compressed-tensors 0.10.2a20250616__py3-none-any.whl → 0.10.2a20250620__py3-none-any.whl

compressed_tensors/utils/offload.py

@@ -85,6 +85,7 @@ __all__ = [
     "delete_offload_module",
     "offloaded_dispatch",
     "disable_offloading",
+    "remove_dispatch",
 ]
 
 
@@ -170,22 +171,22 @@ def update_parameter_data(
 
 def get_execution_device(module: torch.nn.Module) -> torch.device:
     """
-    Get the device which inputs should be moved to before module execution
+    Get the device which inputs should be moved to before module execution.
+    Assume that modules execute in the same order as returned by `model.modules()`
 
     :param module: module to check, may be offloaded
     :return: onload device of module
     """
-    if has_offloaded_params(module):
-        return module._hf_hook.execution_device
+    for submodule in module.modules():
+        if has_offloaded_params(submodule):
+            return submodule._hf_hook.execution_device
 
-    first_param = next(module.parameters(), None)
-    if first_param is None:
-        warnings.warn(
-            f"Unable able to infer execution device of {module}, falling back to CPU"
-        )
-        return torch.device("cpu")
+        param = next(submodule.parameters(recurse=False), None)
+        if param is not None:
+            return param.device
 
-    return first_param.device
+    warnings.warn(f"Unable to get execution device of {module}, falling back to CPU")
+    return torch.device("cpu")
 
 
 def register_offload_parameter(
@@ -514,6 +515,9 @@ def offloaded_dispatch(
     if offload_device == "disk":
         raise NotImplementedError("Disk offloading is not currently supported")
 
+    # remove any existing hooks
+    remove_dispatch(module)
+
     # create weights map
     state_dict = module.state_dict()
     state_dict = {key: val.to(offload_device) for key, val in state_dict.items()}
@@ -535,6 +539,33 @@ def offloaded_dispatch(
         weights_map=weights_map,
         tied_params_map=tied_params_map,
     )
+
+    # when saving a model, `PretrainedModel.save_pretrained` will only
+    # onload weights if the following requirements are met
+    # if (
+    #     hasattr(self, "hf_device_map")
+    #     and len(set(self.hf_device_map.values())) > 1
+    #     and ("cpu" in self.hf_device_map.values()
+    #          or "disk" in self.hf_device_map.values())
+    # ):
+    # because this function always offloads, disregard actual devices and
+    # always use `cpu` and `cuda:0` to guarantee this condition passes
+    setattr(module, "hf_device_map", {"fake_offload": "cpu", "fake_exec": "cuda:0"})
+
+    return module
+
+
+def remove_dispatch(module: torch.nn.Module) -> torch.nn.Module:
+    """
+    Remove any existing dispatches from module
+
+    :param module: module which may be dispatched with hf hooks
+    :return: module without dispatch
+    """
+    remove_hook_from_module(module, recurse=True)
+    if hasattr(module, "hf_device_map"):
+        delattr(module, "hf_device_map")
+
     return module
 
 
@@ -563,7 +594,7 @@ def disable_offloading():
     # update any parameters which may have changed
     for module, (hook, offload) in onloaded_modules.items():
         hook.offload = offload
-        for name, param in module.named_parameters():
+        for name, param in module.named_parameters(recurse=False):
             update_offload_parameter(module, name, param.data)
         hook.post_forward(module, None)
 

compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.10.2.a20250616'
+__version__ = version = '0.10.2.a20250620'
 __version_tuple__ = version_tuple = (0, 10, 2)

{compressed_tensors-0.10.2a20250616.dist-info → compressed_tensors-0.10.2a20250620.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.2a20250616
+Version: 0.10.2a20250620
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.2a20250616.dist-info → compressed_tensors-0.10.2a20250620.dist-info}/RECORD

@@ -1,6 +1,6 @@
 compressed_tensors/__init__.py,sha256=UtKmifNeBCSE2TZSAfduVNNzHY-3V7bLjZ7n7RuXLOE,812
 compressed_tensors/base.py,sha256=73HYH7HY7O2roC89yG_piPFnZwrBfn_i7HmKl90SKc0,875
-compressed_tensors/version.py,sha256=luzxcDdA8EA_-n05GudjzinnLHUkR5eQCjyTTNLxDnw,523
+compressed_tensors/version.py,sha256=fdaxTQXKxDCy4sL7AzY19gLcpfD2GYpzFhuqxROV3Bo,523
 compressed_tensors/compressors/__init__.py,sha256=smSygTSfcfuujRrAXDc6uZm4L_ccV1tWZewqVnOb4lM,825
 compressed_tensors/compressors/base.py,sha256=nvWsv4xEw1Tkxkxth6TmHplDYXfBeP22xWxOsZERyDY,7204
 compressed_tensors/compressors/helpers.py,sha256=OK6qxX9j3bHwF9JfIYSGMgBJe2PWjlTA3byXKCJaTIQ,5431
@@ -54,13 +54,13 @@ compressed_tensors/transform/utils/hadamards.safetensors,sha256=mFd1GzNodGG-ifA1
 compressed_tensors/transform/utils/utils.py,sha256=PRPTYwPs2nnNaQMq2GEbC4QYKHFKlZwaRyPgdDhl66g,2992
 compressed_tensors/utils/__init__.py,sha256=gS4gSU2pwcAbsKj-6YMaqhm25udFy6ISYaWBf-myRSM,808
 compressed_tensors/utils/helpers.py,sha256=cPg-ikdeA92aIGwBONg8GmPNvcGlFhozyJVwsRiXBTA,11981
-compressed_tensors/utils/offload.py,sha256=11k8p3y_0WVbrozYsrPjJb_4Ln8YkeO_4Q8W4ibFVlc,22710
+compressed_tensors/utils/offload.py,sha256=3XiBuWbUkBAt8v1t5i57qDcbB3VJQs_FDeayi-JzIWg,23896
 compressed_tensors/utils/permutations_24.py,sha256=kx6fsfDHebx94zsSzhXGyCyuC9sVyah6BUUir_StT28,2530
 compressed_tensors/utils/permute.py,sha256=V6tJLKo3Syccj-viv4F7ZKZgJeCB-hl-dK8RKI_kBwI,2355
 compressed_tensors/utils/safetensors_load.py,sha256=DMfZBuUbA6qp_BG_zIWT3ckiEE33K9ob34s-OgzReO4,12057
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
-compressed_tensors-0.10.2a20250616.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.10.2a20250616.dist-info/METADATA,sha256=V_L8-0yZM5t7FG0knLacaaIAHNTuEmTe1f7CqmDCAXg,7005
-compressed_tensors-0.10.2a20250616.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-compressed_tensors-0.10.2a20250616.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.10.2a20250616.dist-info/RECORD,,
+compressed_tensors-0.10.2a20250620.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.10.2a20250620.dist-info/METADATA,sha256=hc0YFMpgLrIwHZUEvRbpaShjSDbczmVgRm4pEDst6QM,7005
+compressed_tensors-0.10.2a20250620.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+compressed_tensors-0.10.2a20250620.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.10.2a20250620.dist-info/RECORD,,