mmgp 3.5.0__tar.gz → 3.5.1__tar.gz

{mmgp-3.5.0/src/mmgp.egg-info → mmgp-3.5.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mmgp
-Version: 3.5.0
+Version: 3.5.1
 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 3.5.0 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.5.1 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

{mmgp-3.5.0 → mmgp-3.5.1}/README.md

@@ -1,6 +1,6 @@
 
 <p align="center">
-  <H2>Memory Management 3.5.0 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.5.1 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

{mmgp-3.5.0 → mmgp-3.5.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "mmgp"
-version = "3.5.0"
+version = "3.5.1"
 authors = [
   { name = "deepbeepmeep", email = "deepbeepmeep@yahoo.com" },
 ]

{mmgp-3.5.0 → mmgp-3.5.1}/src/mmgp/offload.py

@@ -1,4 +1,4 @@
-# ------------------ Memory Management 3.5.0 for the GPU Poor by DeepBeepMeep (mmgp)------------------
+# ------------------ Memory Management 3.5.1 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
@@ -658,7 +658,7 @@ def _welcome():
     if welcome_displayed:
          return 
     welcome_displayed = True
-    print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.5.0) by DeepBeepMeep ************{ENDC}{UNBOLD}")
+    print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.5.1) by DeepBeepMeep ************{ENDC}{UNBOLD}")
 
 def change_dtype(model, new_dtype, exclude_buffers = False):
     for submodule_name, submodule in model.named_modules():  
@@ -2014,6 +2014,8 @@ class offload:
                     weight = weight.clone()
                 for active_adapter in active_adapters:
                     data = loras_data.get(active_adapter + '_GPU', None)
+                    if data == None:
+                        continue                    
                     lora_A_weight, lora_B_weight, diff_b, alpha = data
                     scaling = self._get_lora_scaling(loras_scaling, model, active_adapter) * alpha
                     if lora_A_weight != None:
@@ -2028,7 +2030,6 @@ class offload:
                             original_bias = False
                         bias.add_(diff_b, alpha=scaling)
                     # base_weight += scaling * lora_B_weight @ lora_A_weight
-                    break
             if training:
                 pass
                 # result = torch.nn.functional.linear(dropout(x), base_weight, bias=submodule.bias)

{mmgp-3.5.0 → mmgp-3.5.1/src/mmgp.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mmgp
-Version: 3.5.0
+Version: 3.5.1
 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 3.5.0 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.5.1 for the GPU Poor by DeepBeepMeep</H2>	
 </p>