Add support for the latest GPTQ models with group-size

README.md

@@ -178,10 +178,10 @@ Optionally, you can use the following command-line flags:
 | `--cai-chat`     | Launch the web UI in chat mode with a style similar to Character.AI's. If the file `img_bot.png` or `img_bot.jpg` exists in the same folder as server.py, this image will be used as the bot's profile picture. Similarly, `img_me.png` or `img_me.jpg` will be used as your profile picture. |
 | `--cpu`          | Use the CPU to generate text.|
 | `--load-in-8bit` | Load the model with 8-bit precision.|
-| `--load-in-4bit` | DEPRECATED: use `--gptq-bits 4` instead. |
-| `--gptq-bits GPTQ_BITS` |  GPTQ: Load a pre-quantized model with specified precision. 2, 3, 4 and 8 (bit) are supported. Currently only works with LLaMA and OPT. |
-| `--gptq-model-type MODEL_TYPE` |  GPTQ: Model type of pre-quantized model. Currently only LLaMa and OPT are supported. |
-| `--gptq-pre-layer GPTQ_PRE_LAYER` |  GPTQ: The number of layers to preload. |
+| `--wbits WBITS`            | GPTQ: Load a pre-quantized model with specified precision in bits. 2, 3, 4 and 8 are supported. |
+| `--model_type MODEL_TYPE`  | GPTQ: Model type of pre-quantized model. Currently only LLaMA and OPT are supported. |
+| `--groupsize GROUPSIZE`    | GPTQ: Group size. |
+| `--pre_layer PRE_LAYER`    | GPTQ: The number of layers to preload. |
 | `--bf16`         | Load the model with bfloat16 precision. Requires NVIDIA Ampere GPU. |
 | `--auto-devices` | Automatically split the model across the available GPU(s) and CPU.|
 | `--disk`         | If the model is too large for your GPU(s) and CPU combined, send the remaining layers to the disk. |

download-model.py

@@ -116,10 +116,11 @@ def get_download_links_from_huggingface(model, branch):
 
             is_pytorch = re.match("(pytorch|adapter)_model.*\.bin", fname)
             is_safetensors = re.match("model.*\.safetensors", fname)
+            is_pt = re.match(".*\.pt", fname)
             is_tokenizer = re.match("tokenizer.*\.model", fname)
             is_text = re.match(".*\.(txt|json|py)", fname) or is_tokenizer
 
-            if any((is_pytorch, is_safetensors, is_text, is_tokenizer)):
+            if any((is_pytorch, is_safetensors, is_pt, is_tokenizer, is_text)):
                 if is_text:
                     links.append(f"https://huggingface.co/{model}/resolve/{branch}/{fname}")
                     classifications.append('text')
@@ -132,7 +133,8 @@ def get_download_links_from_huggingface(model, branch):
                     elif is_pytorch:
                         has_pytorch = True
                         classifications.append('pytorch')
-
+                    elif is_pt:
+                        classifications.append('pt')
 
         cursor = base64.b64encode(f'{{"file_name":"{dict[-1]["path"]}"}}'.encode()) + b':50'
         cursor = base64.b64encode(cursor)

modules/GPTQ_loader.py

@@ -14,18 +14,21 @@
 
 
 def load_quantized(model_name):
-    if not shared.args.gptq_model_type:
+    if not shared.args.model_type:
         # Try to determine model type from model name
-        model_type = model_name.split('-')[0].lower()
-        if model_type not in ('llama', 'opt'):
-            print("Can't determine model type from model name. Please specify it manually using --gptq-model-type "
+        if model_name.lower().startswith(('llama', 'alpaca')):
+            model_type = 'llama'
+        elif model_name.lower().startswith(('opt', 'galactica')):
+            model_type = 'opt'
+        else:
+            print("Can't determine model type from model name. Please specify it manually using --model_type "
                   "argument")
             exit()
     else:
-        model_type = shared.args.gptq_model_type.lower()
+        model_type = shared.args.model_type.lower()
 
     if model_type == 'llama':
-        if not shared.args.gptq_pre_layer:
+        if not shared.args.pre_layer:
             load_quant = llama.load_quant
         else:
             load_quant = llama_inference_offload.load_quant
@@ -35,33 +38,44 @@ def load_quantized(model_name):
         print("Unknown pre-quantized model type specified. Only 'llama' and 'opt' are supported")
         exit()
 
+    # Now we are going to try to locate the quantized model file.
     path_to_model = Path(f'models/{model_name}')
-    if path_to_model.name.lower().startswith('llama-7b'):
-        pt_model = f'llama-7b-{shared.args.gptq_bits}bit.pt'
-    elif path_to_model.name.lower().startswith('llama-13b'):
-        pt_model = f'llama-13b-{shared.args.gptq_bits}bit.pt'
-    elif path_to_model.name.lower().startswith('llama-30b'):
-        pt_model = f'llama-30b-{shared.args.gptq_bits}bit.pt'
-    elif path_to_model.name.lower().startswith('llama-65b'):
-        pt_model = f'llama-65b-{shared.args.gptq_bits}bit.pt'
+    found_pts = list(path_to_model.glob("*.pt"))
+    found_safetensors = list(path_to_model.glob("*.safetensors"))
+    pt_path = None
+
+    if len(found_pts) == 1:
+        pt_path = found_pts[0]
+    elif len(found_safetensors) == 1:
+        pt_path = found_safetensors[0]
     else:
-        pt_model = f'{model_name}-{shared.args.gptq_bits}bit.pt'
+        if path_to_model.name.lower().startswith('llama-7b'):
+            pt_model = f'llama-7b-{shared.args.wbits}bit'
+        elif path_to_model.name.lower().startswith('llama-13b'):
+            pt_model = f'llama-13b-{shared.args.wbits}bit'
+        elif path_to_model.name.lower().startswith('llama-30b'):
+            pt_model = f'llama-30b-{shared.args.wbits}bit'
+        elif path_to_model.name.lower().startswith('llama-65b'):
+            pt_model = f'llama-65b-{shared.args.wbits}bit'
+        else:
+            pt_model = f'{model_name}-{shared.args.wbits}bit'
 
-    # Try to find the .pt both in models/ and in the subfolder
-    pt_path = None
-    for path in [Path(p) for p in [f"models/{pt_model}", f"{path_to_model}/{pt_model}"]]:
-        if path.exists():
-            pt_path = path
+        # Try to find the .safetensors or .pt both in models/ and in the subfolder
+        for path in [Path(p+ext) for ext in ['.safetensors', '.pt'] for p in [f"models/{pt_model}", f"{path_to_model}/{pt_model}"]]:
+            if path.exists():
+                print(f"Found {path}")
+                pt_path = path
+                break
 
     if not pt_path:
-        print(f"Could not find {pt_model}, exiting...")
+        print("Could not find the quantized model in .pt or .safetensors format, exiting...")
         exit()
 
     # qwopqwop200's offload
-    if shared.args.gptq_pre_layer:
-        model = load_quant(str(path_to_model), str(pt_path), shared.args.gptq_bits, shared.args.gptq_pre_layer)
+    if shared.args.pre_layer:
+        model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits, shared.args.groupsize, shared.args.pre_layer)
     else:
-        model = load_quant(str(path_to_model), str(pt_path), shared.args.gptq_bits)
+        model = load_quant(str(path_to_model), str(pt_path), shared.args.wbits, shared.args.groupsize)
 
         # accelerate offload (doesn't work properly)
         if shared.args.gpu_memory:

modules/models.py

@@ -44,7 +44,7 @@ def load_model(model_name):
     shared.is_RWKV = model_name.lower().startswith('rwkv-')
 
     # Default settings
-    if not any([shared.args.cpu, shared.args.load_in_8bit, shared.args.gptq_bits, shared.args.auto_devices, shared.args.disk, shared.args.gpu_memory is not None, shared.args.cpu_memory is not None, shared.args.deepspeed, shared.args.flexgen, shared.is_RWKV]):
+    if not any([shared.args.cpu, shared.args.load_in_8bit, shared.args.wbits, shared.args.auto_devices, shared.args.disk, shared.args.gpu_memory is not None, shared.args.cpu_memory is not None, shared.args.deepspeed, shared.args.flexgen, shared.is_RWKV]):
         if any(size in shared.model_name.lower() for size in ('13b', '20b', '30b')):
             model = AutoModelForCausalLM.from_pretrained(Path(f"models/{shared.model_name}"), device_map='auto', load_in_8bit=True)
         else:
@@ -95,7 +95,7 @@ def load_model(model_name):
         return model, tokenizer
 
     # Quantized model
-    elif shared.args.gptq_bits > 0:
+    elif shared.args.wbits > 0:
         from modules.GPTQ_loader import load_quantized
 
         model = load_quantized(model_name)

modules/shared.py

@@ -52,7 +52,8 @@
         'default': 'Common sense questions and answers\n\nQuestion: \nFactual answer:',
         '^(gpt4chan|gpt-4chan|4chan)': '-----\n--- 865467536\nInput text\n--- 865467537\n',
         '(rosey|chip|joi)_.*_instruct.*': 'User: \n',
-        'oasst-*': '<|prompter|>Write a story about future of AI development<|endoftext|><|assistant|>'
+        'oasst-*': '<|prompter|>Write a story about future of AI development<|endoftext|><|assistant|>',
+        'alpaca-*': "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n### Instruction:\nWrite a poem about the transformers Python library. \nMention the word \"large language models\" in that poem.\n### Response:\n",
     },
     'lora_prompts': {
         'default': 'Common sense questions and answers\n\nQuestion: \nFactual answer:',
@@ -78,10 +79,15 @@ def str2bool(v):
 parser.add_argument('--cai-chat', action='store_true', help='Launch the web UI in chat mode with a style similar to Character.AI\'s. If the file img_bot.png or img_bot.jpg exists in the same folder as server.py, this image will be used as the bot\'s profile picture. Similarly, img_me.png or img_me.jpg will be used as your profile picture.')
 parser.add_argument('--cpu', action='store_true', help='Use the CPU to generate text.')
 parser.add_argument('--load-in-8bit', action='store_true', help='Load the model with 8-bit precision.')
-parser.add_argument('--load-in-4bit', action='store_true', help='DEPRECATED: use --gptq-bits 4 instead.')
-parser.add_argument('--gptq-bits', type=int, default=0, help='GPTQ: Load a pre-quantized model with specified precision. 2, 3, 4 and 8bit are supported. Currently only works with LLaMA and OPT.')
-parser.add_argument('--gptq-model-type', type=str, help='GPTQ: Model type of pre-quantized model. Currently only LLaMa and OPT are supported.')
-parser.add_argument('--gptq-pre-layer', type=int, default=0, help='GPTQ: The number of layers to preload.')
+
+parser.add_argument('--gptq-bits', type=int, default=0, help='DEPRECATED: use --wbits instead.')
+parser.add_argument('--gptq-model-type', type=str, help='DEPRECATED: use --model_type instead.')
+parser.add_argument('--gptq-pre-layer', type=int, default=0, help='DEPRECATED: use --pre_layer instead.')
+parser.add_argument('--wbits', type=int, default=0, help='GPTQ: Load a pre-quantized model with specified precision in bits. 2, 3, 4 and 8 are supported.')
+parser.add_argument('--model_type', type=str, help='GPTQ: Model type of pre-quantized model. Currently only LLaMA and OPT are supported.')
+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 preload.')
+
 parser.add_argument('--bf16', action='store_true', help='Load the model with bfloat16 precision. Requires NVIDIA Ampere GPU.')
 parser.add_argument('--auto-devices', action='store_true', help='Automatically split the model across the available GPU(s) and CPU.')
 parser.add_argument('--disk', action='store_true', help='If the model is too large for your GPU(s) and CPU combined, send the remaining layers to the disk.')
@@ -109,6 +115,8 @@ def str2bool(v):
 args = parser.parse_args()
 
 # Provisional, this will be deleted later
-if args.load_in_4bit:
-    print("Warning: --load-in-4bit is deprecated and will be removed. Use --gptq-bits 4 instead.\n")
-    args.gptq_bits = 4
+deprecated_dict = {'gptq_bits': ['wbits', 0], 'gptq_model_type': ['model_type', None], 'gptq_pre_layer': ['prelayer', 0]}
+for k in deprecated_dict:
+    if eval(f"args.{k}") != deprecated_dict[k][1]:
+        print(f"Warning: --{k} is deprecated and will be removed. Use --{deprecated_dict[k][0]} instead.")
+        exec(f"args.{deprecated_dict[k][0]} = args.{k}")

server.py

@@ -239,8 +239,9 @@ def set_interface_arguments(interface_mode, extensions, cmd_active):
 
 # Default UI settings
 default_preset = shared.settings['presets'][next((k for k in shared.settings['presets'] if re.match(k.lower(), shared.model_name.lower())), 'default')]
-default_text = shared.settings['lora_prompts'][next((k for k in shared.settings['lora_prompts'] if re.match(k.lower(), shared.lora_name.lower())), 'default')]
-if default_text == '':
+if shared.lora_name != "None":
+    default_text = shared.settings['lora_prompts'][next((k for k in shared.settings['lora_prompts'] if re.match(k.lower(), shared.lora_name.lower())), 'default')]
+else:
     default_text = shared.settings['prompts'][next((k for k in shared.settings['prompts'] if re.match(k.lower(), shared.model_name.lower())), 'default')]
 title ='Text generation web UI'
 description = '\n\n# Text generation lab\nGenerate text using Large Language Models.\n'