Add 4-bit LoRA support (#1200)

README.md

@@ -237,6 +237,7 @@ Optionally, you can use the following command-line flags:
 | `--groupsize GROUPSIZE`   | GPTQ: Group size. |
 | `--pre_layer PRE_LAYER`   | GPTQ: The number of layers to allocate to the GPU. Setting this parameter enables CPU offloading for 4-bit models. |
 | `--no-warmup_autotune`    | GPTQ: Disable warmup autotune for triton. |
+| `--monkey-patch`          | GPTQ: Apply the monkey patch for using LoRAs with quantized models. |
 
 #### FlexGen
 

modules/GPTQ_loader.py

@@ -16,6 +16,8 @@ from modelutils import find_layers
 from quant import make_quant
 
 
+# This function is a replacement for the load_quant function in the
+# GPTQ-for_LLaMa repository. It supports more models and branches.
 def _load_quant(model, checkpoint, wbits, groupsize=-1, faster_kernel=False, exclude_layers=['lm_head'], kernel_switch_threshold=128):
 
     def noop(*args, **kwargs):
@@ -64,6 +66,7 @@ def _load_quant(model, checkpoint, wbits, groupsize=-1, faster_kernel=False, exc
 
     try:
         from quant import autotune_warmup, make_quant_attn
+
         # triton branch
         make_quant_attn(model)
         if not shared.args.no_warmup_autotune:
@@ -77,6 +80,41 @@ def _load_quant(model, checkpoint, wbits, groupsize=-1, faster_kernel=False, exc
     return model
 
 
+# Used to locate the .pt/.safetensors quantized file
+def find_quantized_model_file(model_name):
+    path_to_model = Path(f'{shared.args.model_dir}/{model_name}')
+    pt_path = None
+    priority_name_list = [
+        Path(f'{shared.args.model_dir}/{model_name}{hyphen}{shared.args.wbits}bit{group}{ext}')
+        for group in ([f'-{shared.args.groupsize}g', ''] if shared.args.groupsize > 0 else [''])
+        for ext in ['.safetensors', '.pt']
+        for hyphen in ['-', f'/{model_name}-', '/']
+    ]
+    for path in priority_name_list:
+        if path.exists():
+            pt_path = path
+            break
+
+    # If the model hasn't been found with a well-behaved name, pick the last .pt
+    # or the last .safetensors found in its folder as a last resort
+    if not pt_path:
+        found_pts = list(path_to_model.glob("*.pt"))
+        found_safetensors = list(path_to_model.glob("*.safetensors"))
+        pt_path = None
+
+        if len(found_pts) > 0:
+            if len(found_pts) > 1:
+                print('Warning: more than one .pt model has been found. The last one will be selected. It could be wrong.')
+            pt_path = found_pts[-1]
+        elif len(found_safetensors) > 0:
+            if len(found_pts) > 1:
+                print('Warning: more than one .safetensors model has been found. The last one will be selected. It could be wrong.')
+            pt_path = found_safetensors[-1]
+
+    return pt_path
+
+
+# The function that loads the model in modules/models.py
 def load_quantized(model_name):
 
     # Find the model type
@@ -106,37 +144,9 @@ def load_quantized(model_name):
         print("Unknown pre-quantized model type specified. Only 'llama', 'opt' and 'gptj' are supported")
         exit()
 
-    # Locate the quantized model file
+    # Find the quantized model weights file (.pt/.safetensors)
     path_to_model = Path(f'{shared.args.model_dir}/{model_name}')
-    pt_path = None
+    pt_path = find_quantized_model_file(model_name)
-    priority_name_list = [
-        Path(f'{shared.args.model_dir}/{model_name}{hyphen}{shared.args.wbits}bit{group}{ext}')
-        for group in ([f'-{shared.args.groupsize}g', ''] if shared.args.groupsize > 0 else [''])
-        for ext in ['.safetensors', '.pt']
-        for hyphen in ['-', f'/{model_name}-', '/']
-    ]
-    for path in priority_name_list:
-        if path.exists():
-            pt_path = path
-            break
-
-    # If the model hasn't been found with a well-behaved name, pick the last .pt
-    # or the last .safetensors found in its folder as a last resort
-    if not pt_path:
-        path_to_model = Path(f'{shared.args.model_dir}/{model_name}')
-        found_pts = list(path_to_model.glob("*.pt"))
-        found_safetensors = list(path_to_model.glob("*.safetensors"))
-        pt_path = None
-
-        if len(found_pts) > 0:
-            if len(found_pts) > 1:
-                print('Warning: more than one .pt model has been found. The last one will be selected. It could be wrong.')
-            pt_path = found_pts[-1]
-        elif len(found_safetensors) > 0:
-            if len(found_pts) > 1:
-                print('Warning: more than one .safetensors model has been found. The last one will be selected. It could be wrong.')
-            pt_path = found_safetensors[-1]
-
     if not pt_path:
         print("Could not find the quantized model in .pt or .safetensors format, exiting...")
         exit()

modules/LoRA.py

@@ -43,6 +43,7 @@ def add_lora_to_model(lora_names):
             shared.model.load_adapter(Path(f"{shared.args.lora_dir}/{lora}"), lora)
 
         if not shared.args.load_in_8bit and not shared.args.cpu:
+            if not shared.args.monkey_patch:
                 shared.model.half()
             if not hasattr(shared.model, "hf_device_map"):
                 if torch.has_mps:

modules/models.py

@@ -101,6 +101,17 @@ def load_model(model_name):
 
     # Quantized model
     elif shared.args.wbits > 0:
+
+        # Monkey patch
+        if shared.args.monkey_patch:
+            print("Warning: applying the monkey patch for using LoRAs in 4-bit mode.\nIt may cause undefined behavior outside its intended scope.")
+            from modules.monkey_patch_gptq_lora import load_model_llama
+
+            model, tokenizer = load_model_llama(model_name)
+            return model, tokenizer
+
+        # No monkey patch
+        else:
             from modules.GPTQ_loader import load_quantized
 
             model = load_quantized(model_name)

modules/monkey_patch_gptq_lora.py

@@ -0,0 +1,41 @@
+# Copied from https://github.com/johnsmith0031/alpaca_lora_4bit
+
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path("repositories/alpaca_lora_4bit")))
+
+import autograd_4bit
+from autograd_4bit import (Autograd4bitQuantLinear,
+                           load_llama_model_4bit_low_ram)
+from monkeypatch.peft_tuners_lora_monkey_patch import (
+    Linear4bitLt, replace_peft_model_with_gptq_lora_model)
+
+from modules import shared
+from modules.GPTQ_loader import find_quantized_model_file
+
+replace_peft_model_with_gptq_lora_model()
+
+def load_model_llama(model_name):
+
+    config_path = str(Path(f'{shared.args.model_dir}/{model_name}'))
+    model_path = str(find_quantized_model_file(model_name))
+    model, tokenizer = load_llama_model_4bit_low_ram(config_path, model_path, groupsize=shared.args.groupsize, is_v1_model=False)
+
+    for n, m in model.named_modules():
+        if isinstance(m, Autograd4bitQuantLinear) or isinstance(m, Linear4bitLt):
+            if m.is_v1_model:
+                m.zeros = m.zeros.half()
+            m.scales = m.scales.half()
+            m.bias = m.bias.half()
+    autograd_4bit.use_new = True
+    autograd_4bit.auto_switch = True
+
+    try:
+        tokenizer.eos_token_id = 2
+        tokenizer.bos_token_id = 1
+        tokenizer.pad_token_id = 0
+    except:
+        pass
+
+    return model, tokenizer

modules/shared.py

@@ -124,6 +124,7 @@ parser.add_argument('--model_type', type=str, help='GPTQ: Model type of pre-quan
 parser.add_argument('--groupsize', type=int, default=-1, help='GPTQ: Group size.')
 parser.add_argument('--pre_layer', type=int, default=0, help='GPTQ: The number of layers to allocate to the GPU. Setting this parameter enables CPU offloading for 4-bit models.')
 parser.add_argument('--no-warmup_autotune', action='store_true', help='GPTQ: Disable warmup autotune for triton.')
+parser.add_argument('--monkey-patch', action='store_true', help='GPTQ: Apply the monkey patch for using LoRAs with quantized models.')
 
 # FlexGen
 parser.add_argument('--flexgen', action='store_true', help='Enable the use of FlexGen offloading.')

requirements.txt

@@ -1,7 +1,8 @@
 accelerate==0.18.0
+colorama
 datasets
 flexgen==0.1.7
-gradio==3.25
+gradio==3.25.0
 markdown
 numpy
 Pillow>=9.5.0