFactory(HadamardFactory):
     def _create_weight(
         self,
         size: int,
+        device: device,
         construct_device: device,
     ) -> Parameter:
-        # construct on execution device, cache on offload device
         precision = self.scheme.precision
         data = random_hadamard_matrix(size, precision, construct_device, self.generator)
-        data = data.to(device="cpu")
+        data = data.to(device=device)
         return Parameter(data, requires_grad=self.scheme.requires_grad)

src/compressed_tensors/utils/offload.py

@@ -130,7 +130,8 @@ def get_offloaded_device(module: torch.nn.Module) -> torch.device:
         first_key = list(module._hf_hook.weights_map.keys())[0]
         prefix_dataset = module._hf_hook.weights_map.dataset
         return prefix_dataset[first_key].device
-    return next(module.parameters()).device
+    else:
+        return get_execution_device(module)
 
 
 @check_accelerate(fallback=None)