Remove GGML support

README.md

@@ -158,7 +158,7 @@ text-generation-webui
 │   │   └── tokenizer.model
 ```
 
-* GGML/GGUF models are a single file and should be placed directly into `models`. Example:
+* GGUF models are a single file and should be placed directly into `models`. Example:
 
 ```
 text-generation-webui
@@ -260,7 +260,7 @@ 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/GGUF (for llama.cpp and ctransformers)
+#### GGUF (for llama.cpp and ctransformers)
 
 | Flag        | Description |
 |-------------|-------------|
@@ -279,8 +279,6 @@ Optionally, you can use the following command-line flags:
 | `--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 |
 | `--llama_cpp_seed SEED`        | Seed for llama-cpp models. Default 0 (random). |
-| `--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`  | 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. |
 

docs/llama.cpp.md

@@ -9,16 +9,14 @@ llama.cpp is the best backend in two important scenarios:
 
 #### Pre-converted
 
-Download the GGUF or GGML models directly into your `text-generation-webui/models` folder. It will be a single file.
+Download the GGUF models directly into your `text-generation-webui/models` folder. It will be a single file.
 
-* For GGUF models, make sure its name contains `.gguf`.
+* Make sure its name ends in `.gguf`.
-* For GGML models, make sure its name contains `ggml` and ends in `.bin`.
+* `q4_K_M` quantization is recommended.
-
-`q4_K_M` quantization is recommended.
 
 #### Convert Llama yourself
 
-Follow the instructions in the llama.cpp README to generate a ggml: https://github.com/ggerganov/llama.cpp#prepare-data--run
+Follow the instructions in the llama.cpp README to generate a GGUF: https://github.com/ggerganov/llama.cpp#prepare-data--run
 
 ## GPU acceleration
 

download-model.py

@@ -58,7 +58,6 @@ class ModelDownloader:
         has_pytorch = False
         has_pt = False
         has_gguf = False
-        has_ggml = False
         has_safetensors = False
         is_lora = False
         while True:
@@ -83,10 +82,9 @@ class ModelDownloader:
                 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_gguf, is_ggml, is_tokenizer, is_text)):
+                if any((is_pytorch, is_safetensors, is_pt, is_gguf, is_tokenizer, is_text)):
                     if 'lfs' in dict[i]:
                         sha256.append([fname, dict[i]['lfs']['oid']])
 
@@ -109,9 +107,6 @@ class ModelDownloader:
                         elif is_gguf:
                             has_gguf = True
                             classifications.append('gguf')
-                        elif is_ggml:
-                            has_ggml = True
-                            classifications.append('ggml')
 
             cursor = base64.b64encode(f'{{"file_name":"{dict[-1]["path"]}"}}'.encode()) + b':50'
             cursor = base64.b64encode(cursor)
@@ -123,13 +118,8 @@ class ModelDownloader:
                 if classifications[i] in ['pytorch', 'pt']:
                     links.pop(i)
 
-        # If both GGML and GGUF are available, download GGUF only
+        is_llamacpp = has_gguf and specific_file is not None
-        if has_ggml and has_gguf:
+        return links, sha256, is_lora, is_llamacpp
-            for i in range(len(classifications) - 1, -1, -1):
-                if classifications[i] == 'ggml':
-                    links.pop(i)
-
-        return links, sha256, is_lora, ((has_ggml or has_gguf) and specific_file is not None)
 
     def get_output_folder(self, model, branch, is_lora, is_llamacpp=False, base_folder=None):
         if base_folder is None:

models/config.yaml

@@ -63,7 +63,6 @@ llama-65b-gptq-3bit:
 .*vicuna.*(1.5|1_5):
   instruction_template: 'Vicuna-v1.1'
   truncation_length: 4096
-  rms_norm_eps: 5.0e-6  
 .*stable.*vicuna:
   instruction_template: 'StableVicuna'
 (?!.*chat).*chinese-vicuna:
@@ -211,24 +210,19 @@ llama-65b-gptq-3bit:
   instruction_template: 'Alpaca'
 .*llama-(2|v2):
   truncation_length: 4096
-  rms_norm_eps: 5.0e-6
 .*llama-(2|v2).*chat:
   instruction_template: 'Llama-v2'
-.*70b.*ggml.*\.bin:
-  n_gqa: 8
 .*newhope:
   instruction_template: 'NewHope'
 .*stablebeluga2:
   instruction_template: 'StableBeluga2'
   truncation_length: 4096
-  rms_norm_eps: 5.0e-6
 .*openchat:
   instruction_template: 'OpenChat'
 .*falcon.*-instruct:
 .*(openorca-platypus2):
   instruction_template: 'OpenOrca-Platypus2'
   custom_stopping_strings: '"### Instruction:", "### Response:"'
-  rms_norm_eps: 5.0e-6
 .*codellama:
   rope_freq_base: 1000000
 .*codellama.*instruct:

modules/llamacpp_hf.py

@@ -9,39 +9,23 @@ 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(model_file: Union[str, Path] = None):
+def llama_cpp_lib():
-    if model_file is not None:
-        gguf_model = is_gguf(model_file)
-    else:
-        gguf_model = True
-
     if shared.args.cpu or llama_cpp_cuda is None:
-        return llama_cpp if gguf_model else llama_cpp_ggml
+        return llama_cpp
     else:
-        return llama_cpp_cuda if gguf_model else llama_cpp_ggml_cuda
+        return llama_cpp_cuda
 
 
 class LlamacppHF(PreTrainedModel):
@@ -64,7 +48,7 @@ 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(path).llama_new_context_with_model(model.model, model.params)
+                'ctx': llama_cpp_lib().llama_new_context_with_model(model.model, model.params)
             }
 
     def _validate_model_class(self):
@@ -181,7 +165,7 @@ class LlamacppHF(PreTrainedModel):
         if path.is_file():
             model_file = path
         else:
-            model_file = (list(path.glob('*.gguf*')) + list(path.glob('*ggml*.bin')))[0]
+            model_file = list(path.glob('*.gguf'))[0]
 
         logger.info(f"llama.cpp weights detected: {model_file}\n")
 
@@ -207,14 +191,7 @@ class LlamacppHF(PreTrainedModel):
             'logits_all': True,
         }
 
-        if not is_gguf(model_file):
+        Llama = llama_cpp_lib().Llama
-            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, model_file)

modules/llamacpp_model.py

@@ -1,7 +1,5 @@
 import re
 from functools import partial
-from pathlib import Path
-from typing import Union
 
 import torch
 
@@ -9,39 +7,23 @@ 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(model_file: Union[str, Path] = None):
+def llama_cpp_lib():
-    if model_file is not None:
-        gguf_model = is_gguf(model_file)
-    else:
-        gguf_model = True
-
     if shared.args.cpu or llama_cpp_cuda is None:
-        return llama_cpp if gguf_model else llama_cpp_ggml
+        return llama_cpp
     else:
-        return llama_cpp_cuda if gguf_model else llama_cpp_ggml_cuda
+        return llama_cpp_cuda
 
 
 def ban_eos_logits_processor(eos_token, input_ids, logits):
@@ -59,8 +41,8 @@ class LlamaCppModel:
     @classmethod
     def from_pretrained(self, path):
 
-        Llama = llama_cpp_lib(path).Llama
+        Llama = llama_cpp_lib().Llama
-        LlamaCache = llama_cpp_lib(path).LlamaCache
+        LlamaCache = llama_cpp_lib().LlamaCache
 
         result = self()
         cache_capacity = 0
@@ -95,13 +77,6 @@ class LlamaCppModel:
             'rope_freq_scale': 1.0 / shared.args.compress_pos_emb,
         }
 
-        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/loaders.py

@@ -68,8 +68,6 @@ loaders_and_params = OrderedDict({
     ],
     'llama.cpp': [
         'n_ctx',
-        'n_gqa',
-        'rms_norm_eps',
         'n_gpu_layers',
         'tensor_split',
         'n_batch',
@@ -86,8 +84,6 @@ loaders_and_params = OrderedDict({
     ],
     'llamacpp_HF': [
         'n_ctx',
-        'n_gqa',
-        'rms_norm_eps',
         'n_gpu_layers',
         'tensor_split',
         'n_batch',

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('*.gguf*')) + 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'))[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('*.gguf*')) + list(path_to_model.glob('*ggml*.bin'))) > 0:
+    elif len(list(path_to_model.glob('*.gguf'))) > 0:
         loader = 'llama.cpp'
-    elif re.match(r'.*\.gguf|.*ggml.*\.bin', model_name.lower()):
+    elif re.match(r'.*\.gguf', model_name.lower()):
         loader = 'llama.cpp'
     elif re.match(r'.*rwkv.*\.pth', model_name.lower()):
         loader = 'RWKV'

modules/shared.py

@@ -126,8 +126,6 @@ parser.add_argument('--n-gpu-layers', type=int, default=0, help='Number of layer
 parser.add_argument('--tensor_split', type=str, default=None, help="Split the model across multiple GPUs, comma-separated list of proportions, e.g. 18,17")
 parser.add_argument('--n_ctx', type=int, default=2048, help='Size of the prompt context.')
 parser.add_argument('--llama_cpp_seed', type=int, default=0, help='Seed for llama-cpp models. Default 0 (random)')
-parser.add_argument('--n_gqa', type=int, default=0, help='grouped-query attention. Must be 8 for llama-2 70b.')
-parser.add_argument('--rms_norm_eps', type=float, default=0, help='5e-6 is a good value for llama-2 models.')
 
 # GPTQ
 parser.add_argument('--wbits', type=int, default=0, help='Load a pre-quantized model with specified precision in bits. 2, 3, 4 and 8 are supported.')

modules/ui.py

@@ -73,8 +73,6 @@ def list_model_elements():
         'n_gpu_layers',
         'tensor_split',
         'n_ctx',
-        'n_gqa',
-        'rms_norm_eps',
         'llama_cpp_seed',
         'gpu_split',
         'max_seq_len',

modules/ui_model_menu.py

@@ -82,8 +82,6 @@ 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='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='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")
@@ -128,7 +126,7 @@ def create_ui():
                     shared.gradio['autoload_model'] = gr.Checkbox(value=shared.settings['autoload_model'], label='Autoload the model', info='Whether to load the model as soon as it is selected in the Model dropdown.')
 
                 shared.gradio['custom_model_menu'] = gr.Textbox(label="Download model or LoRA", info="Enter the Hugging Face username/model path, for instance: facebook/galactica-125m. To specify a branch, add it at the end after a \":\" character like this: facebook/galactica-125m:main. To download a single file, enter its name in the second box.")
-                shared.gradio['download_specific_file'] = gr.Textbox(placeholder="File name (for GGUF/GGML)", show_label=False, max_lines=1)
+                shared.gradio['download_specific_file'] = gr.Textbox(placeholder="File name (for GGUF models)", show_label=False, max_lines=1)
                 with gr.Row():
                     shared.gradio['download_model_button'] = gr.Button("Download", variant='primary')
                     shared.gradio['get_file_list'] = gr.Button("Get file list")

modules/utils.py

@@ -2,7 +2,6 @@ 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
@@ -125,15 +124,3 @@ 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

@@ -44,12 +44,6 @@ https://github.com/abetlen/llama-cpp-python/releases/download/v0.1.84/llama_cpp_
 https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels/releases/download/textgen-webui/llama_cpp_python_cuda-0.1.84+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-0.1.84+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"
 https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.0/gptq_for_llama-0.1.0+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"

server.py

@@ -177,8 +177,6 @@ if __name__ == "__main__":
         'skip_special_tokens': shared.settings['skip_special_tokens'],
         'custom_stopping_strings': shared.settings['custom_stopping_strings'],
         'truncation_length': shared.settings['truncation_length'],
-        'n_gqa': 0,
-        'rms_norm_eps': 0,
         'rope_freq_base': 0,
     }