mmgp 3.5.6__tar.gz → 3.5.8__tar.gz

{mmgp-3.5.6/src/mmgp.egg-info → mmgp-3.5.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mmgp
-Version: 3.5.6
+Version: 3.5.8
 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 3.5.6 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.5.8 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

{mmgp-3.5.6 → mmgp-3.5.8}/README.md

@@ -1,6 +1,6 @@
 
 <p align="center">
-  <H2>Memory Management 3.5.6 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.5.8 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 

{mmgp-3.5.6 → mmgp-3.5.8}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "mmgp"
-version = "3.5.6"
+version = "3.5.8"
 authors = [
   { name = "deepbeepmeep", email = "deepbeepmeep@yahoo.com" },
 ]

{mmgp-3.5.6 → mmgp-3.5.8}/src/mmgp/offload.py

@@ -1,4 +1,4 @@
-# ------------------ Memory Management 3.5.6 for the GPU Poor by DeepBeepMeep (mmgp)------------------
+# ------------------ Memory Management 3.5.8 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
@@ -64,6 +64,11 @@ import psutil
 import builtins
 from accelerate import init_empty_weights
 
+import functools
+import types
+from functools import lru_cache
+import torch
+
 
 from mmgp import safetensors2
 from mmgp import profile_type
@@ -122,8 +127,6 @@ class clock:
     def format_time_gap(self):
         return f"{self.stop_time - self.start_time:.2f}s"
 
-
-
 # useful functions to move a group of tensors (to design custom offload patches)
 def move_tensors(obj, device):
     if torch.is_tensor(obj):
@@ -668,7 +671,7 @@ def _welcome():
     if welcome_displayed:
          return 
     welcome_displayed = True
-    print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.5.6) by DeepBeepMeep ************{ENDC}{UNBOLD}")
+    print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.5.8) by DeepBeepMeep ************{ENDC}{UNBOLD}")
 
 def change_dtype(model, new_dtype, exclude_buffers = False):
     for submodule_name, submodule in model.named_modules():  
@@ -1295,7 +1298,7 @@ def move_loras_to_device(model, device="cpu" ):
         if ".lora_" in k:
             m.to(device)
 
-def fast_load_transformers_model(model_path: str,  do_quantize = False, quantizationType =  qint8, pinToMemory = False, partialPinning = False, forcedConfigPath = None, defaultConfigPath = None, modelClass=None, modelPrefix = None, writable_tensors = True, verboseLevel = -1, modules = None,  return_shared_modules = None,  configKwargs ={}):
+def fast_load_transformers_model(model_path: str,  do_quantize = False, quantizationType =  qint8, pinToMemory = False, partialPinning = False, forcedConfigPath = None, defaultConfigPath = None, modelClass=None, modelPrefix = None, writable_tensors = True, verboseLevel = -1, preprocess_sd  = None, modules = None,  return_shared_modules = None,  configKwargs ={}):
     """
     quick version of .LoadfromPretrained of  the transformers library
     used to build a model and load the corresponding weights (quantized or not)
@@ -1383,13 +1386,13 @@ def fast_load_transformers_model(model_path: str,  do_quantize = False, quantiza
 
     model._config = transformer_config
             
-    load_model_data(model,model_path, do_quantize = do_quantize, quantizationType = quantizationType, pinToMemory= pinToMemory, partialPinning= partialPinning, modelPrefix = modelPrefix, writable_tensors =writable_tensors, modules = modules, return_shared_modules =  return_shared_modules, verboseLevel=verboseLevel )
+    load_model_data(model,model_path, do_quantize = do_quantize, quantizationType = quantizationType, pinToMemory= pinToMemory, partialPinning= partialPinning, modelPrefix = modelPrefix, writable_tensors =writable_tensors, preprocess_sd = preprocess_sd , modules = modules, return_shared_modules =  return_shared_modules, verboseLevel=verboseLevel )
 
     return model
 
 
 
-def load_model_data(model, file_path, do_quantize = False, quantizationType = qint8, pinToMemory = False, partialPinning = False, modelPrefix = None, writable_tensors = True,  modules = None, return_shared_modules = None, verboseLevel = -1):
+def load_model_data(model, file_path, do_quantize = False, quantizationType = qint8, pinToMemory = False, partialPinning = False, modelPrefix = None, writable_tensors = True,  preprocess_sd = None, modules = None, return_shared_modules = None, verboseLevel = -1):
     """
     Load a model, detect if it has been previously quantized using quanto and do the extra setup if necessary
     """
@@ -1506,6 +1509,9 @@ def load_model_data(model, file_path, do_quantize = False, quantizationType = qi
     full_state_dict, full_quantization_map, full_tied_weights_map = None, None, None
 
     # deal if we are trying to load just a sub part of a larger model
+    if preprocess_sd != None:
+        state_dict, quantization_map = preprocess_sd(state_dict, quantization_map)
+        
     if modelPrefix != None:
         base_model_prefix = modelPrefix + "."
         state_dict = filter_state_dict(state_dict,base_model_prefix)
@@ -1756,6 +1762,7 @@ class offload:
         global last_offload_obj
         last_offload_obj = self
 
+        self._type_wrappers = {}
         
     def add_module_to_blocks(self, model_id, blocks_name, submodule, prev_block_name, submodule_name):
 
@@ -2203,7 +2210,7 @@ class offload:
                 if len(loras_data) == 0:
                     return old_forward(*args, **kwargs)
                 else:
-                    submodule.aaa = submodule_name
+                    #submodule.aaa = submodule_name # just for debugging if uncommented will cause pytorch recompilation
                     return self._lora_linear_forward(current_model, submodule, loras_data,  *args, **kwargs)
             target_fn = lora_linear_forward
         else:
@@ -2235,10 +2242,63 @@ class offload:
 
         # need to be registered before the forward not to be break the efficiency of the compilation chain
         # it should be at the top of the compilation as this type of hook in the middle of a chain seems to break memory performance
-        target_module.register_forward_pre_hook(preload_blocks_for_compile)        
+        target_module.register_forward_pre_hook(preload_blocks_for_compile)
 
 
-    def hook_check_empty_cache_needed(self, target_module, model, model_id, blocks_name, previous_method,  context):
+
+
+    @torch._dynamo.disable
+    def _pre_check(self, module):
+        model_id    = getattr(module, "_mm_model_id", None)
+        blocks_name = getattr(module, "_mm_blocks_name", None)
+
+        self.ensure_model_loaded(model_id)
+        if blocks_name is None:
+            if self.ready_to_check_mem():
+                self.empty_cache_if_needed()
+        elif blocks_name != self.loaded_blocks[model_id] and \
+             blocks_name not in self.preloaded_blocks_per_model[model_id]:
+            self.gpu_load_blocks(model_id, blocks_name)
+
+    def _get_wrapper_for_type(self, mod_cls):
+        fn = self._type_wrappers.get(mod_cls)
+        if fn is not None:
+            return fn
+
+        # Unique function name per class -> unique compiled code object
+        fname = f"_mm_wrap_{mod_cls.__module__.replace('.', '_')}_{mod_cls.__name__}"
+
+        # Keep body minimal; all heavy/offload logic runs out-of-graph in _pre_check
+        # Include __TYPE_CONST in the code so the bytecode/consts differ per class.
+        src = f"""
+def {fname}(module, *args, **kwargs):
+    _ = __TYPE_CONST  # anchor type as a constant to make code object unique per class
+    mgr = module._mm_manager
+    mgr._pre_check(module)
+    return module._mm_forward(*args, **kwargs)
+"""
+        ns = {"__TYPE_CONST": mod_cls}
+        exec(src, ns)                   # compile a new function object/code object for this class
+        fn = ns[fname]
+        self._type_wrappers[mod_cls] = fn
+        return fn
+
+    def hook_check_load_into_GPU_if_needed(
+        self, target_module, model, model_id, blocks_name, previous_method, context
+    ):
+        # store instance data on the module (not captured by the wrapper)
+        target_module._mm_manager     = self
+        target_module._mm_model_id    = model_id
+        target_module._mm_blocks_name = blocks_name
+        target_module._mm_forward     = previous_method
+
+        # per-TYPE wrapper (unique bytecode per class, reused across instances of that class)
+        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)
+
+    def hook_check_load_into_GPU_if_needed_default(self, target_module, model, model_id, blocks_name, previous_method,  context):
 
         dtype = model._dtype
         qint4quantization =  isinstance(target_module, QModuleMixin) and  target_module.weight!= None and  target_module.weight.qtype == qint4 
@@ -2258,25 +2318,35 @@ class offload:
             target_module.forward = target_module._mm_forward
             return
 
-        def check_empty_cuda_cache(module, *args, **kwargs):
+        def check_load_into_GPU_needed():
             self.ensure_model_loaded(model_id)
             if blocks_name == None:
                 if self.ready_to_check_mem():
                     self.empty_cache_if_needed()
             elif blocks_name != self.loaded_blocks[model_id] and blocks_name not in self.preloaded_blocks_per_model[model_id]:
                 self.gpu_load_blocks(model_id, blocks_name)
-            if qint4quantization and dtype !=None:
-                args, kwargs = self.move_args_to_gpu(dtype, *args, **kwargs)
-
-            return previous_method(*args, **kwargs) 
+            # if qint4quantization and dtype !=None:
+            #     args, kwargs = self.move_args_to_gpu(dtype, *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) 
+            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) 
+            check_load_into_GPU_needed_module = check_load_into_GPU_needed_other
 
         setattr(target_module, "_mm_id", model_id)
         setattr(target_module, "_mm_forward", previous_method)
 
-        setattr(target_module, "forward", functools.update_wrapper(functools.partial(check_empty_cuda_cache, target_module), previous_method) )
+        setattr(target_module, "forward", functools.update_wrapper(functools.partial(check_load_into_GPU_needed_module, target_module), previous_method) )
+        # target_module.register_forward_pre_hook(check_empty_cuda_cache)
 
         
-    def hook_change_module(self, target_module, model, model_id, module_id, previous_method):
+    def hook_change_module(self, target_module, model, model_id, module_id, previous_method, previous_method_name ):
         if hasattr(target_module, "_lock_dtype"):
             dtype = target_module._lock_dtype 
         else:
@@ -2289,16 +2359,17 @@ class offload:
                 args, kwargs = self.move_args_to_gpu(dtype, *args, **kwargs)
             return previous_method(*args, **kwargs) 
   
-        if hasattr(target_module, "_mm_id"):
+        if hasattr(target_module, "_mm_" + previous_method_name):
             return
-        setattr(target_module, "_mm_id", model_id)
+        setattr(target_module, "_mm_Id", model_id)
+        setattr(target_module, "_mm_" + previous_method_name, previous_method)
 
-        setattr(target_module, "forward", functools.update_wrapper(functools.partial(check_change_module, target_module), previous_method) )
+        setattr(target_module, previous_method_name, functools.update_wrapper(functools.partial(check_change_module, target_module), previous_method) )
 
         if not self.verboseLevel >=1:
             return
 
-        if module_id == None or module_id =='':
+        if previous_method_name =="forward" and (module_id == None or module_id ==''):
             model_name = model._get_name()
             print(f"Hooked to model '{model_id}' ({model_name})")
 
@@ -2605,19 +2676,7 @@ def all(pipe_or_dict_of_modules, pinnedMemory = False, pinnedPEFTLora = False, p
     for model_id in models: 
         current_model: torch.nn.Module = models[model_id] 
         towers_names, towers_modules = _detect_main_towers(current_model)
-        # compile main iterative modules stacks ("towers")
         compilationInThisOne = compileAllModels or model_id in modelsToCompile 
-        if compilationInThisOne:
-            if self.verboseLevel>=1:
-                if len(towers_modules)>0:
-                    formated_tower_names = [name + '*' for name in towers_names]
-                    print(f"Pytorch compilation of '{model_id}' is scheduled for these modules : {formated_tower_names}.")
-                else:
-                    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",  
-                    #dynamic=True,
                 
         if pinAllModels or model_id in modelsToPin:
             if hasattr(current_model,"_already_pinned"):
@@ -2661,24 +2720,42 @@ def all(pipe_or_dict_of_modules, pinnedMemory = False, pinnedPEFTLora = False, p
                         cur_blocks_prefix, prev_blocks_name, cur_blocks_seq, is_mod_seq = pre, None, num, False
                         cur_blocks_name = submodule_name
                         # print(f"new block: {model_id}/{cur_blocks_name} - {submodule_name}")
-                          
- 
-            if hasattr(submodule, "forward"):
-                # if  any_lora and isinstance(submodule, ( torch.nn.Linear, torch.nn.Conv3d, torch.nn.LayerNorm)):
-                if  any_lora and  hasattr(submodule,"weight"):
+            top_submodule = len(submodule_name.split("."))==1
+            offload_hooks = submodule._offload_hooks if hasattr(submodule, "_offload_hooks") else [] 
+            assert top_submodule or len(offload_hooks) == 0, "custom offload hooks can only be set at the of the module"
+            submodule_method_names = ["forward"] +  offload_hooks
+            for submodule_method_name in submodule_method_names:
+                if not hasattr(submodule, submodule_method_name ): continue
+                if submodule_method_name == "forward" and any_lora and hasattr(submodule,"weight"):
                     submodule_method = self.hook_lora(submodule, current_model, model_id, loras_model_data, loras_model_shortcuts, submodule_name)                
                 else:
-                    submodule_method = getattr(submodule, "forward")
-                if callable(submodule_method):   
-                    if len(submodule_name.split("."))==1:
-                        self.hook_change_module(submodule, current_model, model_id, submodule_name, submodule_method)
+                    submodule_method = getattr(submodule, submodule_method_name)
+                if callable(submodule_method):
+                    if top_submodule and cur_blocks_name is None:
+                        self.hook_change_module(submodule, current_model, model_id, submodule_name, submodule_method, submodule_method_name)
                     elif compilationInThisOne and submodule in towers_modules: 
                         self.hook_preload_blocks_for_compilation(submodule, model_id, cur_blocks_name, context = submodule_name )
                     else:
-                        self.hook_check_empty_cache_needed(submodule, current_model, model_id, cur_blocks_name, submodule_method, context = submodule_name )
-                
-                self.add_module_to_blocks(model_id, cur_blocks_name, submodule, prev_blocks_name, submodule_name)
+                        if compilationInThisOne and False:
+                            self.hook_check_load_into_GPU_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 )
 
+                    self.add_module_to_blocks(model_id, cur_blocks_name, submodule, prev_blocks_name, submodule_name)
+
+
+        # compile main iterative modules stacks ("towers")
+        if compilationInThisOne:
+            if self.verboseLevel>=1:
+                if len(towers_modules)>0:
+                    formated_tower_names = [name + '*' for name in towers_names]
+                    print(f"Pytorch compilation of '{model_id}' is scheduled for these modules : {formated_tower_names}.")
+                else:
+                    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",  
+                    #dynamic=True,
 
         self.tune_preloading(model_id, current_budget, towers_names)
         self.parameters_ref  = {} 

{mmgp-3.5.6 → mmgp-3.5.8/src/mmgp.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mmgp
-Version: 3.5.6
+Version: 3.5.8
 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 3.5.6 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.5.8 for the GPU Poor by DeepBeepMeep</H2>	
 </p>