mmgp 3.6.5__tar.gz → 3.6.6__tar.gz

{mmgp-3.6.5 → mmgp-3.6.6}/LICENSE.md

@@ -1,2 +1,2 @@
-                    GNU GENERAL PUBLIC LICENSE
+                    GNU GENERAL PUBLIC LICENSE
                        Version 3, 29 June 2007

{mmgp-3.6.5/src/mmgp.egg-info → mmgp-3.6.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mmgp
-Version: 3.6.5
+Version: 3.6.6
 Summary: Memory Management for the GPU Poor
 Author-email: deepbeepmeep <deepbeepmeep@yahoo.com>
 Requires-Python: >=3.10
@@ -15,7 +15,7 @@ Dynamic: license-file
 
 
 <p align="center">
-  <H2>Memory Management 3.6.5 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.6.4 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

{mmgp-3.6.5 → mmgp-3.6.6}/README.md

@@ -1,6 +1,6 @@
 
 <p align="center">
-  <H2>Memory Management 3.6.5 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.6.4 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

{mmgp-3.6.5 → mmgp-3.6.6}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "mmgp"
-version = "3.6.5"
+version = "3.6.6"
 authors = [
   { name = "deepbeepmeep", email = "deepbeepmeep@yahoo.com" },
 ]

{mmgp-3.6.5 → mmgp-3.6.6}/src/mmgp/offload.py

@@ -1,4 +1,4 @@
-# ------------------ Memory Management 3.6.5 for the GPU Poor by DeepBeepMeep (mmgp)------------------
+# ------------------ Memory Management 3.6.6 for the GPU Poor by DeepBeepMeep (mmgp)------------------
 #
 # This module contains multiples optimisations so that models such as Flux (and derived), Mochi, CogView, HunyuanVideo, ...  can run smoothly on a 24 GB GPU limited card. 
 # This a replacement for the accelerate library that should in theory manage offloading, but doesn't work properly with models that are loaded / unloaded several
@@ -63,7 +63,7 @@ import json
 import psutil
 import builtins
 from accelerate import init_empty_weights
-
+from functools import wraps
 import functools
 import types
 
@@ -86,6 +86,14 @@ class QEmbedding(QModuleMixin, torch.nn.Embedding):
 
 
 
+def cudacontext(device):
+    def decorator(func):
+        @wraps(func)
+        def wrapper(*args, **kwargs):
+            with torch.device(device):
+                return func(*args, **kwargs)
+        return wrapper
+    return decorator
 
 
 shared_state = {}
@@ -689,7 +697,7 @@ def _welcome():
     if welcome_displayed:
          return 
     welcome_displayed = True
-    print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.6.5) by DeepBeepMeep ************{ENDC}{UNBOLD}")
+    print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.6.6) by DeepBeepMeep ************{ENDC}{UNBOLD}")
 
 def change_dtype(model, new_dtype, exclude_buffers = False):
     for submodule_name, submodule in model.named_modules():  
@@ -1421,7 +1429,6 @@ def fast_load_transformers_model(model_path: str,  do_quantize = False, quantiza
             model = transfomer_class.from_config(transformer_config )
 
 
-    torch.set_default_device('cpu')
     model.eval().requires_grad_(False)
 
     model._config = transformer_config
@@ -1432,6 +1439,7 @@ def fast_load_transformers_model(model_path: str,  do_quantize = False, quantiza
 
 
 
+@cudacontext("cpu")
 def load_model_data(model, file_path, do_quantize = False, quantizationType = qint8, pinToMemory = False, partialPinning = False, modelPrefix = None, writable_tensors = True,  preprocess_sd = None, postprocess_sd = None, modules = None, return_shared_modules = None, default_dtype = torch.bfloat16, ignore_unused_weights = False, verboseLevel = -1):
     """
     Load a model, detect if it has been previously quantized using quanto and do the extra setup if necessary

{mmgp-3.6.5 → mmgp-3.6.6/src/mmgp.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mmgp
-Version: 3.6.5
+Version: 3.6.6
 Summary: Memory Management for the GPU Poor
 Author-email: deepbeepmeep <deepbeepmeep@yahoo.com>
 Requires-Python: >=3.10
@@ -15,7 +15,7 @@ Dynamic: license-file
 
 
 <p align="center">
-  <H2>Memory Management 3.6.5 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.6.4 for the GPU Poor by DeepBeepMeep</H2>	
 </p>