mmgp 3.5.8__tar.gz → 3.5.9__tar.gz

{mmgp-3.5.8/src/mmgp.egg-info → mmgp-3.5.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mmgp
-Version: 3.5.8
+Version: 3.5.9
 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 3.5.8 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.5.9 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

{mmgp-3.5.8 → mmgp-3.5.9}/README.md

@@ -1,6 +1,6 @@
 
 <p align="center">
-  <H2>Memory Management 3.5.8 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.5.9 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

{mmgp-3.5.8 → mmgp-3.5.9}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "mmgp"
-version = "3.5.8"
+version = "3.5.9"
 authors = [
   { name = "deepbeepmeep", email = "deepbeepmeep@yahoo.com" },
 ]

{mmgp-3.5.8 → mmgp-3.5.9}/src/mmgp/offload.py

@@ -1,4 +1,4 @@
-# ------------------ Memory Management 3.5.8 for the GPU Poor by DeepBeepMeep (mmgp)------------------
+# ------------------ Memory Management 3.5.9 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
@@ -671,7 +671,7 @@ def _welcome():
     if welcome_displayed:
          return 
     welcome_displayed = True
-    print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.5.8) by DeepBeepMeep ************{ENDC}{UNBOLD}")
+    print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.5.9) by DeepBeepMeep ************{ENDC}{UNBOLD}")
 
 def change_dtype(model, new_dtype, exclude_buffers = False):
     for submodule_name, submodule in model.named_modules():  
@@ -1179,11 +1179,10 @@ def load_loras_into_model(model, lora_path, lora_multi = None, activate_all_lora
                         loras_adapter_data[1] = lora_B.to(module.weight.dtype) 
                     else:
                         loras_adapter_data[2] = diff_b.to(module.weight.dtype) 
-                    if rank != None:
-                        alpha_key = k[:-len("lora_X.weight")] + "alpha"
+                    if rank != None and "lora" in lora_key:
+                        alpha_key = k[:-len(lora_key)] + "alpha"
                         alpha = lora_alphas.get(alpha_key, None)
-                        alpha = 1. if alpha == None else alpha / rank  
-                        loras_adapter_data[3] = alpha
+                        if alpha is not None: loras_adapter_data[3] = alpha / rank 
             lora_A = lora_B = diff = diff_b = v = loras_module_data = loras_adapter_data = lora_alphas = None
 
             if len(invalid_keys)  > 0:
@@ -2179,6 +2178,7 @@ class offload:
                     scaling = self._get_lora_scaling(loras_scaling, model, active_adapter) * alpha
                     if scaling == 0:
                         continue
+
                     if lora_A == None:
                         result.add_(diff_b, alpha=scaling)
                     else:

{mmgp-3.5.8 → mmgp-3.5.9/src/mmgp.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mmgp
-Version: 3.5.8
+Version: 3.5.9
 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 3.5.8 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.5.9 for the GPU Poor by DeepBeepMeep</H2>	
 </p>