mmgp 3.6.5__py3-none-any.whl → 3.6.7__py3-none-any.whl

mmgp/offload.py

@@ -1,4 +1,4 @@
-# ------------------ Memory Management 3.6.5 for the GPU Poor by DeepBeepMeep (mmgp)------------------
+# ------------------ Memory Management 3.6.7 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 3.6.7) by DeepBeepMeep ************{ENDC}{UNBOLD}")
 
 def change_dtype(model, new_dtype, exclude_buffers = False):
     for submodule_name, submodule in model.named_modules():  
@@ -1075,15 +1083,10 @@ def load_loras_into_model(model, lora_path, lora_multi = None, activate_all_lora
 
         if not fail:
             pos = first_key.find(".")
-            prefix = first_key[0:pos]
-            if prefix not in ["diffusion_model", "transformer"]:
-                msg = f"No compatible weight was found in Lora file '{path}'. Please check that it is compatible with the Diffusers format."
-                error_msg = append(error_msg, msg) 
-                fail = True
+            prefix = first_key[0:pos+1]
+            if prefix in ["diffusion_model.", "transformer."]:
+                state_dict = { k[ len(prefix):]: v for k, v in state_dict.items() if k.startswith(prefix) }
 
-        if not fail:
-
-            state_dict = { k[ len(prefix) + 1:]: v for k, v in state_dict.items() if k.startswith(prefix) }
             clean_up = True
 
             keys = list(state_dict.keys())
@@ -1421,7 +1424,6 @@ 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 +1434,7 @@ 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.dist-info → mmgp-3.6.7.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mmgp
-Version: 3.6.5
+Version: 3.6.7
 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 3.6.7 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

mmgp-3.6.7.dist-info/RECORD

@@ -0,0 +1,10 @@
+__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+mmgp/__init__.py,sha256=A9qBwyQMd1M7vshSTOBnFGP1MQvS2hXmTcTCMUcmyzE,509
+mmgp/fp8_quanto_bridge.py,sha256=VtUaD6wzo7Yn9vGY0LMtbhwt6KMWRpSWLc65bU_sfZU,21155
+mmgp/offload.py,sha256=_-fgKyHyIZ7mWgjzZ_PsmShStl0i7soo4ZXpzA55b6A,138996
+mmgp/safetensors2.py,sha256=zYNMprt1KoxgVALbcz6DawxsQDNNRImvgO9cYRChUiY,19028
+mmgp-3.6.7.dist-info/licenses/LICENSE.md,sha256=DD-WIS0BkPoWJ_8hQO3J8hMP9K_1-dyrYv1YCbkxcDU,94
+mmgp-3.6.7.dist-info/METADATA,sha256=PwOn_JzoyOYqoBZtPhnT9B-wNFH7fGHszW0Ypg7LZFI,16309
+mmgp-3.6.7.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+mmgp-3.6.7.dist-info/top_level.txt,sha256=waGaepj2qVfnS2yAOkaMu4r9mJaVjGbEi6AwOUogU_U,14
+mmgp-3.6.7.dist-info/RECORD,,

{mmgp-3.6.5.dist-info → mmgp-3.6.7.dist-info}/licenses/LICENSE.md

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

mmgp-3.6.5.dist-info/RECORD

@@ -1,10 +0,0 @@
-__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-mmgp/__init__.py,sha256=A9qBwyQMd1M7vshSTOBnFGP1MQvS2hXmTcTCMUcmyzE,509
-mmgp/fp8_quanto_bridge.py,sha256=VtUaD6wzo7Yn9vGY0LMtbhwt6KMWRpSWLc65bU_sfZU,21155
-mmgp/offload.py,sha256=GwM0o0rWUwY3tb_HDl_pO6S0XvMpVZsre3DzwFCLQh4,138988
-mmgp/safetensors2.py,sha256=zYNMprt1KoxgVALbcz6DawxsQDNNRImvgO9cYRChUiY,19028
-mmgp-3.6.5.dist-info/licenses/LICENSE.md,sha256=HjzvY2grdtdduZclbZ46B2M-XpT4MDCxFub5ZwTWq2g,93
-mmgp-3.6.5.dist-info/METADATA,sha256=9wXPfJYiHExhfM-kSeA4mLoNh0laDRXJ3ZrBFSpQlUg,16309
-mmgp-3.6.5.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-mmgp-3.6.5.dist-info/top_level.txt,sha256=waGaepj2qVfnS2yAOkaMu4r9mJaVjGbEi6AwOUogU_U,14
-mmgp-3.6.5.dist-info/RECORD,,