ctransformers: another attempt

Generalized ctransformers based on:
#2892
Credits to randoentity

modules/ctransformers_model.py

@@ -0,0 +1,76 @@
+from ctransformers import AutoModelForCausalLM
+from ctransformers import AutoConfig
+
+from modules import shared
+from modules.callbacks import Iteratorize
+from modules.logging_colors import logger
+
+class CtransformersModel:
+    def __init__(self):
+        pass
+
+    @classmethod
+    def from_pretrained(self, path):
+        result = self()
+        stops = shared.settings['custom_stopping_strings']
+        stops.append("<|end|>")
+
+        # ctransformers uses -1 for random seed
+        config = AutoConfig.from_pretrained(
+            str(path),
+            stop=stops,
+            threads=shared.args.threads,
+            gpu_layers=shared.args.n_gpu_layers,
+            batch_size=shared.args.n_batch,
+            stream=not shared.args.no_stream,
+            seed=(-1 if shared.args.llama_cpp_seed == 0 else shared.args.llama_cpp_seed)
+        )
+        self.model = AutoModelForCausalLM.from_pretrained(
+            str(result.model_dir(path) if result.model_type_is_auto() else path),
+            model_type=(None if result.model_type_is_auto() else shared.args.model_type),
+            config=config
+        )
+        logger.info(f'Using ctransformers model_type: {self.model.model_type} for {self.model.model_path}')
+        return result, result
+
+    def model_type_is_auto(self):
+        return shared.args.model_type == "Auto" or shared.args.model_type == "None"
+
+    def model_dir(self, path):
+        if path.is_file():
+            return path.parent
+        return path
+
+    def encode(self, string, **kwargs):
+        return self.model.tokenize(string)
+
+    def decode(self, ids):
+        return self.model.detokenize(ids)
+
+
+    def generate(self, prompt, state, callback=None):
+        prompt = prompt if type(prompt) is str else prompt.decode()
+        generator = self.model._stream(
+            prompt=prompt,
+            max_new_tokens=state['max_new_tokens'],
+            temperature=state['temperature'],
+            top_p=state['top_p'],
+            top_k=state['top_k'],
+            repetition_penalty=state['repetition_penalty'],
+            threads=shared.args.threads
+        )
+
+        output = ""
+        for token in generator:
+            if callback:
+                callback(token)
+            output += token
+        return output
+
+
+    def generate_with_streaming(self, *args, **kwargs):
+        with Iteratorize(self.generate, args, kwargs, callback=None) as generator:
+            reply = ''
+            for token in generator:
+                reply += token
+                yield reply

modules/loaders.py

@@ -86,6 +86,16 @@
         'compress_pos_emb',
         'alpha_value',
         'exllama_HF_info',
+    ],
+    'ctransformers': [
+        'n_ctx',
+        'n_gpu_layers',
+        'n_batch',
+        'threads',
+        'no_mmap',
+        'mlock',
+        'model_type',
+        'llama_cpp_seed',
     ]
 }
 
@@ -244,6 +254,13 @@
         'skip_special_tokens',
         'auto_max_new_tokens',
     },
+    'ctransformers': {
+        'temperature',
+        'top_p',
+        'top_k',
+        'repetition_penalty',
+        'seed'
+    }
 }
 
 
@@ -264,6 +281,30 @@ def blacklist_samplers(loader):
     else:
         return [gr.update(visible=True) if sampler in loaders_samplers[loader] else gr.update(visible=False) for sampler in all_samplers]
 
+model_loader_type_table = {
+    'GPTQ-for-LLaMa': [
+        "None",
+        "llama",
+        "opt",
+        "gptj"
+    ],
+    'ctransformers': [
+        "None",
+        "gptj",
+        "gpt_neox",
+        "llama",
+        "mpt",
+        "dolly-v2"
+        "replit",
+        "starcoder",
+        "falcon"
+    ],
+}
+
+def model_loader_type(loader):
+    if loader in model_loader_type_table:
+        return model_loader_type_table[loader]
+    return ["None"]
 
 def get_gpu_memory_keys():
     return [k for k in shared.gradio if k.startswith('gpu_memory')]

modules/models.py

@@ -58,7 +58,8 @@ def load_model(model_name, loader=None):
         'llamacpp_HF': llamacpp_HF_loader,
         'RWKV': RWKV_loader,
         'ExLlama': ExLlama_loader,
-        'ExLlama_HF': ExLlama_HF_loader
+        'ExLlama_HF': ExLlama_HF_loader,
+        'ctransformers': CtransformorsModel_loader,
     }
 
     p = Path(model_name)
@@ -263,6 +264,25 @@ def llamacpp_HF_loader(model_name):
     return model, tokenizer
 
 
+def CtransformorsModel_loader(model_name):
+    from modules.ctransformers_model import CtransformersModel
+
+    path = Path(f'{shared.args.model_dir}/{model_name}')
+    logger.info(f'ctransformers loading: {path}\n')
+    ctrans = CtransformersModel()
+    if ctrans.model_type_is_auto():
+        model_file = path
+    else:
+        if path.is_file():
+            model_file = path
+        else:
+            model_file = list(
+                Path(f'{shared.args.model_dir}/{model_name}').glob('*.bin')
+            )[0]
+            logger.info(f'ctransformers weights detected: {model_file}\n')
+    model, tokenizer = ctrans.from_pretrained(model_file)
+    return model, tokenizer
+
 def GPTQ_loader(model_name):
 
     # Monkey patch

modules/shared.py

@@ -35,7 +35,7 @@
     'autoload_model': False,
     'max_new_tokens': 200,
     'max_new_tokens_min': 1,
-    'max_new_tokens_max': 4096,
+    'max_new_tokens_max': 8000,
     'auto_max_new_tokens': False,
     'seed': -1,
     'character': 'None',

modules/text_generation.py

@@ -34,9 +34,10 @@ def generate_reply(*args, **kwargs):
 def get_max_prompt_length(state):
     return state['truncation_length'] - state['max_new_tokens']
 
-
+encode_llama_prompts = ['LlamaCppModel', 'RWKVModel', 'CtransformersModel']
+encode_llama_truncation = ['LlamaCppModel', 'RWKVModel', 'ExllamaModel', 'CtransformersModel']
 def encode(prompt, add_special_tokens=True, add_bos_token=True, truncation_length=None):
-    if shared.model.__class__.__name__ in ['LlamaCppModel', 'RWKVModel']:
+    if shared.model.__class__.__name__ in encode_llama_prompts:
         input_ids = shared.tokenizer.encode(str(prompt))
         input_ids = np.array(input_ids).reshape(1, len(input_ids))
         return input_ids
@@ -51,7 +52,7 @@ def encode(prompt, add_special_tokens=True, add_bos_token=True, truncation_lengt
     if truncation_length is not None:
         input_ids = input_ids[:, -truncation_length:]
 
-    if shared.model.__class__.__name__ in ['LlamaCppModel', 'RWKVModel', 'ExllamaModel'] or shared.args.cpu:
+    if shared.model.__class__.__name__ in encode_llama_truncation or shared.args.cpu:
         return input_ids
     elif shared.args.deepspeed:
         return input_ids.to(device=local_rank)
@@ -169,7 +170,12 @@ def apply_stopping_strings(reply, all_stop_strings):
 
     return reply, stop_found
 
-
+_generate_reply_use_custom = [
+    'LlamaCppModel',
+    'RWKVModel',
+    'ExllamaModel',
+    'CtransformersModel'
+]
 def _generate_reply(question, state, stopping_strings=None, is_chat=False):
     generate_func = apply_extensions('custom_generate_reply')
     if generate_func is None:
@@ -178,7 +184,7 @@ def _generate_reply(question, state, stopping_strings=None, is_chat=False):
             yield ''
             return
 
-        if shared.model.__class__.__name__ in ['LlamaCppModel', 'RWKVModel', 'ExllamaModel']:
+        if shared.model.__class__.__name__ in _generate_reply_use_custom:
             generate_func = generate_reply_custom
         else:
             generate_func = generate_reply_HF

requirements.txt

@@ -31,3 +31,4 @@ https://github.com/abetlen/llama-cpp-python/releases/download/v0.1.77/llama_cpp_
 # llama-cpp-python with CUDA support
 https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_cuda-0.1.77+cu117-cp310-cp310-win_amd64.whl; platform_system == "Windows"
 https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_cuda-0.1.77+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/jllllll/ctransformers-cuBLAS-wheels/releases/download/AVX2/ctransformers-0.2.16+cu117-py3-none-any.whl

server.py

@@ -204,7 +204,7 @@ def create_model_menus():
 
     with gr.Row():
         with gr.Column():
-            shared.gradio['loader'] = gr.Dropdown(label="Model loader", choices=["Transformers", "ExLlama_HF", "ExLlama", "AutoGPTQ", "GPTQ-for-LLaMa", "llama.cpp", "llamacpp_HF"], value=None)
+            shared.gradio['loader'] = gr.Dropdown(label="Model loader", choices=loaders.loaders_and_params.keys(), value=None)
             with gr.Box():
                 with gr.Row():
                     with gr.Column():
@@ -225,7 +225,7 @@ def create_model_menus():
 
                         shared.gradio['wbits'] = gr.Dropdown(label="wbits", choices=["None", 1, 2, 3, 4, 8], value=str(shared.args.wbits) if shared.args.wbits > 0 else "None")
                         shared.gradio['groupsize'] = gr.Dropdown(label="groupsize", choices=["None", 32, 64, 128, 1024], value=str(shared.args.groupsize) if shared.args.groupsize > 0 else "None")
-                        shared.gradio['model_type'] = gr.Dropdown(label="model_type", choices=["None", "llama", "opt", "gptj"], value=shared.args.model_type or "None")
+                        shared.gradio['model_type'] = gr.Dropdown(label="model_type", choices=["None"], value=shared.args.model_type or "None")
                         shared.gradio['pre_layer'] = gr.Slider(label="pre_layer", minimum=0, maximum=100, value=shared.args.pre_layer[0] if shared.args.pre_layer is not None else 0)
                         shared.gradio['autogptq_info'] = gr.Markdown('* ExLlama_HF is recommended over AutoGPTQ for models derived from LLaMA.')
                         shared.gradio['gpu_split'] = gr.Textbox(label='gpu-split', info='Comma-separated list of VRAM (in GB) to use per GPU. Example: 20,7,7')
@@ -267,6 +267,7 @@ def create_model_menus():
                 shared.gradio['model_status'] = gr.Markdown('No model is loaded' if shared.model_name == 'None' else 'Ready')
 
     shared.gradio['loader'].change(loaders.make_loader_params_visible, gradio('loader'), gradio(loaders.get_all_params()))
+    shared.gradio['loader'].change(fn=lambda value: gr.update(choices=loaders.model_loader_type(value)), inputs=shared.gradio['loader'], outputs=shared.gradio['model_type'])
 
     # In this event handler, the interface state is read and updated
     # with the model defaults (if any), and then the model is loaded