Merge pull request #3695 from oobabooga/gguf2

GGUF

README.md

@@ -156,7 +156,7 @@ text-generation-webui
 
 In the "Model" tab of the UI, those models can be automatically downloaded from Hugging Face. You can also download them via the command-line with `python download-model.py organization/model`.
 
-* GGML models are a single file and should be placed directly into `models`. Example:
+* GGML/GGUF models are a single file and should be placed directly into `models`. Example:
 
 ```
 text-generation-webui
@@ -258,7 +258,7 @@ Optionally, you can use the following command-line flags:
 | `--quant_type QUANT_TYPE`                   | quant_type for 4-bit. Valid options: nf4, fp4. |
 | `--use_double_quant`                        | use_double_quant for 4-bit. |
 
-#### GGML (for llama.cpp and ctransformers)
+#### GGML/GGUF (for llama.cpp and ctransformers)
 
 | Flag        | Description |
 |-------------|-------------|
@@ -269,16 +269,16 @@ Optionally, you can use the following command-line flags:
 
 #### llama.cpp
 
-| Flag        | Description |
+| Flag          | Description |
-|-------------|-------------|
+|---------------|---------------|
-| `--no-mmap` | Prevent mmap from being used. |
+| `--no-mmap`   | Prevent mmap from being used. |
-| `--mlock`   | Force the system to keep the model in RAM. |
+| `--mlock`     | Force the system to keep the model in RAM. |
 | `--mul_mat_q` | Activate new mulmat kernels. |
 | `--cache-capacity CACHE_CAPACITY`   | Maximum cache capacity. Examples: 2000MiB, 2GiB. When provided without units, bytes will be assumed. |
-| `--tensor_split TENSOR_SPLIT` | Split the model across multiple GPUs, comma-separated list of proportions, e.g. 18,17 |
+| `--tensor_split TENSOR_SPLIT`  | Split the model across multiple GPUs, comma-separated list of proportions, e.g. 18,17 |
-| `--llama_cpp_seed SEED` | Seed for llama-cpp models. Default 0 (random). |
+| `--llama_cpp_seed SEED`        | Seed for llama-cpp models. Default 0 (random). |
-| `--n_gqa N_GQA`         | grouped-query attention. Must be 8 for llama-2 70b. |
+| `--n_gqa N_GQA`                | GGML only (not used by GGUF): Grouped-Query Attention. Must be 8 for llama-2 70b. |
-| `--rms_norm_eps RMS_NORM_EPS`  | 5e-6 is a good value for llama-2 models. |
+| `--rms_norm_eps RMS_NORM_EPS`  | GGML only (not used by GGUF): 5e-6 is a good value for llama-2 models. |
 | `--cpu`                        | Use the CPU version of llama-cpp-python instead of the GPU-accelerated version. |
 |`--cfg-cache`                   | llamacpp_HF: Create an additional cache for CFG negative prompts. |
 

download-model.py

@@ -57,7 +57,8 @@ class ModelDownloader:
         classifications = []
         has_pytorch = False
         has_pt = False
-        # has_ggml = False
+        has_gguf = False
+        has_ggml = False
         has_safetensors = False
         is_lora = False
         while True:
@@ -78,10 +79,11 @@ class ModelDownloader:
                 is_pytorch = re.match(r"(pytorch|adapter|gptq)_model.*\.bin", fname)
                 is_safetensors = re.match(r".*\.safetensors", fname)
                 is_pt = re.match(r".*\.pt", fname)
+                is_gguf = re.match(r'.*\.gguf', fname)
                 is_ggml = re.match(r".*ggml.*\.bin", fname)
                 is_tokenizer = re.match(r"(tokenizer|ice|spiece).*\.model", fname)
                 is_text = re.match(r".*\.(txt|json|py|md)", fname) or is_tokenizer
-                if any((is_pytorch, is_safetensors, is_pt, is_ggml, is_tokenizer, is_text)):
+                if any((is_pytorch, is_safetensors, is_pt, is_gguf, is_ggml, is_tokenizer, is_text)):
                     if 'lfs' in dict[i]:
                         sha256.append([fname, dict[i]['lfs']['oid']])
 
@@ -101,8 +103,11 @@ class ModelDownloader:
                         elif is_pt:
                             has_pt = True
                             classifications.append('pt')
+                        elif is_gguf:
+                            has_gguf = True
+                            classifications.append('gguf')
                         elif is_ggml:
-                            # has_ggml = True
+                            has_ggml = True
                             classifications.append('ggml')
 
             cursor = base64.b64encode(f'{{"file_name":"{dict[-1]["path"]}"}}'.encode()) + b':50'
@@ -115,6 +120,12 @@ class ModelDownloader:
                 if classifications[i] in ['pytorch', 'pt']:
                     links.pop(i)
 
+        # If both GGML and GGUF are available, download GGUF only
+        if has_ggml and has_gguf:
+            for i in range(len(classifications) - 1, -1, -1):
+                if classifications[i] == 'ggml':
+                    links.pop(i)
+
         return links, sha256, is_lora
 
     def get_output_folder(self, model, branch, is_lora, base_folder=None):

modules/llamacpp_hf.py

@@ -9,27 +9,43 @@ from transformers.modeling_outputs import CausalLMOutputWithPast
 
 from modules import RoPE, shared
 from modules.logging_colors import logger
+from modules.utils import is_gguf
 
 import llama_cpp
 
+try:
+    import llama_cpp_ggml
+except:
+    llama_cpp_ggml = llama_cpp
+
 if torch.cuda.is_available() and not torch.version.hip:
     try:
         import llama_cpp_cuda
     except:
         llama_cpp_cuda = None
+    try:
+        import llama_cpp_ggml_cuda
+    except:
+        llama_cpp_ggml_cuda = llama_cpp_cuda
 else:
     llama_cpp_cuda = None
+    llama_cpp_ggml_cuda = None
 
 
-def llama_cpp_lib():
+def llama_cpp_lib(model_file: Union[str, Path] = None):
-    if shared.args.cpu or llama_cpp_cuda is None:
+    if model_file is not None:
-        return llama_cpp
+        gguf_model = is_gguf(model_file)
     else:
-        return llama_cpp_cuda
+        gguf_model = True
+
+    if shared.args.cpu or llama_cpp_cuda is None:
+        return llama_cpp if gguf_model else llama_cpp_ggml
+    else:
+        return llama_cpp_cuda if gguf_model else llama_cpp_ggml_cuda
 
 
 class LlamacppHF(PreTrainedModel):
-    def __init__(self, model):
+    def __init__(self, model, path):
         super().__init__(PretrainedConfig())
         self.model = model
         self.generation_config = GenerationConfig()
@@ -48,7 +64,7 @@ class LlamacppHF(PreTrainedModel):
                 'n_tokens': self.model.n_tokens,
                 'input_ids': self.model.input_ids.copy(),
                 'scores': self.model.scores.copy(),
-                'ctx': llama_cpp_lib().llama_new_context_with_model(model.model, model.params)
+                'ctx': llama_cpp_lib(path).llama_new_context_with_model(model.model, model.params)
             }
 
     def _validate_model_class(self):
@@ -165,7 +181,7 @@ class LlamacppHF(PreTrainedModel):
         if path.is_file():
             model_file = path
         else:
-            model_file = list(path.glob('*ggml*.bin'))[0]
+            model_file = (list(path.glob('*.gguf*')) + list(path.glob('*ggml*.bin')))[0]
 
         logger.info(f"llama.cpp weights detected: {model_file}\n")
 
@@ -188,12 +204,17 @@ class LlamacppHF(PreTrainedModel):
             'rope_freq_base': RoPE.get_rope_freq_base(shared.args.alpha_value, shared.args.rope_freq_base),
             'tensor_split': tensor_split_list,
             'rope_freq_scale': 1.0 / shared.args.compress_pos_emb,
-            'n_gqa': shared.args.n_gqa or None,
-            'rms_norm_eps': shared.args.rms_norm_eps or None,
             'logits_all': True,
         }
 
-        Llama = llama_cpp_lib().Llama
+        if not is_gguf(model_file):
+            ggml_params = {
+                'n_gqa': shared.args.n_gqa or None,
+                'rms_norm_eps': shared.args.rms_norm_eps or None,
+            }
+            params = params | ggml_params
+
+        Llama = llama_cpp_lib(model_file).Llama
         model = Llama(**params)
 
-        return LlamacppHF(model)
+        return LlamacppHF(model, model_file)

modules/llamacpp_model.py

@@ -1,5 +1,7 @@
 import re
 from functools import partial
+from pathlib import Path
+from typing import Union
 
 import torch
 
@@ -7,23 +9,39 @@ from modules import RoPE, shared
 from modules.callbacks import Iteratorize
 from modules.logging_colors import logger
 from modules.text_generation import get_max_prompt_length
+from modules.utils import is_gguf
 
 import llama_cpp
 
+try:
+    import llama_cpp_ggml
+except:
+    llama_cpp_ggml = llama_cpp
+
 if torch.cuda.is_available() and not torch.version.hip:
     try:
         import llama_cpp_cuda
     except:
         llama_cpp_cuda = None
+    try:
+        import llama_cpp_ggml_cuda
+    except:
+        llama_cpp_ggml_cuda = llama_cpp_cuda
 else:
     llama_cpp_cuda = None
+    llama_cpp_ggml_cuda = None
 
 
-def llama_cpp_lib():
+def llama_cpp_lib(model_file: Union[str, Path] = None):
-    if shared.args.cpu or llama_cpp_cuda is None:
+    if model_file is not None:
-        return llama_cpp
+        gguf_model = is_gguf(model_file)
     else:
-        return llama_cpp_cuda
+        gguf_model = True
+
+    if shared.args.cpu or llama_cpp_cuda is None:
+        return llama_cpp if gguf_model else llama_cpp_ggml
+    else:
+        return llama_cpp_cuda if gguf_model else llama_cpp_ggml_cuda
 
 
 def ban_eos_logits_processor(eos_token, input_ids, logits):
@@ -41,8 +59,8 @@ class LlamaCppModel:
     @classmethod
     def from_pretrained(self, path):
 
-        Llama = llama_cpp_lib().Llama
+        Llama = llama_cpp_lib(path).Llama
-        LlamaCache = llama_cpp_lib().LlamaCache
+        LlamaCache = llama_cpp_lib(path).LlamaCache
 
         result = self()
         cache_capacity = 0
@@ -75,10 +93,15 @@ class LlamaCppModel:
             'rope_freq_base': RoPE.get_rope_freq_base(shared.args.alpha_value, shared.args.rope_freq_base),
             'tensor_split': tensor_split_list,
             'rope_freq_scale': 1.0 / shared.args.compress_pos_emb,
-            'n_gqa': shared.args.n_gqa or None,
-            'rms_norm_eps': shared.args.rms_norm_eps or None,
         }
 
+        if not is_gguf(path):
+            ggml_params = {
+                'n_gqa': shared.args.n_gqa or None,
+                'rms_norm_eps': shared.args.rms_norm_eps or None,
+            }
+            params = params | ggml_params
+
         result.model = Llama(**params)
         if cache_capacity > 0:
             result.model.set_cache(LlamaCache(capacity_bytes=cache_capacity))

modules/models.py

@@ -241,7 +241,7 @@ def llamacpp_loader(model_name):
     if path.is_file():
         model_file = path
     else:
-        model_file = list(Path(f'{shared.args.model_dir}/{model_name}').glob('*ggml*.bin'))[0]
+        model_file = (list(Path(f'{shared.args.model_dir}/{model_name}').glob('*.gguf*')) + list(Path(f'{shared.args.model_dir}/{model_name}').glob('*ggml*.bin')))[0]
 
     logger.info(f"llama.cpp weights detected: {model_file}")
     model, tokenizer = LlamaCppModel.from_pretrained(model_file)

modules/models_settings.py

@@ -24,9 +24,9 @@ def infer_loader(model_name):
         loader = None
     elif Path(f'{shared.args.model_dir}/{model_name}/quantize_config.json').exists() or ('wbits' in model_settings and type(model_settings['wbits']) is int and model_settings['wbits'] > 0):
         loader = 'AutoGPTQ'
-    elif len(list(path_to_model.glob('*ggml*.bin'))) > 0:
+    elif len(list(path_to_model.glob('*.gguf*')) + list(path_to_model.glob('*ggml*.bin'))) > 0:
         loader = 'llama.cpp'
-    elif re.match(r'.*ggml.*\.bin', model_name.lower()):
+    elif re.match(r'.*\.gguf|.*ggml.*\.bin', model_name.lower()):
         loader = 'llama.cpp'
     elif re.match(r'.*rwkv.*\.pth', model_name.lower()):
         loader = 'RWKV'

modules/ui_model_menu.py

@@ -80,8 +80,8 @@ def create_ui():
                             shared.gradio['n_ctx'] = gr.Slider(minimum=0, maximum=16384, step=256, label="n_ctx", value=shared.args.n_ctx)
                             shared.gradio['threads'] = gr.Slider(label="threads", minimum=0, step=1, maximum=32, value=shared.args.threads)
                             shared.gradio['n_batch'] = gr.Slider(label="n_batch", minimum=1, maximum=2048, value=shared.args.n_batch)
-                            shared.gradio['n_gqa'] = gr.Slider(minimum=0, maximum=16, step=1, label="n_gqa", value=shared.args.n_gqa, info='grouped-query attention. Must be 8 for llama-2 70b.')
+                            shared.gradio['n_gqa'] = gr.Slider(minimum=0, maximum=16, step=1, label="n_gqa", value=shared.args.n_gqa, info='GGML only (not used by GGUF): Grouped-Query Attention. Must be 8 for llama-2 70b.')
-                            shared.gradio['rms_norm_eps'] = gr.Slider(minimum=0, maximum=1e-5, step=1e-6, label="rms_norm_eps", value=shared.args.rms_norm_eps, info='5e-6 is a good value for llama-2 models.')
+                            shared.gradio['rms_norm_eps'] = gr.Slider(minimum=0, maximum=1e-5, step=1e-6, label="rms_norm_eps", value=shared.args.rms_norm_eps, info='GGML only (not used by GGUF): 5e-6 is a good value for llama-2 models.')
 
                             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")

modules/utils.py

@@ -2,6 +2,7 @@ import os
 import re
 from datetime import datetime
 from pathlib import Path
+from typing import Union
 
 from modules import shared
 from modules.logging_colors import logger
@@ -124,3 +125,15 @@ def get_datasets(path: str, ext: str):
 
 def get_available_chat_styles():
     return sorted(set(('-'.join(k.stem.split('-')[1:]) for k in Path('css').glob('chat_style*.css'))), key=natural_keys)
+
+
+def is_gguf(path: Union[str, Path]) -> bool:
+    '''
+    Determines if a llama.cpp model is in GGUF format
+    Copied from ctransformers utils.py
+    '''
+    path = str(Path(path).resolve())
+    with open(path, "rb") as f:
+        magic = f.read(4)
+
+    return magic == "GGUF".encode()

requirements.txt

@@ -22,19 +22,31 @@ tensorboard
 tqdm
 wandb
 
+# bitsandbytes
 bitsandbytes==0.41.1; platform_system != "Windows"
 https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.41.1-py3-none-win_amd64.whl; platform_system == "Windows"
+
+# AutoGPTQ
 https://github.com/PanQiWei/AutoGPTQ/releases/download/v0.4.2/auto_gptq-0.4.2+cu117-cp310-cp310-win_amd64.whl; platform_system == "Windows"
 https://github.com/PanQiWei/AutoGPTQ/releases/download/v0.4.2/auto_gptq-0.4.2+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+
+# ExLlama
 https://github.com/jllllll/exllama/releases/download/0.0.10/exllama-0.0.10+cu117-cp310-cp310-win_amd64.whl; platform_system == "Windows"
 https://github.com/jllllll/exllama/releases/download/0.0.10/exllama-0.0.10+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
 
 # llama-cpp-python without GPU support
-llama-cpp-python==0.1.78; platform_system != "Windows"
+llama-cpp-python==0.1.79; platform_system != "Windows"
-https://github.com/abetlen/llama-cpp-python/releases/download/v0.1.78/llama_cpp_python-0.1.78-cp310-cp310-win_amd64.whl; platform_system == "Windows"
+https://github.com/abetlen/llama-cpp-python/releases/download/v0.1.79/llama_cpp_python-0.1.79-cp310-cp310-win_amd64.whl; platform_system == "Windows"
+
 # llama-cpp-python with CUDA support
-https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_cuda-0.1.78+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-0.1.79+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-0.1.78+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_cuda-0.1.79+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+
+# llama-cpp-python with GGML support
+https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/cpu/llama_cpp_python_ggml-0.1.78+cpuavx2-cp310-cp310-win_amd64.whl; platform_system == "Windows"
+https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/cpu/llama_cpp_python_ggml-0.1.78+cpuavx2-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_ggml_cuda-0.1.78+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_ggml_cuda-0.1.78+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
 
 # GPTQ-for-LLaMa
 https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.0/gptq_for_llama-0.1.0+cu117-cp310-cp310-win_amd64.whl; platform_system == "Windows"