mmgp 3.5.12__tar.gz → 3.6.0__tar.gz

{mmgp-3.5.12/src/mmgp.egg-info → mmgp-3.6.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mmgp
-Version: 3.5.12
+Version: 3.6.0
 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 3.5.12 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.6.0 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

{mmgp-3.5.12 → mmgp-3.6.0}/README.md

@@ -1,6 +1,6 @@
 
 <p align="center">
-  <H2>Memory Management 3.5.12 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.6.0 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

{mmgp-3.5.12 → mmgp-3.6.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "mmgp"
-version = "3.5.12"
+version = "3.6.0"
 authors = [
   { name = "deepbeepmeep", email = "deepbeepmeep@yahoo.com" },
 ]

{mmgp-3.5.12 → mmgp-3.6.0}/src/mmgp/offload.py

@@ -1,4 +1,4 @@
-# ------------------ Memory Management 3.5.12 for the GPU Poor by DeepBeepMeep (mmgp)------------------
+# ------------------ Memory Management 3.6.0 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 3.5.12) by DeepBeepMeep ************{ENDC}{UNBOLD}")
+    print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.6.0) by DeepBeepMeep ************{ENDC}{UNBOLD}")
 
 def change_dtype(model, new_dtype, exclude_buffers = False):
     for submodule_name, submodule in model.named_modules():  
@@ -2287,9 +2287,10 @@ class offload:
         src = f"""
 def {fname}(module, *args, **kwargs):
     _ = __TYPE_CONST  # anchor type as a constant to make code object unique per class
+    nada = "{fname}"
     mgr = module._mm_manager
     mgr._pre_check(module)
-    return module._mm_forward(*args, **kwargs)
+    return module._mm_forward(*args, **kwargs) #{fname}
 """
         ns = {"__TYPE_CONST": mod_cls}
         exec(src, ns)                   # compile a new function object/code object for this class
@@ -2310,7 +2311,8 @@ def {fname}(module, *args, **kwargs):
         wrapper_fn = self._get_wrapper_for_type(type(target_module))
 
         # bind as a bound method (no partial/closures)
-        target_module.forward = types.MethodType(wrapper_fn, target_module)
+        # target_module.forward = types.MethodType(wrapper_fn, target_module)
+        target_module.forward = functools.update_wrapper(functools.partial(wrapper_fn, target_module), previous_method) 
 
     def hook_check_load_into_GPU_if_needed_default(self, target_module, model, model_id, blocks_name, previous_method,  context):
 
@@ -2345,12 +2347,12 @@ def {fname}(module, *args, **kwargs):
         if isinstance(target_module, torch.nn.Linear):
             def check_load_into_GPU_needed_linear(module, *args, **kwargs):
                 check_load_into_GPU_needed()
-                return previous_method(*args, **kwargs) 
+                return previous_method(*args, **kwargs) # linear
             check_load_into_GPU_needed_module = check_load_into_GPU_needed_linear
         else:
             def check_load_into_GPU_needed_other(module, *args, **kwargs):
                 check_load_into_GPU_needed()
-                return previous_method(*args, **kwargs) 
+                return previous_method(*args, **kwargs) # other
             check_load_into_GPU_needed_module = check_load_into_GPU_needed_other
 
         setattr(target_module, "_mm_id", model_id)
@@ -2498,7 +2500,7 @@ def {fname}(module, *args, **kwargs):
 
 
 
-def all(pipe_or_dict_of_modules, pinnedMemory = False, pinnedPEFTLora = False, partialPinning = False, loras = None, quantizeTransformer = True,  extraModelsToQuantize = None, quantizationType = qint8, budgets= 0, workingVRAM = None, asyncTransfers = True, compile = False, convertWeightsFloatTo = torch.bfloat16, perc_reserved_mem_max = 0, coTenantsMap = None, vram_safety_coefficient = 0.8, verboseLevel = -1):
+def all(pipe_or_dict_of_modules, pinnedMemory = False, pinnedPEFTLora = False, partialPinning = False, loras = None, quantizeTransformer = True,  extraModelsToQuantize = None, quantizationType = qint8, budgets= 0, workingVRAM = None, asyncTransfers = True, compile = False, convertWeightsFloatTo = torch.bfloat16, perc_reserved_mem_max = 0, coTenantsMap = None, vram_safety_coefficient = 0.8, compile_mode ="default", verboseLevel = -1):
     """Hook to a pipeline or a group of modules in order to reduce their VRAM requirements:
     pipe_or_dict_of_modules : the pipeline object or a dictionary of modules of the model
     quantizeTransformer: set True by default will quantize on the fly the video / image model
@@ -2771,8 +2773,8 @@ def all(pipe_or_dict_of_modules, pinnedMemory = False, pinnedPEFTLora = False, p
                     elif compilationInThisOne and submodule in towers_modules: 
                         self.hook_preload_blocks_for_compilation(submodule, model_id, cur_blocks_name, context = submodule_name )
                     else:
-                        if compilationInThisOne and False:
-                            self.hook_check_load_into_GPU_needed(submodule, current_model, model_id, cur_blocks_name, submodule_method, context = submodule_name )
+                        if compilationInThisOne: #and False
+                            self.hook_check_load_into_GPU_if_needed(submodule, current_model, model_id, cur_blocks_name, submodule_method, context = submodule_name )
                         else:
                             self.hook_check_load_into_GPU_if_needed_default(submodule, current_model, model_id, cur_blocks_name, submodule_method, context = submodule_name )
 
@@ -2789,7 +2791,7 @@ def all(pipe_or_dict_of_modules, pinnedMemory = False, pinnedPEFTLora = False, p
                     print(f"Pytorch compilation of model '{model_id}' is not yet supported.")
 
             for submodel in towers_modules:
-                submodel.forward= torch.compile(submodel.forward,  backend= "inductor", mode="default" ) # , fullgraph= True, mode= "reduce-overhead", "max-autotune", "max-autotune-no-cudagraphs",  
+                submodel.forward= torch.compile(submodel.forward,  backend= "inductor", mode= compile_mode) # , fullgraph= True, mode= "reduce-overhead", "max-autotune", "max-autotune-no-cudagraphs",  
                     #dynamic=True,
 
         self.tune_preloading(model_id, current_budget, towers_names)

{mmgp-3.5.12 → mmgp-3.6.0/src/mmgp.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mmgp
-Version: 3.5.12
+Version: 3.6.0
 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 3.5.12 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.6.0 for the GPU Poor by DeepBeepMeep</H2>	
 </p>