mmgp 3.3.2__py3-none-any.whl → 3.3.3__py3-none-any.whl

mmgp/offload.py

@@ -1,4 +1,4 @@
-# ------------------ Memory Management 3.3.2 for the GPU Poor by DeepBeepMeep (mmgp)------------------
+# ------------------ Memory Management 3.3.3 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
@@ -401,7 +401,7 @@ def _pin_sd_to_memory(sd, sd_name, tied_weights = None, gig_tensor_size = BIG_TE
     return 
 
 
-def _pin_to_memory(model, model_id, partialPinning = False, pinnedPEFTLora = True, gig_tensor_size = BIG_TENSOR_MAX_SIZE, verboseLevel = 1):
+def _pin_to_memory(model, model_id, partialPinning = False, pinnedPEFTLora = True, big_tensor_size = BIG_TENSOR_MAX_SIZE, verboseLevel = 1):
 
     global max_pinnable_bytes, total_pinned_bytes
     if max_pinnable_bytes > 0 and  max_pinnable_bytes >= max_pinnable_bytes:
@@ -474,7 +474,7 @@ def _pin_to_memory(model, model_id, partialPinning = False, pinnedPEFTLora = Tru
                 length = torch.numel(p.data) * p.data.element_size() 
 
             ref_cache[ref] = (n, length)
-            if current_big_tensor_size + length > gig_tensor_size :
+            if current_big_tensor_size + length > big_tensor_size and current_big_tensor_size !=0  :
                 big_tensors_sizes.append(current_big_tensor_size)
                 current_big_tensor_size = 0
                 big_tensor_no += 1
@@ -498,28 +498,11 @@ def _pin_to_memory(model, model_id, partialPinning = False, pinnedPEFTLora = Tru
     big_tensors_sizes.append(current_big_tensor_size)
 
     big_tensors = []
-    last_big_tensor = 0
     total = 0
     
 
     failed_planned_allocation = False
 
-    # for size in big_tensors_sizes:
-    #     try:
-    #         # if total > 7000 * ONE_MB:
-    #         #     raise  Exception ("test no more reserved RAM")
-    #         current_big_tensor = torch.empty( size, dtype= torch.uint8, pin_memory=True, device="cpu")
-    #         big_tensors.append(current_big_tensor)
-    #     except:
-    #         print(f"Unable to pin more tensors for this model as the maximum reservable memory has been reached ({total/ONE_MB:.2f})")
-    #         max_pinnable_bytes = total + total_pinned_bytes
-    #         failed_planned_allocation = True
-    #         break
-
-    #     last_big_tensor += 1
-    #     total += size
-
-
     gc.collect()
 
     last_allocated_big_tensor = -1        
@@ -561,13 +544,6 @@ def _pin_to_memory(model, model_id, partialPinning = False, pinnedPEFTLora = Tru
 
                 total += size
 
-            # if big_tensor_no != prev_big_tensor:
-            #     gc.collect()
-            #     prev_big_tensor = big_tensor_no
-            # match_param, match_isbuffer = tied_weights.get(n, (None, False))
-            # if match_param != None:
-
-            # if big_tensor_no>=0 and big_tensor_no < last_big_tensor:
             current_big_tensor = big_tensors[big_tensor_no]
             if is_buffer :
                 _force_load_buffer(p) # otherwise potential memory leak
@@ -600,9 +576,9 @@ def _pin_to_memory(model, model_id, partialPinning = False, pinnedPEFTLora = Tru
 
     if verboseLevel >=1:
         if partialPinning or failed_planned_allocation:        
-            print(f"The model was partially pinned to reserved RAM: {last_big_tensor} large blocks spread across {total/ONE_MB:.2f} MB")
+            print(f"The model was partially pinned to reserved RAM: {last_allocated_big_tensor + 1} large blocks spread across {total/ONE_MB:.2f} MB")
         else:
-            print(f"The whole model was pinned to reserved RAM: {last_big_tensor} large blocks spread across {total/ONE_MB:.2f} MB")
+            print(f"The whole model was pinned to reserved RAM: {last_allocated_big_tensor + 1} large blocks spread across {total/ONE_MB:.2f} MB")
 
     model._already_pinned = True
 
@@ -949,7 +925,7 @@ def load_loras_into_model(model, lora_path, lora_multi = None, activate_all_lora
             continue
         fail = False
         skip = False
-        state_dict = safetensors2.torch_load_file(path)
+        state_dict = safetensors2.torch_load_file(path, writable_tensors= False)
 
 
 

{mmgp-3.3.2.dist-info → mmgp-3.3.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mmgp
-Version: 3.3.2
+Version: 3.3.3
 Summary: Memory Management for the GPU Poor
 Author-email: deepbeepmeep <deepbeepmeep@yahoo.com>
 License:                     GNU GENERAL PUBLIC LICENSE
@@ -17,7 +17,7 @@ Dynamic: license-file
 
 
 <p align="center">
-  <H2>Memory Management 3.3.2 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.3.3 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

mmgp-3.3.3.dist-info/RECORD

@@ -0,0 +1,9 @@
+__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+mmgp/__init__.py,sha256=A9qBwyQMd1M7vshSTOBnFGP1MQvS2hXmTcTCMUcmyzE,509
+mmgp/offload.py,sha256=R0UbOXEGAFKd_6090o8v5CkVmJiWmHDQsww7A3-LZEU,106550
+mmgp/safetensors2.py,sha256=rmUbBmK3Dra5prUTTRSVi6-XUFAa9Mj6B5CNPgzt9To,17333
+mmgp-3.3.3.dist-info/licenses/LICENSE.md,sha256=HjzvY2grdtdduZclbZ46B2M-XpT4MDCxFub5ZwTWq2g,93
+mmgp-3.3.3.dist-info/METADATA,sha256=xcODp7uhIfvy7Il1xEp8ed2VYmH1Eln-EnLy3MM4VGM,16153
+mmgp-3.3.3.dist-info/WHEEL,sha256=DK49LOLCYiurdXXOXwGJm6U4DkHkg4lcxjhqwRa0CP4,91
+mmgp-3.3.3.dist-info/top_level.txt,sha256=waGaepj2qVfnS2yAOkaMu4r9mJaVjGbEi6AwOUogU_U,14
+mmgp-3.3.3.dist-info/RECORD,,

mmgp-3.3.2.dist-info/RECORD

@@ -1,9 +0,0 @@
-__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-mmgp/__init__.py,sha256=A9qBwyQMd1M7vshSTOBnFGP1MQvS2hXmTcTCMUcmyzE,509
-mmgp/offload.py,sha256=43FnFfWqwhh2qz0uykqEpxb_XP9Jx8MPGzN31PExT2w,107470
-mmgp/safetensors2.py,sha256=rmUbBmK3Dra5prUTTRSVi6-XUFAa9Mj6B5CNPgzt9To,17333
-mmgp-3.3.2.dist-info/licenses/LICENSE.md,sha256=HjzvY2grdtdduZclbZ46B2M-XpT4MDCxFub5ZwTWq2g,93
-mmgp-3.3.2.dist-info/METADATA,sha256=mVMLkutqhUihIeo8uo_LK71ithm84_AEaNvnyRnzmEA,16153
-mmgp-3.3.2.dist-info/WHEEL,sha256=DK49LOLCYiurdXXOXwGJm6U4DkHkg4lcxjhqwRa0CP4,91
-mmgp-3.3.2.dist-info/top_level.txt,sha256=waGaepj2qVfnS2yAOkaMu4r9mJaVjGbEi6AwOUogU_U,14
-mmgp-3.3.2.dist-info/RECORD,,