mmgp 3.0.3__tar.gz → 3.0.9__tar.gz

{mmgp-3.0.3/src/mmgp.egg-info → mmgp-3.0.9}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.2
 Name: mmgp
-Version: 3.0.3
+Version: 3.0.9
 Summary: Memory Management for the GPU Poor
 Author-email: deepbeepmeep <deepbeepmeep@yahoo.com>
 License:                     GNU GENERAL PUBLIC LICENSE
@@ -13,10 +13,11 @@ Requires-Dist: optimum-quanto
 Requires-Dist: accelerate
 Requires-Dist: safetensors
 Requires-Dist: psutil
+Requires-Dist: peft
 
 
 <p align="center">
-  <H2>Memory Management 3.0 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.0.9 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 
@@ -38,8 +39,9 @@ Each profile may use a combination of the following:
 - Ability to pin models to reserved RAM to accelerate transfers to VRAM
 - Async transfers to VRAM to avoid a pause when loading a new slice of a model
 - Automated on the fly quantization or ability to load pre quantized models
-- support for pytorch compilation on Linux and WSL (not supported so far on pure Windows).
-
+- Pretrained Lora support with low RAM requirements
+- Support for pytorch compilation on Linux and WSL (supported on pure Windows but requires a complex Triton Installation).
+- 
 ## Installation
 First you need to install the module in your current project with:
 ```shell
@@ -105,20 +107,22 @@ If you are short on RAM and plan to work with quantized models, it is recommende
 ##  Going further
 
 The module includes several tools to package a light version of your favorite video / image generator:
-- *save_model(model, file_path, do_quantize = False, quantization_type = qint8 )*\
+- *save_model(model, file_path, do_quantize = False, quantizationType = qint8 )*\
 Save tensors of a model already loaded in memory in a safetensor format (much faster to reload). You can save it in a quantized format (default qint8 quantization recommended).
 The resulting safetensor file will contain extra fields in its metadata such as the quantization map and its configuration, so you will be able to move the file around without files such as *config.json* or *file_map.json*.
 You will need *load_model_data* or *fast_load_transformers_model* to read the file again . You may also load it using the default *safetensor* librar however you will need to provide in the same directory any complementary file that are usually requested (for instance *config.json*)
 
-- *load_model_data(model, file_path: str, do_quantize = False, quantization_type = qint8, pinToRAM = False, partialPin = False)*\
+- *load_model_data(model, file_path: str, do_quantize = False, quantizationType = qint8, pinToRAM = False, partialPin = False)*\
 Load the tensors data of a model in RAM of a model already initialized with no data. Detect and handle quantized models saved previously with *save_model*.A model can also be quantized on the fly while being loaded. The model which is loaded can be pinned to RAM while it is loaded, this is more RAM efficient than pinning tensors later using *offline.all* or *offline.profile*
 
-- *fast_load_transformers_model(model_path: str, do_quantize = False, quantization_type = qint8, pinToRAM = False, partialPin = False)*\
+- *fast_load_transformers_model(model_path: str, do_quantize = False, quantizationType = qint8, pinToRAM = False, partialPin = False)*\
 Initialize (build the model hierarchy in memory) and fast load the corresponding tensors of a 'transformers' or 'diffusers' library model.
 The advantages over the original *from_pretrained* method is that a full model can fit into a single file with a filename of your choosing (thefore you can have multiple 'transformers' versions of the same model in the same directory) and prequantized models are processed in a transparent way. 
 Last but not least, you can also on the fly pin to RAM the whole model or the most important part of it (partialPin = True) in a more efficient way (faster and requires less RAM) than if you did through *offload.all* or *offload.profile*.
 
-
+- *load_loras_into_model(model, lora_path, lora_multi)
+Load in a model a list of Lora described by a list of path *lora_path* and a list of *weights coefficients*.
+The Lora file must be in the *diffusers* format. This function works also on non diffusers models. However if there is already an official Lora support for a model it is recommended to use the official diffusers functions.
 
 The typical workflow wil be:
 1) temporarly insert the *save_model* function just after a model has been fully loaded to save a copy of the model / quantized model.

{mmgp-3.0.3 → mmgp-3.0.9}/README.md

@@ -1,6 +1,6 @@
 
 <p align="center">
-  <H2>Memory Management 3.0 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.0.9 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 
@@ -22,8 +22,9 @@ Each profile may use a combination of the following:
 - Ability to pin models to reserved RAM to accelerate transfers to VRAM
 - Async transfers to VRAM to avoid a pause when loading a new slice of a model
 - Automated on the fly quantization or ability to load pre quantized models
-- support for pytorch compilation on Linux and WSL (not supported so far on pure Windows).
-
+- Pretrained Lora support with low RAM requirements
+- Support for pytorch compilation on Linux and WSL (supported on pure Windows but requires a complex Triton Installation).
+- 
 ## Installation
 First you need to install the module in your current project with:
 ```shell
@@ -89,20 +90,22 @@ If you are short on RAM and plan to work with quantized models, it is recommende
 ##  Going further
 
 The module includes several tools to package a light version of your favorite video / image generator:
-- *save_model(model, file_path, do_quantize = False, quantization_type = qint8 )*\
+- *save_model(model, file_path, do_quantize = False, quantizationType = qint8 )*\
 Save tensors of a model already loaded in memory in a safetensor format (much faster to reload). You can save it in a quantized format (default qint8 quantization recommended).
 The resulting safetensor file will contain extra fields in its metadata such as the quantization map and its configuration, so you will be able to move the file around without files such as *config.json* or *file_map.json*.
 You will need *load_model_data* or *fast_load_transformers_model* to read the file again . You may also load it using the default *safetensor* librar however you will need to provide in the same directory any complementary file that are usually requested (for instance *config.json*)
 
-- *load_model_data(model, file_path: str, do_quantize = False, quantization_type = qint8, pinToRAM = False, partialPin = False)*\
+- *load_model_data(model, file_path: str, do_quantize = False, quantizationType = qint8, pinToRAM = False, partialPin = False)*\
 Load the tensors data of a model in RAM of a model already initialized with no data. Detect and handle quantized models saved previously with *save_model*.A model can also be quantized on the fly while being loaded. The model which is loaded can be pinned to RAM while it is loaded, this is more RAM efficient than pinning tensors later using *offline.all* or *offline.profile*
 
-- *fast_load_transformers_model(model_path: str, do_quantize = False, quantization_type = qint8, pinToRAM = False, partialPin = False)*\
+- *fast_load_transformers_model(model_path: str, do_quantize = False, quantizationType = qint8, pinToRAM = False, partialPin = False)*\
 Initialize (build the model hierarchy in memory) and fast load the corresponding tensors of a 'transformers' or 'diffusers' library model.
 The advantages over the original *from_pretrained* method is that a full model can fit into a single file with a filename of your choosing (thefore you can have multiple 'transformers' versions of the same model in the same directory) and prequantized models are processed in a transparent way. 
 Last but not least, you can also on the fly pin to RAM the whole model or the most important part of it (partialPin = True) in a more efficient way (faster and requires less RAM) than if you did through *offload.all* or *offload.profile*.
 
-
+- *load_loras_into_model(model, lora_path, lora_multi)
+Load in a model a list of Lora described by a list of path *lora_path* and a list of *weights coefficients*.
+The Lora file must be in the *diffusers* format. This function works also on non diffusers models. However if there is already an official Lora support for a model it is recommended to use the official diffusers functions.
 
 The typical workflow wil be:
 1) temporarly insert the *save_model* function just after a model has been fully loaded to save a copy of the model / quantized model.

{mmgp-3.0.3 → mmgp-3.0.9}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "mmgp"
-version = "3.0.3"
+version = "3.0.9"
 authors = [
   { name = "deepbeepmeep", email = "deepbeepmeep@yahoo.com" },
 ]
@@ -13,6 +13,7 @@ dependencies = [
   "optimum-quanto",
   "accelerate",
   "safetensors",
-  "psutil"
+  "psutil",
+  "peft"
 ]
 

{mmgp-3.0.3 → mmgp-3.0.9}/src/mmgp/offload.py

@@ -61,10 +61,22 @@ import sys
 import os
 import json
 import psutil
+try:    
+    from diffusers.utils.peft_utils import set_weights_and_activate_adapters, get_peft_kwargs
+except:
+    set_weights_and_activate_adapters = None
+    get_peft_kwargs = None
+    pass
+try:    
+    from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
+except:
+    inject_adapter_in_model = None
+    pass
+
 from mmgp import safetensors2
 from mmgp import profile_type
 
-from optimum.quanto import freeze, qfloat8, qint8, quantize, QModuleMixin, QTensor, WeightQBytesTensor, quantize_module 
+from optimum.quanto import freeze,  qfloat8, qint4 , qint8, quantize, QModuleMixin, QTensor,  quantize_module 
 
 
 
@@ -127,6 +139,9 @@ def move_tensors(obj, device):
         return _list
     else:
         raise TypeError("Tensor or list / dict of tensors expected")
+def _get_module_name(v):
+    return v.__module__.lower()
+
 
 def _compute_verbose_level(level):
     if level <0:        
@@ -263,7 +278,13 @@ def _pin_to_memory(model, model_id, partialPinning = False, perc_reserved_mem_ma
     # print(f"num params to pin {model_id}: {len(params_list)}")
     for p in params_list:
         if isinstance(p, QTensor):
-            length = torch.numel(p._data) * p._data.element_size() + torch.numel(p._scale) * p._scale.element_size() 
+            if p._qtype == qint4:
+                if hasattr(p,"_scale_shift"):
+                    length = torch.numel(p._data._data) * p._data._data.element_size() + torch.numel(p._scale_shift) * p._scale_shift.element_size() 
+                else:
+                    length = torch.numel(p._data._data) * p._data._data.element_size() + torch.numel(p._scale) * p._scale.element_size() + torch.numel(p._shift) * p._shift.element_size()                     
+            else:
+                length = torch.numel(p._data) * p._data.element_size() + torch.numel(p._scale) * p._scale.element_size() 
         else:
             length = torch.numel(p.data) * p.data.element_size() 
 
@@ -306,10 +327,22 @@ def _pin_to_memory(model, model_id, partialPinning = False, perc_reserved_mem_ma
         if big_tensor_no>=0 and big_tensor_no < last_big_tensor:
             current_big_tensor = big_tensors[big_tensor_no]
             if isinstance(p, QTensor):
-                length1 = torch.numel(p._data) * p._data.element_size() 
-                p._data = _move_to_pinned_tensor(p._data, current_big_tensor, offset, length1)
-                length2 = torch.numel(p._scale) * p._scale.element_size() 
-                p._scale = _move_to_pinned_tensor(p._scale, current_big_tensor, offset + length1, length2)
+                if p._qtype == qint4:
+                    length1 = torch.numel(p._data._data) * p._data._data.element_size()
+                    p._data._data =  _move_to_pinned_tensor(p._data._data, current_big_tensor, offset, length1)
+                    if hasattr(p,"_scale_shift"):
+                        length2 = torch.numel(p._scale_shift) * p._scale_shift.element_size() 
+                        p._scale_shift = _move_to_pinned_tensor(p._scale_shift, current_big_tensor, offset + length1, length2)
+                    else:
+                        length2 = torch.numel(p._scale) * p._scale.element_size() 
+                        p._scale = _move_to_pinned_tensor(p._scale, current_big_tensor, offset + length1, length2)
+                        length3 = torch.numel(p._shift) * p._shift.element_size() 
+                        p._shift = _move_to_pinned_tensor(p._shift, current_big_tensor, offset + length1 + length2, length3)
+                else:
+                    length1 = torch.numel(p._data) * p._data.element_size() 
+                    p._data = _move_to_pinned_tensor(p._data, current_big_tensor, offset, length1)
+                    length2 = torch.numel(p._scale) * p._scale.element_size() 
+                    p._scale = _move_to_pinned_tensor(p._scale, current_big_tensor, offset + length1, length2)
             else:
                 length = torch.numel(p.data) * p.data.element_size() 
                 p.data = _move_to_pinned_tensor(p.data, current_big_tensor, offset, length)
@@ -339,98 +372,6 @@ def _welcome():
     print(f"{BOLD}{HEADER}************ Memory Management for the GPU Poor (mmgp 3.0) by DeepBeepMeep ************{ENDC}{UNBOLD}")
 
 
-# def _pin_to_memory_sd(model, sd, model_id, partialPinning = False, perc_reserved_mem_max = 0, verboseLevel = 1):
-#     if  verboseLevel>=1 :
-#         if partialPinning:
-#             print(f"Partial pinning to reserved RAM of data of file '{model_id}' while loading it")
-#         else:            
-#             print(f"Pinning data to reserved RAM of file '{model_id}' while loading it")
-
-#     max_reservable_memory = _get_max_reservable_memory(perc_reserved_mem_max)
-#     if partialPinning:
-#         towers_names, _ = _detect_main_towers(model)
-#         towers_names = [n +"." for n in towers_names]
-
-#     BIG_TENSOR_MAX_SIZE = 2**28 # 256 MB
-#     current_big_tensor_size = 0
-#     big_tensor_no  = 0
-#     big_tensors_sizes = []
-#     tensor_map_indexes = []
-#     total_tensor_bytes = 0
-
-#     for k,t in  sd.items():
-#         include = True
-#         # if isinstance(p, QTensor):
-#         #     length = torch.numel(p._data) * p._data.element_size() + torch.numel(p._scale) * p._scale.element_size() 
-#         # else:
-#         #     length = torch.numel(p.data) * p.data.element_size() 
-#         length = torch.numel(t) * t.data.element_size() 
-
-#         if partialPinning:
-#             include = any(k.startswith(pre) for pre in towers_names) if partialPinning else True
-
-#         if include:
-#             if current_big_tensor_size + length > BIG_TENSOR_MAX_SIZE:
-#                 big_tensors_sizes.append(current_big_tensor_size)
-#                 current_big_tensor_size = 0
-#                 big_tensor_no += 1
-#             tensor_map_indexes.append((big_tensor_no, current_big_tensor_size, length  ))
-#             current_big_tensor_size += length
-#         else:
-#             tensor_map_indexes.append((-1, 0, 0  ))
-#         total_tensor_bytes += length
-
-#     big_tensors_sizes.append(current_big_tensor_size)
-
-#     big_tensors = []
-#     last_big_tensor = 0
-#     total = 0
-    
-
-#     for size in big_tensors_sizes:
-#         try:
-#             currrent_big_tensor = torch.empty( size, dtype= torch.uint8, pin_memory=True)
-#             big_tensors.append(currrent_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})")
-#             break
-
-#         last_big_tensor += 1
-#         total += size
-
-
-#     tensor_no = 0
-#     for k,t in sd.items():
-#         big_tensor_no, offset, length = tensor_map_indexes[tensor_no]
-#         if big_tensor_no>=0 and big_tensor_no < last_big_tensor:
-#             current_big_tensor = big_tensors[big_tensor_no]
-#             # if isinstance(p, QTensor):
-#             #     length1 = torch.numel(p._data) * p._data.element_size() 
-#             #     p._data = _move_to_pinned_tensor(p._data, current_big_tensor, offset, length1)
-#             #     length2 = torch.numel(p._scale) * p._scale.element_size() 
-#             #     p._scale = _move_to_pinned_tensor(p._scale, current_big_tensor, offset + length1, length2)
-#             # else:
-#             #     length = torch.numel(p.data) * p.data.element_size() 
-#             #     p.data = _move_to_pinned_tensor(p.data, current_big_tensor, offset, length)
-#             length = torch.numel(t) * t.data.element_size() 
-#             t = _move_to_pinned_tensor(t, current_big_tensor, offset, length)
-#             sd[k] = t
-#         tensor_no += 1
-
-#     global total_pinned_bytes
-#     total_pinned_bytes += total
-
-#     if verboseLevel >=1:
-#         if total_tensor_bytes == total:        
-#             print(f"The whole model was pinned to reserved RAM: {last_big_tensor} large blocks spread across {total/ONE_MB:.2f} MB")
-#         else:
-#             print(f"{total/ONE_MB:.2f} MB were pinned to reserved RAM out of {total_tensor_bytes/ONE_MB:.2f} MB")
-
-#     model._already_pinned = True
-
-
-#     return 
-
 def  _quantize_dirty_hack(model):
     # dirty hack: add a hook on state_dict() to return a fake non quantized state_dict if called by Lora Diffusers initialization functions
     setattr( model, "_real_state_dict", model.state_dict)
@@ -536,10 +477,14 @@ def _quantize(model_to_quantize, weights=qint8, verboseLevel = 1, threshold = 10
     prev_blocks_prefix = None
 
     if hasattr(model_to_quantize, "_quanto_map"):
+        for k, entry in model_to_quantize._quanto_map.items():
+            weights  =  entry["weights"]
+            print(f"Model '{model_id}' is already quantized to format '{weights}'")
+            return False
         print(f"Model '{model_id}' is already quantized")
         return False
-    
-    print(f"Quantization of model '{model_id}' started")
+
+    print(f"Quantization of model '{model_id}' started to format '{weights}'")
 
     for submodule_name, submodule in model_to_quantize.named_modules():  
         if isinstance(submodule, QModuleMixin):
@@ -594,18 +539,18 @@ def _quantize(model_to_quantize, weights=qint8, verboseLevel = 1, threshold = 10
     if verboseLevel >=2:
         print(f"Total Excluded {total_excluded/ONE_MB:.1f} MB oF {total_size/ONE_MB:.1f} that is {perc_excluded*100:.2f}%")
     if perc_excluded >= 0.10:
-        print(f"Too many many modules are excluded, there is something wrong with the selection, switch back to full quantization.")
+        print(f"Too many modules are excluded, there is something wrong with the selection, switch back to full quantization.")
         exclude_list = None
 
 
     #quantize(model_to_quantize,weights, exclude= exclude_list)
-    pass
+
     for name, m in model_to_quantize.named_modules():
         if exclude_list is None or not any( name == module_name for module_name in exclude_list):
             _quantize_submodule(model_to_quantize, name, m, weights=weights, activations=None, optimizer=None)
 
-    # force read non quantized parameters so that their lazy tensors and corresponding mmap are released
-    # otherwise we may end up to keep in memory both the quantized and the non quantize model
+    # force to read non quantized parameters so that their lazy tensors and corresponding mmap are released
+    # otherwise we may end up keeping in memory both the quantized and the non quantize model
     for m in model_to_quantize.modules():
         # do not read quantized weights (detected them directly or behind an adapter)
         if isinstance(m, QModuleMixin) or hasattr(m, "base_layer") and  isinstance(m.base_layer, QModuleMixin): 
@@ -620,12 +565,16 @@ def _quantize(model_to_quantize, weights=qint8, verboseLevel = 1, threshold = 10
             b.data = b.data + 0 
 
     
+
     freeze(model_to_quantize)
     torch.cuda.empty_cache()
     gc.collect()         
     quantization_map = _quantization_map(model_to_quantize)
     model_to_quantize._quanto_map = quantization_map
 
+    if hasattr(model_to_quantize, "_already_pinned"):
+        delattr(model_to_quantize, "_already_pinned")
+
     _quantize_dirty_hack(model_to_quantize)
 
     print(f"Quantization of model '{model_id}' done")
@@ -684,15 +633,25 @@ class offload:
 
         for k,p in submodule.named_parameters(recurse=False):
             if isinstance(p, QTensor):
-                blocks_params.append( (submodule, k, p._data, p._scale) )
-                blocks_params_size += p._data.nbytes
-                blocks_params_size += p._scale.nbytes
+                blocks_params.append( (submodule, k, p ) )
+
+                if p._qtype == qint4:
+                    if hasattr(p,"_scale_shift"):
+                        blocks_params_size += torch.numel(p._scale_shift) * p._scale_shift.element_size()
+                        blocks_params_size += torch.numel(p._data._data) * p._data._data.element_size()
+                    else:
+                        blocks_params_size += torch.numel(p._scale) * p._scale.element_size()
+                        blocks_params_size += torch.numel(p._shift) * p._shift.element_size()
+                        blocks_params_size += torch.numel(p._data._data) * p._data._data.element_size()
+                else:
+                    blocks_params_size += torch.numel(p._scale) * p._scale.element_size()
+                    blocks_params_size += torch.numel(p._data) * p._data.element_size()
             else:
-                blocks_params.append( (submodule, k, p.data, None) )
-                blocks_params_size += p.data.nbytes
+                blocks_params.append( (submodule, k, p  ) )
+                blocks_params_size += torch.numel(p.data) * p.data.element_size()
 
         for k, p in submodule.named_buffers(recurse=False):
-            blocks_params.append( (submodule, k, p.data, None) )
+            blocks_params.append( (submodule, k, p) )
             blocks_params_size += p.data.nbytes
 
 
@@ -710,34 +669,28 @@ class offload:
                 return False    
         return True
 
-    def gpu_load_blocks(self, model_id, blocks_name, async_load = False):
+
+    def gpu_load_blocks(self, model_id, blocks_name):
         # cl = clock.start()
 
         if blocks_name != None:
             self.loaded_blocks[model_id] = blocks_name           
 
         entry_name = model_id if blocks_name is None else model_id + "/" + blocks_name
-
-        def cpu_to_gpu(stream_to_use, blocks_params, record_for_stream = None):
+        
+        def cpu_to_gpu(stream_to_use, blocks_params): #, record_for_stream = None
             with torch.cuda.stream(stream_to_use):
                 for param in blocks_params:
-                    parent_module, n, data, scale  = param
-                    p = getattr(parent_module, n)
-                    if isinstance(p, QTensor):
-                        q = WeightQBytesTensor.create(p.qtype, p.axis, p.size(), p.stride(), data.cuda(non_blocking=True), scale.cuda(non_blocking=True), activation_qtype=p.activation_qtype, requires_grad=p.requires_grad )
-                        #q = p.to("cuda", non_blocking=True)
-                        q = torch.nn.Parameter(q , requires_grad=False)
-                        setattr(parent_module, n , q)
-                        del p
-                    else:
-                        p.data = p.data.cuda(non_blocking=True) 
-                    
-                    if record_for_stream != None:
-                        if isinstance(p, QTensor):
-                            q._data.record_stream(record_for_stream)
-                            q._scale.record_stream(record_for_stream)
-                        else:
-                            p.data.record_stream(record_for_stream)
+                    parent_module, n, p  = param
+                    q = p.to("cuda", non_blocking=True)
+                    q = torch.nn.Parameter(q , requires_grad=False)
+                    setattr(parent_module, n , q)
+                    # if record_for_stream != None:
+                    #     if isinstance(p, QTensor):
+                    #         q._data.record_stream(record_for_stream)
+                    #         q._scale.record_stream(record_for_stream)
+                    #     else:
+                    #         p.data.record_stream(record_for_stream)
 
 
         if self.verboseLevel >=2:
@@ -776,19 +729,10 @@ class offload:
             print(f"Unloading model {blocks_name} ({model_name}) from GPU")
  
         blocks_params = self.blocks_of_modules[blocks_name]
-
         for param in blocks_params:
-            parent_module, n, data, scale = param
-            p = getattr(parent_module, n) 
-            if isinstance(p, QTensor):
-                # need to change the parameter directly from the module as it can't be swapped in place due to a memory leak in the pytorch compiler
-                q = WeightQBytesTensor.create(p.qtype, p.axis, p.size(), p.stride(), data, scale, activation_qtype=p.activation_qtype, requires_grad=p.requires_grad )
-                q = torch.nn.Parameter(q , requires_grad=False)
-                setattr(parent_module, n , q)
-                del p
-            else:
-                p.data = data
-
+            parent_module, n, p  = param
+            q = torch.nn.Parameter(p , requires_grad=False)
+            setattr(parent_module, n , q)
         # cl.stop()
         # print(f"unload time: {cl.format_time_gap()}")
 
@@ -824,8 +768,8 @@ class offload:
             if torch.is_tensor(arg):    
                 if arg.dtype == torch.float32:
                     arg = arg.to(torch.bfloat16).cuda(non_blocking=True)             
-                else:
-                    arg = arg.cuda(non_blocking=True)             
+                elif not arg.is_cuda:
+                    arg = arg.cuda(non_blocking=True)
             new_args.append(arg)
 
         for k in kwargs:
@@ -833,7 +777,7 @@ class offload:
             if torch.is_tensor(arg):
                 if arg.dtype == torch.float32:
                     arg = arg.to(torch.bfloat16).cuda(non_blocking=True)             
-                else:
+                elif not arg.is_cuda:
                     arg = arg.cuda(non_blocking=True)             
             new_kwargs[k]= arg
         
@@ -897,6 +841,10 @@ class offload:
 
     def hook_check_empty_cache_needed(self, target_module, model_id,blocks_name, previous_method,  context):
 
+        qint4quantization =  isinstance(target_module, QModuleMixin) and  target_module.weight!= None and  target_module.weight.qtype == qint4 
+        if qint4quantization:
+            pass
+
         def check_empty_cuda_cache(module, *args, **kwargs):
             # if self.ready_to_check_mem():
             #     self.empty_cache_if_needed()
@@ -912,6 +860,8 @@ class offload:
                             self.empty_cache_if_needed()
                     self.loaded_blocks[model_id] = blocks_name
                     self.gpu_load_blocks(model_id, blocks_name)
+            if qint4quantization:
+                args, kwargs = self.move_args_to_gpu(*args, **kwargs)
 
             return previous_method(*args, **kwargs) 
 
@@ -970,7 +920,105 @@ class offload:
     #     for module in parent_module.components.items():
     #         self.unhook_module(module)
 
-def fast_load_transformers_model(model_path: str, do_quantize = False, quantization_type = qint8, pinToMemory = False, partialPinning = False,  verboseLevel = -1):
+import torch
+
+
+
+
+def load_loras_into_model(model, lora_path, lora_multi = None, verboseLevel = -1):
+    verboseLevel = _compute_verbose_level(verboseLevel)
+
+    if inject_adapter_in_model == None or set_weights_and_activate_adapters == None or  get_peft_kwargs == None:
+        raise Exception("Unable to load Lora, missing 'peft' and / or 'diffusers' modules")
+    
+    if not isinstance(lora_path, list):
+        lora_path = [lora_path]
+    
+    if lora_multi is None:
+        lora_multi = [1. for _ in lora_path]
+
+    for i, path in enumerate(lora_path):
+        adapter_name = str(i)
+
+        state_dict = safetensors2.torch_load_file(path)
+        
+        keys = list(state_dict.keys())
+        if len(keys) == 0:
+            raise Exception(f"Empty Lora '{path}'")
+
+
+        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]
+        if not any( prefix.startswith(some_prefix) for some_prefix 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."
+            raise Exception(msg)
+
+        transformer = model
+
+        transformer_keys = [k for k in keys if k.startswith(prefix)]
+        state_dict = {
+            k.replace(f"{prefix}.", ""): v for k, v in state_dict.items() if k in transformer_keys
+        }
+
+        sd_keys = state_dict.keys()
+        if len(sd_keys) == 0:
+            print(f"No compatible weight was found in Lora file '{path}'. Please check that it is compatible with the Diffusers format.")
+            return
+
+        # is_correct_format = all("lora" in key for key in state_dict.keys())
+
+
+
+
+        # check with first key if is not in peft format
+        # first_key = next(iter(state_dict.keys()))
+        # if "lora_A" not in first_key:
+        #     state_dict = convert_unet_state_dict_to_peft(state_dict)
+
+        if adapter_name in getattr(transformer, "peft_config", {}):
+            raise ValueError(
+                f"Adapter name {adapter_name} already in use in the transformer - please select a new adapter name."
+            )
+
+        rank = {}
+        for key, val in state_dict.items():
+            if "lora_B" in key:
+                rank[key] = val.shape[1]
+
+        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}
+
+        lora_config_kwargs = get_peft_kwargs(rank, network_alpha_dict=network_alphas, peft_state_dict=state_dict)
+        
+        lora_config = LoraConfig(**lora_config_kwargs)
+        peft_kwargs = {}        
+        peft_kwargs["low_cpu_mem_usage"] = True
+        inject_adapter_in_model(lora_config, model, adapter_name=adapter_name, **peft_kwargs)
+
+        incompatible_keys = set_peft_model_state_dict(transformer, state_dict, adapter_name, **peft_kwargs)
+
+        warn_msg = ""
+        if incompatible_keys is not None:
+            # Check only for unexpected keys.
+            unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+            if unexpected_keys:
+                pass
+        if verboseLevel >=1:
+            print(f"Lora '{path}' was loaded in model '{_get_module_name(model)}'")
+    set_weights_and_activate_adapters(model,[ str(i) for i in range(len(lora_multi))], lora_multi)
+
+
+def fast_load_transformers_model(model_path: str, do_quantize = False, quantizationType =  qint8, pinToMemory = False, partialPinning = False, verboseLevel = -1):
     """
     quick version of .LoadfromPretrained of  the transformers library
     used to build a model and load the corresponding weights (quantized or not)
@@ -1040,13 +1088,13 @@ def fast_load_transformers_model(model_path: str, do_quantize = False, quantizat
 
     model._config = transformer_config
             
-    load_model_data(model,model_path, do_quantize = do_quantize, quantization_type = quantization_type, pinToMemory= pinToMemory, partialPinning= partialPinning, verboseLevel=verboseLevel )
+    load_model_data(model,model_path, do_quantize = do_quantize, quantizationType = quantizationType, pinToMemory= pinToMemory, partialPinning= partialPinning, verboseLevel=verboseLevel )
 
     return model
 
 
 
-def load_model_data(model, file_path: str, do_quantize = False, quantization_type = qint8, pinToMemory = False, partialPinning = False, verboseLevel = -1):
+def load_model_data(model, file_path: str, do_quantize = False, quantizationType = qint8, pinToMemory = False, partialPinning = False, verboseLevel = -1):
     """
     Load a model, detect if it has been previously quantized using quanto and do the extra setup if necessary
     """
@@ -1067,14 +1115,6 @@ def load_model_data(model, file_path: str, do_quantize = False, quantization_typ
             state_dict = state_dict["module"]
     else:
         state_dict, metadata = _safetensors_load_file(file_path)
-
-
-        # if pinToMemory:
-        #     _pin_to_memory_sd(model,state_dict, file_path, partialPinning = partialPinning, perc_reserved_mem_max = perc_reserved_mem_max, verboseLevel = verboseLevel)
-
-        # with safetensors2.safe_open(file_path) as f:
-        #     metadata = f.metadata() 
-
             
         if metadata is None:
             quantization_map = None
@@ -1109,7 +1149,7 @@ def load_model_data(model, file_path: str, do_quantize = False, quantization_typ
 
     if do_quantize:
         if quantization_map is None:
-            if _quantize(model, quantization_type, verboseLevel=verboseLevel, model_id=file_path):
+            if _quantize(model, quantizationType, verboseLevel=verboseLevel, model_id=file_path):
                 quantization_map = model._quanto_map  
         else:
             if verboseLevel >=1:
@@ -1120,7 +1160,7 @@ def load_model_data(model, file_path: str, do_quantize = False, quantization_typ
 
     return
 
-def save_model(model, file_path, do_quantize = False, quantization_type = qint8, verboseLevel = -1 ):
+def save_model(model, file_path, do_quantize = False, quantizationType = qint8, verboseLevel = -1 ):
     """save the weights of a model and quantize them if requested
     These weights can be loaded again using 'load_model_data'
     """       
@@ -1147,7 +1187,7 @@ def save_model(model, file_path, do_quantize = False, quantization_type = qint8,
                 config= json.loads(text)
 
     if do_quantize:
-        _quantize(model, weights=quantization_type, model_id=file_path)
+        _quantize(model, weights=quantizationType, model_id=file_path)
     
     quantization_map = getattr(model, "_quanto_map", None)
 
@@ -1160,7 +1200,7 @@ def save_model(model, file_path, do_quantize = False, quantization_type = qint8,
 
 
 
-def all(pipe_or_dict_of_modules, pinnedMemory = False, quantizeTransformer = True,  extraModelsToQuantize = None, budgets= 0, asyncTransfers = True, compile = False, perc_reserved_mem_max = 0, verboseLevel = -1):
+def all(pipe_or_dict_of_modules, pinnedMemory = False, quantizeTransformer = True,  extraModelsToQuantize = None, quantizationType = qint8, budgets= 0, asyncTransfers = True, compile = False, perc_reserved_mem_max = 0, 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
@@ -1238,6 +1278,8 @@ def all(pipe_or_dict_of_modules, pinnedMemory = False, quantizeTransformer = Tru
     self.anyCompiledModule = compileAllModels or len(modelsToCompile)>0
     if self.anyCompiledModule:
         torch._dynamo.config.cache_size_limit = 10000
+        torch.compiler.reset()
+
       #  torch._logging.set_logs(recompiles=True)
       #  torch._inductor.config.realize_opcount_threshold = 100 # workaround bug "AssertionError: increase TRITON_MAX_BLOCK['X'] to 4096."
 
@@ -1252,19 +1294,31 @@ def all(pipe_or_dict_of_modules, pinnedMemory = False, quantizeTransformer = Tru
         
         # if the model has just been quantized so there is no need to quantize it again
         if model_id in models_to_quantize:
-            _quantize(current_model, weights=qint8, verboseLevel = self.verboseLevel, model_id=model_id)
+            _quantize(current_model, weights=quantizationType, verboseLevel = self.verboseLevel, model_id=model_id)
 
         modelPinned = (pinAllModels or model_id in modelsToPin) and not hasattr(current_model,"_already_pinned")
 
-        current_model_size = 0    
-        # load all the remaining unread lazy safetensors in RAM to free open cache files 
-        for p in current_model.parameters():
+        current_model_size = 0
+
+        for n, p in current_model.named_parameters():
+            p.requires_grad = False
+            p = p.detach()
             if isinstance(p, QTensor):
                 # # fix quanto bug (seems to have been fixed)   
                 # if not modelPinned and p._scale.dtype == torch.float32:
                 #     p._scale = p._scale.to(torch.bfloat16) 
-                current_model_size +=  torch.numel(p._scale) * p._scale.element_size()
-                current_model_size +=  torch.numel(p._data) * p._data.element_size()
+                if p._qtype == qint4:
+                    if hasattr(p,"_scale_shift"):
+                        current_model_size +=  torch.numel(p._scale_shift) * p._scale_shift.element_size()
+                    else:
+                        current_model_size +=  torch.numel(p._scale) * p._shift.element_size() + torch.numel(p._scale) * p._shift.element_size()
+
+                    current_model_size +=  torch.numel(p._data._data) * p._data._data.element_size()
+
+                else:
+                    current_model_size +=  torch.numel(p._scale) * p._scale.element_size()
+                    current_model_size +=  torch.numel(p._data) * p._data.element_size()
+
             else:
                 if p.data.dtype == torch.float32:
                     # convert any left overs float32 weight to bloat16 to divide by 2 the model memory footprint
@@ -1272,7 +1326,7 @@ def all(pipe_or_dict_of_modules, pinnedMemory = False, quantizeTransformer = Tru
                 current_model_size +=  torch.numel(p.data) * p.data.element_size()
                             
         for b in current_model.buffers():
-            if b.data.dtype == torch.float32:
+            if b.data.dtype == torch.float32: 
                 # convert any left overs float32 weight to bloat16 to divide by 2 the model memory footprint
                 b.data = b.data.to(torch.bfloat16)
             current_model_size +=  torch.numel(b.data) * b.data.element_size()
@@ -1308,7 +1362,6 @@ def all(pipe_or_dict_of_modules, pinnedMemory = False, quantizeTransformer = Tru
             print(f"Potential iterative blocks found in model '{model_id}':{towers_names}")
         # compile main iterative modules stacks ("towers")
         if compileAllModels or model_id in modelsToCompile :
-            #torch.compiler.reset()
             if self.verboseLevel>=1:
                 print(f"Pytorch compilation of model '{model_id}' is scheduled.")
             for tower in towers_modules:
@@ -1406,7 +1459,7 @@ def profile(pipe_or_dict_of_modules, profile_no: profile_type =  profile_type.Ve
         modules= modules.components 
 
     modules = {k: _remove_model_wrapper(v) for k, v in modules.items() if isinstance(v, torch.nn.Module)}
-    module_names = {k: v.__module__.lower() for k, v in modules.items() }
+    module_names = {k: _get_module_name(v) for k, v in modules.items() }
 
     default_extraModelsToQuantize = []
     quantizeTransformer = True

{mmgp-3.0.3 → mmgp-3.0.9}/src/mmgp/safetensors2.py

@@ -155,7 +155,7 @@ def torch_write_file(sd, file_path, quantization_map = None, config = None):
            torch.bool : 'BOOL' ,  torch.float64 : 'F64' , torch.float32 : 'F32' , torch.float16 : 'F16', torch.float8_e5m2 : "F8_E5M2", torch.float8_e4m3fn: "F8_E4M3" }
     pos = 0
     i = 0
-    mx = 1000000
+    mx = 100000
     for k , t  in sd.items():
         entry = {}
         dtypestr= map[t.dtype]
@@ -186,8 +186,6 @@ def torch_write_file(sd, file_path, quantization_map = None, config = None):
 
     length_of_header_bytes = struct.pack('<Q', size_header)
 
-    empty_tensor = b'\x80\x3f'
-
     with open(file_path, "wb") as writer:
         bytes_written = writer.write(length_of_header_bytes)        
         bytes_written = writer.write(header_bytes)        
@@ -195,12 +193,20 @@ def torch_write_file(sd, file_path, quantization_map = None, config = None):
         i = 0
         for k , t  in sd.items():
             size = torch.numel(t) * t.element_size()
-            if len(t.shape) == 0:
-                bytes_written = writer.write(empty_tensor)
-            else:
-                buffer = t.view(torch.uint8).numpy().tobytes()
-                bytes_written = writer.write(buffer)
-            assert bytes_written == size
+            if size != 0:
+                if len(t.shape) == 0:
+                    dtype = t.dtype
+                    # convert in a friendly format, scalars types not supported by numpy
+                    if  dtype ==  torch.bfloat16:
+                        t = t.view(torch.uint16)
+                    elif  dtype ==  torch.float8_e5m2 or dtype ==  torch.float8_e4m3fn:
+                        t = t.view(torch.uint8)
+                    buffer = t.numpy().tobytes()
+                else:
+                    buffer = t.view(torch.uint8).numpy().tobytes()
+                bytes_written = writer.write(buffer)            
+                assert bytes_written == size
+                
             i+=1
             if i==mx:
                 break
@@ -208,7 +214,7 @@ def torch_write_file(sd, file_path, quantization_map = None, config = None):
 class SafeTensorFile:
     """Main class for accessing safetensors files that provides memory-efficient access"""
     
-    def __init__(self, file_path, metadata, catalog, skip_bytes):
+    def __init__(self, file_path, metadata, catalog, skip_bytes, lazy_loading = True):
         self._file_path = file_path
         self._metadata = metadata
         self._catalog = catalog
@@ -216,20 +222,30 @@ class SafeTensorFile:
         self._keys = None
         self.sd = None
         self.mtracker = None
+        self.lazy_loading = lazy_loading
 
     @classmethod
-    def load_metadata(cls, file_path):    
+    def load_metadata(cls, file_path, lazy_loading = True):
         with open(file_path, 'rb') as f:
             catalog, metadata, skip_bytes = _read_safetensors_header(file_path, f)
 
-        return cls(file_path, metadata, catalog, skip_bytes)
+        return cls(file_path, metadata, catalog, skip_bytes, lazy_loading)
 
-    def init_tensors(self):
+    def init_tensors(self, lazyTensors = True):
         if self.sd is None:
-            self.sd = self.create_tensors()
+            self.lazy_loading = lazyTensors
+            if lazyTensors:
+                self.sd = self.create_tensors_with_mmap()
+            else:
+                self.sd = self.create_tensors_without_mmap()
+        # else:
+        #     if not self.lazy_loading and lazyTensors:
+        #         raise Exception("Every tensor should be either lazy loaded or not lazy loaded")
+
         return self.sd
     
-    def create_tensors(self):
+            
+    def create_tensors_with_mmap(self):
  
         self.mtracker = MmapTracker(self._file_path)
         import mmap
@@ -282,7 +298,12 @@ class SafeTensorFile:
                 map_idx = next(iter_tensor_no)
                 offset = current_pos - maps[map_idx][1]
                 if len(shape) == 0:
-                    t = torch.ones((), dtype=dtype, device="cpu")
+                    if length == 0: 
+                        t = torch.empty(0, dtype=dtype)
+                    else:
+                        # don't waste a memory view for a scalar
+                        t = torch.frombuffer(bytearray(maps[map_idx][0][offset:offset + length]), dtype=torch.uint8)
+                        t = t.view(dtype)                        
                 else:
                     mv = memoryview(maps[map_idx][0])[offset:offset + length]                
                     t = torch.frombuffer(mv, dtype=dtype)
@@ -293,8 +314,33 @@ class SafeTensorFile:
 
         return sd
 
+    def create_tensors_without_mmap(self):
+        sd = OrderedDict()    
+        
+        with open(self._file_path, 'rb') as f:
+            f.seek(self._skip_bytes, 0)
+            for k,v in self._catalog.items():
+                dtypestr =  v["dtype"]
+                dtype= _map_to_dtype[dtypestr]
+                shape = v["shape"]
+                data_offsets = v["data_offsets"]
+                length = data_offsets[1]-data_offsets[0]
+                buffer = f.read(length)
+                if len(shape) == 0:
+                    if length == 0: 
+                        t = torch.empty(0, dtype=dtype)
+                    else:
+                        t = torch.frombuffer(bytearray(buffer), dtype=torch.uint8)
+                        t = t.view(dtype)                        
+                else:
+                    t = torch.frombuffer(bytearray(buffer), dtype=dtype)
+                    t = torch.reshape(t, shape)
+                sd[k] = t
+        return sd
+
     def get_tensor(self, name: str) -> torch.tensor:
         """Get a tensor by name"""
+        # To do : switch to a JIT tensor creation per tensor
         self.init_tensors()
         return self.sd[name]
  
@@ -310,7 +356,7 @@ class SafeTensorFile:
         
     def tensors(self) -> Dict[str, torch.tensor]:
         """Get dictionary of all tensors"""
-        self.init_tensors()
+        self.init_tensors(self.lazy_loading)
         return self.sd
         
     def metadata(self) -> Optional[Dict[str, str]]:
@@ -319,7 +365,7 @@ class SafeTensorFile:
         
     def __len__(self) -> int:
         """Get number of tensors"""
-        self.init_tensors()
+        self.init_tensors(self.lazy_loading)
         return len(self.keys())
         
     def __contains__(self, key: str) -> bool:
@@ -337,10 +383,9 @@ class SafeTensorFile:
 class _SafeTensorLoader:
     """Context manager for loading SafeTensorFile"""
     
-    def __init__(self, filename: str):
+    def __init__(self, filename: str ):
         self.filename = Path(filename)
         self.sft = None
-        
         if not self.filename.exists():
             raise FileNotFoundError(f"File not found: {filename}")
             
@@ -367,7 +412,6 @@ class _SafeTensorLoader:
 
 def safe_open(filename: str, framework: str = "pt",device = "cpu") -> _SafeTensorLoader:
     if device != "cpu" or framework !="pt":
-        pass
         return _old_safe_open(filename =filename, framework=framework, device=device)
     return _SafeTensorLoader(filename)
 

{mmgp-3.0.3 → mmgp-3.0.9/src/mmgp.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.2
 Name: mmgp
-Version: 3.0.3
+Version: 3.0.9
 Summary: Memory Management for the GPU Poor
 Author-email: deepbeepmeep <deepbeepmeep@yahoo.com>
 License:                     GNU GENERAL PUBLIC LICENSE
@@ -13,10 +13,11 @@ Requires-Dist: optimum-quanto
 Requires-Dist: accelerate
 Requires-Dist: safetensors
 Requires-Dist: psutil
+Requires-Dist: peft
 
 
 <p align="center">
-  <H2>Memory Management 3.0 for the GPU Poor by DeepBeepMeep</H2>	
+  <H2>Memory Management 3.0.9 for the GPU Poor by DeepBeepMeep</H2>	
 </p>
 
 
@@ -38,8 +39,9 @@ Each profile may use a combination of the following:
 - Ability to pin models to reserved RAM to accelerate transfers to VRAM
 - Async transfers to VRAM to avoid a pause when loading a new slice of a model
 - Automated on the fly quantization or ability to load pre quantized models
-- support for pytorch compilation on Linux and WSL (not supported so far on pure Windows).
-
+- Pretrained Lora support with low RAM requirements
+- Support for pytorch compilation on Linux and WSL (supported on pure Windows but requires a complex Triton Installation).
+- 
 ## Installation
 First you need to install the module in your current project with:
 ```shell
@@ -105,20 +107,22 @@ If you are short on RAM and plan to work with quantized models, it is recommende
 ##  Going further
 
 The module includes several tools to package a light version of your favorite video / image generator:
-- *save_model(model, file_path, do_quantize = False, quantization_type = qint8 )*\
+- *save_model(model, file_path, do_quantize = False, quantizationType = qint8 )*\
 Save tensors of a model already loaded in memory in a safetensor format (much faster to reload). You can save it in a quantized format (default qint8 quantization recommended).
 The resulting safetensor file will contain extra fields in its metadata such as the quantization map and its configuration, so you will be able to move the file around without files such as *config.json* or *file_map.json*.
 You will need *load_model_data* or *fast_load_transformers_model* to read the file again . You may also load it using the default *safetensor* librar however you will need to provide in the same directory any complementary file that are usually requested (for instance *config.json*)
 
-- *load_model_data(model, file_path: str, do_quantize = False, quantization_type = qint8, pinToRAM = False, partialPin = False)*\
+- *load_model_data(model, file_path: str, do_quantize = False, quantizationType = qint8, pinToRAM = False, partialPin = False)*\
 Load the tensors data of a model in RAM of a model already initialized with no data. Detect and handle quantized models saved previously with *save_model*.A model can also be quantized on the fly while being loaded. The model which is loaded can be pinned to RAM while it is loaded, this is more RAM efficient than pinning tensors later using *offline.all* or *offline.profile*
 
-- *fast_load_transformers_model(model_path: str, do_quantize = False, quantization_type = qint8, pinToRAM = False, partialPin = False)*\
+- *fast_load_transformers_model(model_path: str, do_quantize = False, quantizationType = qint8, pinToRAM = False, partialPin = False)*\
 Initialize (build the model hierarchy in memory) and fast load the corresponding tensors of a 'transformers' or 'diffusers' library model.
 The advantages over the original *from_pretrained* method is that a full model can fit into a single file with a filename of your choosing (thefore you can have multiple 'transformers' versions of the same model in the same directory) and prequantized models are processed in a transparent way. 
 Last but not least, you can also on the fly pin to RAM the whole model or the most important part of it (partialPin = True) in a more efficient way (faster and requires less RAM) than if you did through *offload.all* or *offload.profile*.
 
-
+- *load_loras_into_model(model, lora_path, lora_multi)
+Load in a model a list of Lora described by a list of path *lora_path* and a list of *weights coefficients*.
+The Lora file must be in the *diffusers* format. This function works also on non diffusers models. However if there is already an official Lora support for a model it is recommended to use the official diffusers functions.
 
 The typical workflow wil be:
 1) temporarly insert the *save_model* function just after a model has been fully loaded to save a copy of the model / quantized model.

{mmgp-3.0.3 → mmgp-3.0.9}/src/mmgp.egg-info/requires.txt

@@ -3,3 +3,4 @@ optimum-quanto
 accelerate
 safetensors
 psutil
+peft