Merge pull request oobabooga#358 from mayaeary/8bit-offload

Add support for memory maps with --load-in-8bit

Author: oobabooga

modules/models.py

@@ -7,7 +7,9 @@
 import numpy as np
 import torch
 import transformers
-from transformers import AutoModelForCausalLM, AutoTokenizer
+from accelerate import infer_auto_device_map, init_empty_weights
+from transformers import (AutoConfig, AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
 
 import modules.shared as shared
 
@@ -94,39 +96,58 @@ def load_model(model_name):
 
     # Custom
     else:
-        command = "AutoModelForCausalLM.from_pretrained"
-        params = ["low_cpu_mem_usage=True"]
+        params = {"low_cpu_mem_usage": True}
         if not shared.args.cpu and not torch.cuda.is_available():
             print("Warning: torch.cuda.is_available() returned False.\nThis means that no GPU has been detected.\nFalling back to CPU mode.\n")
             shared.args.cpu = True
 
         if shared.args.cpu:
-            params.append("low_cpu_mem_usage=True")
-            params.append("torch_dtype=torch.float32")
+            params["torch_dtype"] = torch.float32
         else:
-            params.append("device_map='auto'")
-            params.append("load_in_8bit=True" if shared.args.load_in_8bit else "torch_dtype=torch.bfloat16" if shared.args.bf16 else "torch_dtype=torch.float16")
+            params["device_map"] = 'auto'
+            if shared.args.load_in_8bit:
+                params['quantization_config'] = BitsAndBytesConfig(load_in_8bit=True, llm_int8_enable_fp32_cpu_offload=True)
+            elif shared.args.bf16:
+                params["torch_dtype"] = torch.bfloat16
+            else:
+                params["torch_dtype"] = torch.float16
 
             if shared.args.gpu_memory:
                 memory_map = shared.args.gpu_memory
-                max_memory = f"max_memory={{0: '{memory_map[0]}GiB'"
-                for i in range(1, len(memory_map)):
-                    max_memory += (f", {i}: '{memory_map[i]}GiB'")
-                max_memory += (f", 'cpu': '{shared.args.cpu_memory or '99'}GiB'}}")
-                params.append(max_memory)
-            elif not shared.args.load_in_8bit:
-                total_mem = (torch.cuda.get_device_properties(0).total_memory/(1024*1024))
-                suggestion = round((total_mem-1000)/1000)*1000
-                if total_mem-suggestion < 800:
+                max_memory = {}
+                for i in range(len(memory_map)):
+                    max_memory[i] = f'{memory_map[i]}GiB'
+                max_memory['cpu'] = f'{shared.args.cpu_memory or 99}GiB'
+                params['max_memory'] = max_memory
+            else:
+                total_mem = (torch.cuda.get_device_properties(0).total_memory / (1024*1024))
+                suggestion = round((total_mem-1000) / 1000) * 1000
+                if total_mem - suggestion < 800:
                     suggestion -= 1000
                 suggestion = int(round(suggestion/1000))
                 print(f"\033[1;32;1mAuto-assiging --gpu-memory {suggestion} for your GPU to try to prevent out-of-memory errors.\nYou can manually set other values.\033[0;37;0m")
-                params.append(f"max_memory={{0: '{suggestion}GiB', 'cpu': '{shared.args.cpu_memory or '99'}GiB'}}")
-            if shared.args.disk:
-                params.append(f"offload_folder='{shared.args.disk_cache_dir}'")
+                
+                max_memory = {0: f'{suggestion}GiB', 'cpu': f'{shared.args.cpu_memory or 99}GiB'}
+                params['max_memory'] = max_memory
 
-        command = f"{command}(Path(f'models/{shared.model_name}'), {', '.join(set(params))})"
-        model = eval(command)
+            if shared.args.disk:
+                params["offload_folder"] = shared.args.disk_cache_dir
+
+        checkpoint = Path(f'models/{shared.model_name}')
+
+        if shared.args.load_in_8bit and params.get('max_memory', None) is not None and params['device_map'] == 'auto':
+            config = AutoConfig.from_pretrained(checkpoint)
+            with init_empty_weights():
+                model = AutoModelForCausalLM.from_config(config)
+            model.tie_weights()
+            params['device_map'] = infer_auto_device_map(
+                model, 
+                dtype=torch.int8, 
+                max_memory=params['max_memory'],
+                no_split_module_classes = model._no_split_modules
+            )
+
+        model = AutoModelForCausalLM.from_pretrained(checkpoint, **params)
 
     # Loading the tokenizer
     if shared.model_name.lower().startswith(('gpt4chan', 'gpt-4chan', '4chan')) and Path("models/gpt-j-6B/").exists():