mmgp 3.2.7__tar.gz → 3.2.8__tar.gz

{mmgp-3.2.7/src/mmgp.egg-info → mmgp-3.2.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: mmgp
-Version: 3.2.7
+Version: 3.2.8
 Summary: Memory Management for the GPU Poor
 Author-email: deepbeepmeep <deepbeepmeep@yahoo.com>
 License:                     GNU GENERAL PUBLIC LICENSE
@@ -16,7 +16,7 @@ Requires-Dist: psutil
 
 
 <p align="center">
-  <H2>Memory Management 3.2.7 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.2.8 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

{mmgp-3.2.7 → mmgp-3.2.8}/README.md

@@ -1,6 +1,6 @@
 
 <p align="center">
-  <H2>Memory Management 3.2.7 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.2.8 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

{mmgp-3.2.7 → mmgp-3.2.8}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "mmgp"
-version = "3.2.7"
+version = "3.2.8"
 authors = [
   { name = "deepbeepmeep", email = "deepbeepmeep@yahoo.com" },
 ]

{mmgp-3.2.7 → mmgp-3.2.8}/src/mmgp/offload.py

@@ -1,4 +1,4 @@
-# ------------------ Memory Management 3.2.7 for the GPU Poor by DeepBeepMeep (mmgp)------------------
+# ------------------ Memory Management 3.2.8 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
@@ -575,7 +575,7 @@ def _welcome():
     if welcome_displayed:
          return 
     welcome_displayed = True
-    print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.2.7) by DeepBeepMeep ************{ENDC}{UNBOLD}")
+    print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.2.8) by DeepBeepMeep ************{ENDC}{UNBOLD}")
 
 def _extract_num_from_str(num_in_str):
     size = len(num_in_str)
@@ -910,6 +910,9 @@ def load_loras_into_model(model, lora_path, lora_multi = None, activate_all_lora
         skip = False
         state_dict = safetensors2.torch_load_file(path)
 
+
+
+
         if preprocess_sd != None:
             state_dict = preprocess_sd(state_dict)
 
@@ -947,42 +950,34 @@ def load_loras_into_model(model, lora_path, lora_multi = None, activate_all_lora
             # tied_weights = _extract_tie_weights_from_sd(state_dict, path) # to do
 
         clean_up = False
-        keys = list(state_dict.keys())
-        if len(keys) == 0:
+        first_key = next(iter(state_dict), None)
+        if first_key == None:
             msg = f"Empty Lora '{path}'"
             error_msg = append(error_msg, msg) 
             fail = True
 
         if not fail:
-            network_alphas = {}
-            for k in keys:
-                if "alpha" in k:
-                    alpha_value = state_dict.pop(k)
-                    if not ( (torch.is_tensor(alpha_value) and torch.is_floating_point(alpha_value)) or isinstance(
-                        alpha_value, float
-                    )):
-                        network_alphas[k] =  torch.tensor( float(alpha_value.item() ) )
-
-            pos = keys[0].find(".")
-            prefix = keys[0][0:pos]
+            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
 
         if not fail:
+
             state_dict = { k[ len(prefix) + 1:]: v for k, v in state_dict.items() if k.startswith(prefix) }
-            rank = {}
             clean_up = True
 
-            # for key, val in state_dict.items():
-            #     if "lora_B" in key:
-            #         rank[key] = val.shape[1]
+            keys = list(state_dict.keys())
 
-            # if network_alphas is not None and len(network_alphas) >= 1:
-            #     alpha_keys = [k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix]
-            #     network_alphas = {k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys}
-            network_alphas = None
+            lora_alphas = {}
+            for k in keys:
+                if "alpha" in k:
+                    alpha_value = state_dict.pop(k)
+                    if torch.is_tensor(alpha_value):
+                        alpha_value = float(alpha_value.item())
+                    lora_alphas[k] = alpha_value
 
             invalid_keys = []
             unexpected_keys = []
@@ -1037,14 +1032,18 @@ def load_loras_into_model(model, lora_path, lora_multi = None, activate_all_lora
                         loras_module_data = dict()
                         loras_model_data[module] = loras_module_data
                     loras_adapter_data =  loras_module_data.get(adapter_name, None)
+                    lora_A = None if lora_A == None else lora_A.to(torch.bfloat16) 
+                    lora_B = None if lora_B == None else lora_B.to(torch.bfloat16) 
                     if loras_adapter_data == None:
-                        loras_adapter_data = [lora_A, lora_B]
+                        alpha = lora_alphas.get(k[:-len("lora_X.weight")] + "alpha", 1.) 
+                        loras_adapter_data = [lora_A, lora_B, alpha]
                         loras_module_data[adapter_name] = loras_adapter_data
                     elif lora_A != None:
                         loras_adapter_data[0] = lora_A
                     else:
                         loras_adapter_data[1] = lora_B
             lora_A, lora_B, v, loras_module_data, loras_adapter_data = None, None, None, None, None
+            lora_alphas = None
 
             if len(invalid_keys)  > 0:
                 msg = "Lora '{path}' contains non Lora keys '{trunc(invalid_keys,200)}'"
@@ -1548,10 +1547,10 @@ class offload:
                 lora_data = lora_module.get(adapter, None)
                 if lora_data == None:
                     continue                     
-                lora_A, lora_B = lora_data
+                lora_A, lora_B, alpha = lora_data
                 key = adapter + '_GPU'
                 if to_GPU:
-                    lora_module[key] = [lora_A.cuda(), lora_B.cuda()]
+                    lora_module[key] = [lora_A.cuda(), lora_B.cuda(), alpha]
                 elif key in lora_module:
                     del lora_module[key]
             
@@ -1801,8 +1800,8 @@ class offload:
                     data = loras_data.get(active_adapter + '_GPU', None)
                     if data == None:
                         continue
-                    lora_A_weight, lora_B_weight = data
-                    scaling = get_scaling(active_adapter)
+                    lora_A_weight, lora_B_weight, alpha = data
+                    scaling = get_scaling(active_adapter) * alpha
                     weight.addmm_(lora_B_weight, lora_A_weight, alpha= scaling )
                     # base_weight += scaling * lora_B_weight @ lora_A_weight
 
@@ -1822,9 +1821,9 @@ class offload:
                     data = loras_data.get(active_adapter + '_GPU', None)
                     if data == None:
                         continue
-                    lora_A, lora_B = data
+                    lora_A, lora_B, alpha = data
                     # dropout = self.lora_dropout[active_adapter]
-                    scaling = get_scaling(active_adapter)
+                    scaling = get_scaling(active_adapter) * alpha
                     x = x.to(lora_A.dtype)
 
                     if training:        
@@ -2248,7 +2247,7 @@ def all(pipe_or_dict_of_modules, pinnedMemory = False, pinnedPEFTLora = False, l
                     print(f"Model '{model_id}' already pinned to reserved memory")
             else:
                 _pin_to_memory(current_model, model_id, partialPinning= partialPinning, pinnedPEFTLora = pinnedPEFTLora, verboseLevel=verboseLevel)            
-
+    
         current_budget = model_budgets[model_id]
         cur_blocks_prefix, prev_blocks_name, cur_blocks_name,cur_blocks_seq, is_mod_seq = None, None, None, -1, False
         self.loaded_blocks[model_id] = None

{mmgp-3.2.7 → mmgp-3.2.8/src/mmgp.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: mmgp
-Version: 3.2.7
+Version: 3.2.8
 Summary: Memory Management for the GPU Poor
 Author-email: deepbeepmeep <deepbeepmeep@yahoo.com>
 License:                     GNU GENERAL PUBLIC LICENSE
@@ -16,7 +16,7 @@ Requires-Dist: psutil
 
 
 <p align="center">
-  <H2>Memory Management 3.2.7 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.2.8 for the GPU Poor by DeepBeepMeep</H2>	
 </p>