mmgp 3.6.0__py3-none-any.whl → 3.6.1__py3-none-any.whl

mmgp/offload.py

@@ -1,4 +1,4 @@
-# ------------------ Memory Management 3.6.0 for the GPU Poor by DeepBeepMeep (mmgp)------------------
+# ------------------ Memory Management 3.6.1 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
@@ -688,7 +688,7 @@ def _welcome():
     if welcome_displayed:
          return 
     welcome_displayed = True
-    print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.6.0) by DeepBeepMeep ************{ENDC}{UNBOLD}")
+    print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.6.1) by DeepBeepMeep ************{ENDC}{UNBOLD}")
 
 def change_dtype(model, new_dtype, exclude_buffers = False):
     for submodule_name, submodule in model.named_modules():  
@@ -1097,7 +1097,9 @@ def load_loras_into_model(model, lora_path, lora_multi = None, activate_all_lora
 
             invalid_keys = []
             unexpected_keys = []
-            for k, v in state_dict.items():
+            new_state_dict = {}
+            for k in list(state_dict.keys()):
+                v = state_dict.pop(k)
                 lora_A = lora_B = diff_b = diff = lora_key = None
                 if k.endswith(".diff"):
                     diff = v
@@ -1141,6 +1143,7 @@ def load_loras_into_model(model, lora_path, lora_multi = None, activate_all_lora
                             error_msg = append(error_msg, msg) 
                             fail = True
                         break
+                    v = lora_A = lora_A.to(module.weight.dtype)                     
                 elif lora_B != None:
                     rank = lora_B.shape[1] 
                     if module_shape[0] != v.shape[0]:
@@ -1151,6 +1154,7 @@ def load_loras_into_model(model, lora_path, lora_multi = None, activate_all_lora
                             error_msg = append(error_msg, msg) 
                             fail = True
                         break
+                    v = lora_B = lora_B.to(module.weight.dtype)                     
                 elif diff != None:
                     lora_B = diff
                     if module_shape != v.shape:
@@ -1161,6 +1165,7 @@ def load_loras_into_model(model, lora_path, lora_multi = None, activate_all_lora
                             error_msg = append(error_msg, msg) 
                             fail = True
                         break
+                    v = lora_B = lora_B.to(module.weight.dtype)                     
                 elif diff_b != None:
                     rank = diff_b.shape[0] 
                     if not hasattr(module, "bias"):
@@ -1179,8 +1184,11 @@ def load_loras_into_model(model, lora_path, lora_multi = None, activate_all_lora
                                 error_msg = append(error_msg, msg) 
                                 fail = True
                             break
+                    v = diff_b = diff_b.to(module.weight.dtype)                     
                 
                 if not check_only:
+                    new_state_dict[k] = v
+                    v = None
                     loras_module_data = loras_model_data.get(module, None)
                     assert loras_module_data != None
                     loras_adapter_data =  loras_module_data.get(adapter_name, None)
@@ -1188,11 +1196,11 @@ def load_loras_into_model(model, lora_path, lora_multi = None, activate_all_lora
                         loras_adapter_data = [None, None, None, 1.]
                         loras_module_data[adapter_name] = loras_adapter_data
                     if lora_A != None:
-                        loras_adapter_data[0] = lora_A.to(module.weight.dtype) 
+                        loras_adapter_data[0] = lora_A
                     elif lora_B != None:
-                        loras_adapter_data[1] = lora_B.to(module.weight.dtype) 
+                        loras_adapter_data[1] = lora_B 
                     else:
-                        loras_adapter_data[2] = diff_b.to(module.weight.dtype) 
+                        loras_adapter_data[2] = diff_b 
                     if rank != None and lora_key is not None and "lora" in lora_key:
                         alpha_key = k[:-len(lora_key)] + "alpha"
                         alpha = lora_alphas.get(alpha_key, None)
@@ -1220,7 +1228,7 @@ def load_loras_into_model(model, lora_path, lora_multi = None, activate_all_lora
             if not check_only:
                 # model._loras_tied_weights[adapter_name] = tied_weights
                 if pinnedLora:
-                    pinned_sd_list.append(state_dict)
+                    pinned_sd_list.append(new_state_dict)
                     pinned_names_list.append(path)
                     # _pin_sd_to_memory(state_dict, path)
 

{mmgp-3.6.0.dist-info → mmgp-3.6.1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mmgp
-Version: 3.6.0
+Version: 3.6.1
 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 3.6.0 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.6.1 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

mmgp-3.6.1.dist-info/RECORD

@@ -0,0 +1,9 @@
+__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+mmgp/__init__.py,sha256=A9qBwyQMd1M7vshSTOBnFGP1MQvS2hXmTcTCMUcmyzE,509
+mmgp/offload.py,sha256=EPiMRzMLQ_6fnfCKeCLXfaSXKprMNcWilOhJhkdNkO8,132833
+mmgp/safetensors2.py,sha256=zYNMprt1KoxgVALbcz6DawxsQDNNRImvgO9cYRChUiY,19028
+mmgp-3.6.1.dist-info/licenses/LICENSE.md,sha256=HjzvY2grdtdduZclbZ46B2M-XpT4MDCxFub5ZwTWq2g,93
+mmgp-3.6.1.dist-info/METADATA,sha256=lTyBifQB0ApNB_OnIKrGXgPh6Bf66TUyEcqpGWbDTfM,16309
+mmgp-3.6.1.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+mmgp-3.6.1.dist-info/top_level.txt,sha256=waGaepj2qVfnS2yAOkaMu4r9mJaVjGbEi6AwOUogU_U,14
+mmgp-3.6.1.dist-info/RECORD,,

mmgp-3.6.0.dist-info/RECORD

@@ -1,9 +0,0 @@
-__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-mmgp/__init__.py,sha256=A9qBwyQMd1M7vshSTOBnFGP1MQvS2hXmTcTCMUcmyzE,509
-mmgp/offload.py,sha256=UaqWLw8jrNs9ibbIWplbLO5Cym84Txuu4lzttgxmnXs,132411
-mmgp/safetensors2.py,sha256=zYNMprt1KoxgVALbcz6DawxsQDNNRImvgO9cYRChUiY,19028
-mmgp-3.6.0.dist-info/licenses/LICENSE.md,sha256=HjzvY2grdtdduZclbZ46B2M-XpT4MDCxFub5ZwTWq2g,93
-mmgp-3.6.0.dist-info/METADATA,sha256=epm8_KuIB_c4W9iB31KIbHtNjdVuLyvW-DZoc8RR434,16309
-mmgp-3.6.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-mmgp-3.6.0.dist-info/top_level.txt,sha256=waGaepj2qVfnS2yAOkaMu4r9mJaVjGbEi6AwOUogU_U,14
-mmgp-3.6.0.dist-info/RECORD,,