Merge branch 'oobabooga:main' into exllama-module

README.md

@@ -212,7 +212,7 @@ Optionally, you can use the following command-line flags:
 
 | Flag                                       | Description |
 |--------------------------------------------|-------------|
-| `--loader LOADER`                          | Choose the model loader manually, otherwise, it will get autodetected. Valid options: transformers, autogptq, gptq-for-llama, exllama, llamacpp, rwkv, flexgen |
+| `--loader LOADER`                          | Choose the model loader manually, otherwise, it will get autodetected. Valid options: transformers, autogptq, gptq-for-llama, exllama, exllama_hf, llamacpp, rwkv, flexgen |
 
 #### Accelerate/transformers
 

characters/instruction-following/Starchat-Beta.yaml

@@ -0,0 +1,4 @@
+user: "<|user|>"
+bot: "<|assistant|>"
+context: "<|system|>\n<|end|>\n"
+turn_template: "<|user|>\n<|user-message|><|end|>\n<|bot|>\n<|bot-message|><|end|>\n"

characters/instruction-following/Tulu.yaml

@@ -0,0 +1,4 @@
+user: "<|user|>"
+bot: "<|assistant|>"
+context: ""
+turn_template: "<|user|>\n<|user-message|>\n<|bot|>\n<|bot-message|>\n"

css/chat.css

@@ -93,3 +93,22 @@ div.svelte-362y77>*, div.svelte-362y77>.form>* {
 .message-body :not(pre) > code {
     white-space: normal !important;
 }
+
+@media print {
+    body {
+        visibility: hidden;
+    }
+
+    .chat {
+        visibility: visible;
+        position: absolute;
+        left: 0;
+        top: 0;
+        max-width: none;
+        max-height: none;
+        width: 100%;
+        height: fit-content;
+        display: flex;
+        flex-direction: column-reverse;
+    }
+}

css/html_instruct_style.css

@@ -17,6 +17,10 @@
     margin-bottom: 1.25em !important;
 }
 
+.message-body ul, .message-body ol {
+    margin-bottom: 1.25em !important;
+}
+
 .dark .message-body p em {
     color: rgb(198, 202, 214) !important;
 }

docker/Dockerfile

@@ -26,7 +26,7 @@ LABEL maintainer="Your Name <your.email@example.com>"
 LABEL description="Docker image for GPTQ-for-LLaMa and Text Generation WebUI"
 
 RUN apt-get update && \
-    apt-get install --no-install-recommends -y libportaudio2 libasound-dev git python3 python3-pip make g++ && \
+    apt-get install --no-install-recommends -y python3-dev libportaudio2 libasound-dev git python3 python3-pip make g++ && \
     rm -rf /var/lib/apt/lists/*
 
 RUN --mount=type=cache,target=/root/.cache/pip pip3 install virtualenv

download-model.py

@@ -1,5 +1,5 @@
 '''
-Downloads models from Hugging Face to models/model-name.
+Downloads models from Hugging Face to models/username_modelname.
 
 Example:
 python download-model.py facebook/opt-1.3b
@@ -11,8 +11,8 @@ import base64
 import datetime
 import hashlib
 import json
-import re
 import os
+import re
 import sys
 from pathlib import Path
 
@@ -21,63 +21,12 @@ import tqdm
 from tqdm.contrib.concurrent import thread_map
 
 
-def select_model_from_default_options():
-    models = {
-        "OPT 6.7B": ("facebook", "opt-6.7b", "main"),
-        "OPT 2.7B": ("facebook", "opt-2.7b", "main"),
-        "OPT 1.3B": ("facebook", "opt-1.3b", "main"),
-        "OPT 350M": ("facebook", "opt-350m", "main"),
-        "GALACTICA 6.7B": ("facebook", "galactica-6.7b", "main"),
-        "GALACTICA 1.3B": ("facebook", "galactica-1.3b", "main"),
-        "GALACTICA 125M": ("facebook", "galactica-125m", "main"),
-        "Pythia-6.9B-deduped": ("EleutherAI", "pythia-6.9b-deduped", "main"),
-        "Pythia-2.8B-deduped": ("EleutherAI", "pythia-2.8b-deduped", "main"),
-        "Pythia-1.4B-deduped": ("EleutherAI", "pythia-1.4b-deduped", "main"),
-        "Pythia-410M-deduped": ("EleutherAI", "pythia-410m-deduped", "main"),
-    }
-
-    choices = {}
-    print("Select the model that you want to download:\n")
-    for i, name in enumerate(models):
-        char = chr(ord('A') + i)
-        choices[char] = name
-        print(f"{char}) {name}")
-
-    char_hugging = chr(ord('A') + len(models))
-    print(f"{char_hugging}) Manually specify a Hugging Face model")
-    char_exit = chr(ord('A') + len(models) + 1)
-    print(f"{char_exit}) Do not download a model")
-    print()
-    print("Input> ", end='')
-    choice = input()[0].strip().upper()
-    if choice == char_exit:
-        exit()
-    elif choice == char_hugging:
-        print("""\nType the name of your desired Hugging Face model in the format organization/name.
-
-Examples:
-facebook/opt-1.3b
-EleutherAI/pythia-1.4b-deduped
-""")
-
-        print("Input> ", end='')
-        model = input()
-        branch = "main"
-    else:
-        arr = models[choices[choice]]
-        model = f"{arr[0]}/{arr[1]}"
-        branch = arr[2]
-
-    return model, branch
-
-
 class ModelDownloader:
     def __init__(self):
         self.s = requests.Session()
         if os.getenv('HF_USER') is not None and os.getenv('HF_PASS') is not None:
             self.s.auth = (os.getenv('HF_USER'), os.getenv('HF_PASS'))
 
-
     def sanitize_model_and_branch_names(self, model, branch):
         if model[-1] == '/':
             model = model[:-1]
@@ -92,7 +41,6 @@ class ModelDownloader:
 
         return model, branch
 
-
     def get_download_links_from_huggingface(self, model, branch, text_only=False):
         base = "https://huggingface.co"
         page = f"/api/models/{model}/tree/{branch}"
@@ -163,7 +111,6 @@ class ModelDownloader:
 
         return links, sha256, is_lora
 
-
     def get_output_folder(self, model, branch, is_lora, base_folder=None):
         if base_folder is None:
             base_folder = 'models' if not is_lora else 'loras'
@@ -171,59 +118,64 @@ class ModelDownloader:
         output_folder = f"{'_'.join(model.split('/')[-2:])}"
         if branch != 'main':
             output_folder += f'_{branch}'
+
         output_folder = Path(base_folder) / output_folder
         return output_folder
 
-
     def get_single_file(self, url, output_folder, start_from_scratch=False):
         filename = Path(url.rsplit('/', 1)[1])
         output_path = output_folder / filename
+        headers = {}
+        mode = 'wb'
         if output_path.exists() and not start_from_scratch:
+
             # Check if the file has already been downloaded completely
             r = self.s.get(url, stream=True, timeout=20)
             total_size = int(r.headers.get('content-length', 0))
             if output_path.stat().st_size >= total_size:
                 return
+
             # Otherwise, resume the download from where it left off
             headers = {'Range': f'bytes={output_path.stat().st_size}-'}
             mode = 'ab'
-        else:
-            headers = {}
-            mode = 'wb'
 
-        r = self.s.get(url, stream=True, headers=headers, timeout=20)
+        with self.s.get(url, stream=True, headers=headers, timeout=20) as r:
-        with open(output_path, mode) as f:
+            r.raise_for_status()  # Do not continue the download if the request was unsuccessful
             total_size = int(r.headers.get('content-length', 0))
-            block_size = 1024
+            block_size = 1024 * 1024  # 1MB
-            with tqdm.tqdm(total=total_size, unit='iB', unit_scale=True, bar_format='{l_bar}{bar}| {n_fmt:6}/{total_fmt:6} {rate_fmt:6}') as t:
+            with open(output_path, mode) as f:
-                for data in r.iter_content(block_size):
+                with tqdm.tqdm(total=total_size, unit='iB', unit_scale=True, bar_format='{l_bar}{bar}| {n_fmt:6}/{total_fmt:6} {rate_fmt:6}') as t:
-                    t.update(len(data))
+                    count = 0
-                    f.write(data)
+                    for data in r.iter_content(block_size):
-
+                        t.update(len(data))
+                        f.write(data)
+                        if total_size != 0 and self.progress_bar is not None:
+                            count += len(data)
+                            self.progress_bar(float(count) / float(total_size), f"Downloading {filename}")
 
     def start_download_threads(self, file_list, output_folder, start_from_scratch=False, threads=1):
         thread_map(lambda url: self.get_single_file(url, output_folder, start_from_scratch=start_from_scratch), file_list, max_workers=threads, disable=True)
 
+    def download_model_files(self, model, branch, links, sha256, output_folder, progress_bar=None, start_from_scratch=False, threads=1):
+        self.progress_bar = progress_bar
 
-    def download_model_files(self, model, branch, links, sha256, output_folder, start_from_scratch=False, threads=1):
         # Creating the folder and writing the metadata
-        if not output_folder.exists():
+        output_folder.mkdir(parents=True, exist_ok=True)
-            output_folder.mkdir(parents=True, exist_ok=True)
+        metadata = f'url: https://huggingface.co/{model}\n' \
-        with open(output_folder / 'huggingface-metadata.txt', 'w') as f:
+                   f'branch: {branch}\n' \
-            f.write(f'url: https://huggingface.co/{model}\n')
+                   f'download date: {datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")}\n'
-            f.write(f'branch: {branch}\n')
+
-            f.write(f'download date: {str(datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S"))}\n')
+        sha256_str = '\n'.join([f'    {item[1]} {item[0]}' for item in sha256])
-            sha256_str = ''
+        if sha256_str:
-            for i in range(len(sha256)):
+            metadata += f'sha256sum:\n{sha256_str}'
-                sha256_str += f'    {sha256[i][1]} {sha256[i][0]}\n'
+
-            if sha256_str != '':
+        metadata += '\n'
-                f.write(f'sha256sum:\n{sha256_str}')
+        (output_folder / 'huggingface-metadata.txt').write_text(metadata)
 
         # Downloading the files
         print(f"Downloading the model to {output_folder}")
         self.start_download_threads(links, output_folder, start_from_scratch=start_from_scratch, threads=threads)
 
-
     def check_model_files(self, model, branch, links, sha256, output_folder):
         # Validate the checksums
         validated = True
@@ -264,8 +216,6 @@ if __name__ == '__main__':
 
     branch = args.branch
     model = args.MODEL
-    if model is None:
-        model, branch = select_model_from_default_options()
 
     downloader = ModelDownloader()
     # Cleaning up the model/branch names

models/config.yaml

@@ -50,7 +50,7 @@ llama-65b-gptq-3bit:
 .*vicuna.*v0:
   mode: 'instruct'
   instruction_template: 'Vicuna-v0'
-.*vicuna.*(1.1|1_1):
+.*vicuna.*(1.1|1_1|1.3|1_3):
   mode: 'instruct'
   instruction_template: 'Vicuna-v1.1'
 .*wizard.*vicuna:
@@ -184,7 +184,7 @@ llama-65b-gptq-3bit:
 .*Nous-Hermes-13b:
   mode: 'instruct'
   instruction_template: 'Alpaca'
-.*airoboros-13b-gpt4:
+.*airoboros:
   mode: 'instruct'
   instruction_template: 'Vicuna-v1.1'
 .*WizardLM-30B-V1.0:
@@ -193,7 +193,7 @@ llama-65b-gptq-3bit:
 TheBloke_WizardLM-30B-GPTQ:
   mode: 'instruct'
   instruction_template: 'Vicuna-v1.1'
-.*(A|a)lpa(cino|sta):
+.*alpa(cino|sta):
   mode: 'instruct'
   instruction_template: 'Alpaca'
 .*hippogriff:
@@ -202,9 +202,33 @@ TheBloke_WizardLM-30B-GPTQ:
 .*gpt4all-.*-snoozy:
   mode: 'instruct'
   instruction_template: 'WizardLM'
-.*(L|l)azarus:
+.*lazarus:
   mode: 'instruct'
   instruction_template: 'Alpaca'
-.*(G|g)uanaco-.*(7|13|33|65)(b|B):
+.*guanaco-.*(7|13|33|65)b:
   mode: 'instruct'
   instruction_template: 'Guanaco'
+.*hypermantis:
+  mode: 'instruct'
+  instruction_template: 'Alpaca'
+.*open-llama-.*-open-instruct:
+  mode: 'instruct'
+  instruction_template: 'Alpaca'
+.*starcoder-gpteacher-code-instruct:
+  mode: 'instruct'
+  instruction_template: 'Alpaca'
+.*tulu:
+  mode: 'instruct'
+  instruction_template: 'Tulu'
+.*chronos:
+  mode: 'instruct'
+  instruction_template: 'Alpaca'
+.*samantha:
+  mode: 'instruct'
+  instruction_template: 'Samantha'
+.*wizardcoder:
+  mode: 'instruct'
+  instruction_template: 'Alpaca'
+.*starchat-beta:
+  mode: 'instruct'
+  instruction_template: 'Starchat-Beta'

modules/exllama_hf.py

@@ -0,0 +1,82 @@
+import os
+import sys
+from pathlib import Path
+from typing import *
+
+import torch
+from transformers import (
+    GenerationConfig,
+    LlamaTokenizer,
+    PretrainedConfig,
+    PreTrainedModel
+)
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+from modules import shared
+from modules.logging_colors import logger
+from modules.relative_imports import RelativeImport
+
+with RelativeImport("repositories/exllama"):
+    from model import ExLlama, ExLlamaCache, ExLlamaConfig
+
+
+class ExllamaHF(PreTrainedModel):
+    def __init__(self, config: ExLlamaConfig):
+        super().__init__(PretrainedConfig())
+        self.ex_config = config
+        self.ex_model = ExLlama(self.ex_config)
+        self.generation_config = GenerationConfig()
+
+    def _validate_model_class(self):
+        pass
+
+    def _validate_model_kwargs(self, model_kwargs: Dict[str, Any]):
+        pass
+
+    def prepare_inputs_for_generation(self, input_ids, **kwargs):
+        return {'input_ids': input_ids, **kwargs}
+
+    @property
+    def device(self) -> torch.device:
+        # TODO: May cause problem on multi-gpu inference?
+        return torch.device(0)
+
+    def __call__(self, *args, **kwargs):
+        # TODO: Some decoding methods (such as Contrastive Search) may not work at this time
+        assert len(args) == 0, 'no *args should be passed to forward'
+        use_cache = kwargs['use_cache']
+        seq = kwargs['input_ids'][0].tolist()
+        cache = kwargs['past_key_values'] if 'past_key_values' in kwargs else None
+        if cache is None:
+            cache = ExLlamaCache(self.ex_model)
+            self.ex_model.forward(torch.tensor([seq[:-1]], dtype=torch.long), cache, preprocess_only=True)
+        logits = self.ex_model.forward(torch.tensor([seq[-1:]], dtype=torch.long), cache).to(self.device)
+        return CausalLMOutputWithPast(logits=logits, past_key_values=cache if use_cache else None)
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
+        assert len(model_args) == 0 and len(kwargs) == 0, "extra args is currently not supported"
+        if isinstance(pretrained_model_name_or_path, str):
+            pretrained_model_name_or_path = Path(pretrained_model_name_or_path)
+            
+        pretrained_model_name_or_path = Path(f'{shared.args.model_dir}') / Path(pretrained_model_name_or_path)
+        config = ExLlamaConfig(pretrained_model_name_or_path / 'config.json')
+
+        # from 'oobabooga/text-generation-webui/modules/exllama.py'
+        weight_path = None
+        for ext in ['.safetensors', '.pt', '.bin']:
+            found = list(pretrained_model_name_or_path.glob(f"*{ext}"))
+            if len(found) > 0:
+                weight_path = found[-1]
+                break
+        assert weight_path is not None, f'could not find weight in "{pretrained_model_name_or_path}"'
+
+        config.model_path = str(weight_path)
+        
+        # This slowes down a bit but align better with autogptq generation.
+        # TODO: Should give user choice to tune the exllama config
+        config.act_order = True
+        config.fused_attn = False
+        config.fused_mlp_thd = 0
+
+        return ExllamaHF(config)

modules/llamacpp_model.py

@@ -52,9 +52,9 @@ class LlamaCppModel:
             'n_gpu_layers': shared.args.n_gpu_layers
         }
 
-        self.model = Llama(**params)
+        result.model = Llama(**params)
         if cache_capacity > 0:
-            self.model.set_cache(LlamaCache(capacity_bytes=cache_capacity))
+            result.model.set_cache(LlamaCache(capacity_bytes=cache_capacity))
 
         # This is ugly, but the model and the tokenizer are the same object in this library.
         return result, result

modules/loaders.py

@@ -55,6 +55,10 @@ loaders_and_params = {
     'ExLlama' : [
         'gpu_split',
         'exllama_info',
+    ],
+    'ExLlama_HF' : [
+        'gpu_split',
+        'exllama_HF_info',
     ]
 }
 

modules/models.py

@@ -49,7 +49,8 @@ def load_model(model_name, loader=None):
         'llama.cpp': llamacpp_loader,
         'FlexGen': flexgen_loader,
         'RWKV': RWKV_loader,
-        'ExLlama': ExLlama_loader
+        'ExLlama': ExLlama_loader,
+        'ExLlama_HF': ExLlama_HF_loader
     }
 
     if loader is None:
@@ -278,6 +279,12 @@ def ExLlama_loader(model_name):
     return model, tokenizer
 
 
+def ExLlama_HF_loader(model_name):
+    from modules.exllama_hf import ExllamaHF
+
+    return ExllamaHF.from_pretrained(model_name)
+
+
 def get_max_memory_dict():
     max_memory = {}
     if shared.args.gpu_memory:

modules/shared.py

@@ -98,7 +98,7 @@ parser.add_argument('--extensions', type=str, nargs="+", help='The list of exten
 parser.add_argument('--verbose', action='store_true', help='Print the prompts to the terminal.')
 
 # Model loader
-parser.add_argument('--loader', type=str, help='Choose the model loader manually, otherwise, it will get autodetected. Valid options: transformers, autogptq, gptq-for-llama, exllama, llamacpp, rwkv, flexgen')
+parser.add_argument('--loader', type=str, help='Choose the model loader manually, otherwise, it will get autodetected. Valid options: transformers, autogptq, gptq-for-llama, exllama, exllama_hf, llamacpp, rwkv, flexgen')
 
 # Accelerate/transformers
 parser.add_argument('--cpu', action='store_true', help='Use the CPU to generate text. Warning: Training on CPU is extremely slow.')
@@ -218,6 +218,8 @@ def fix_loader_name(name):
         return 'GPTQ-for-LLaMa'
     elif name in ['exllama', 'ex-llama', 'ex_llama', 'exlama']:
         return 'ExLlama'
+    elif name in ['exllama-hf', 'exllama_hf', 'exllama hf', 'ex-llama-hf', 'ex_llama_hf']:
+        return 'ExLlama_HF'
 
 
 if args.loader is not None:

modules/text_generation.py

@@ -104,9 +104,8 @@ def get_reply_from_output_ids(output_ids, input_ids, original_question, state, i
     else:
         new_tokens = len(output_ids) - len(input_ids[0])
         reply = decode(output_ids[-new_tokens:], state['skip_special_tokens'])
-
         # Prevent LlamaTokenizer from skipping a space
-        if type(shared.tokenizer) is transformers.LlamaTokenizer and len(output_ids) > 0:
+        if type(shared.tokenizer) in [transformers.LlamaTokenizer, transformers.LlamaTokenizerFast] and len(output_ids) > 0:
             if shared.tokenizer.convert_ids_to_tokens(int(output_ids[-new_tokens])).startswith('▁'):
                 reply = ' ' + reply
 

modules/training.py

@@ -11,7 +11,7 @@ import gradio as gr
 import torch
 import transformers
 from datasets import Dataset, load_dataset
-from peft import (LoraConfig, get_peft_model, prepare_model_for_int8_training,
+from peft import (LoraConfig, get_peft_model, prepare_model_for_kbit_training,
                   set_peft_model_state_dict)
 
 from modules import shared, ui, utils
@@ -30,14 +30,17 @@ try:
     MODEL_CLASSES = {v: k for k, v in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES}
 except:
     standard_modules = ["q_proj", "v_proj"]
-    model_to_lora_modules = {"llama": standard_modules, "opt": standard_modules, "gptj": standard_modules, "gpt_neox": ["query_key_value"]}
+    model_to_lora_modules = {"llama": standard_modules, "opt": standard_modules, "gptj": standard_modules, "gpt_neox": ["query_key_value"], "rw":["query_key_value"]}
     MODEL_CLASSES = {
         "LlamaForCausalLM": "llama",
         "OPTForCausalLM": "opt",
         "GPTJForCausalLM": "gptj",
-        "GPTNeoXForCausalLM": "gpt_neox"
+        "GPTNeoXForCausalLM": "gpt_neox",
+        "RWForCausalLM": "rw"
+        
     }
 
+train_log = {}
 
 WANT_INTERRUPT = False
 PARAMETERS = ["lora_name", "always_override", "save_steps", "micro_batch_size", "batch_size", "epochs", "learning_rate", "lr_scheduler_type", "lora_rank", "lora_alpha", "lora_dropout", "cutoff_len", "dataset", "eval_dataset", "format", "eval_steps", "raw_text_file", "overlap_len", "newline_favor_len", "higher_rank_limit", "warmup_steps", "optimizer", "hard_cut_string", "train_only_after"]
@@ -357,7 +360,7 @@ def do_train(lora_name: str, always_override: bool, save_steps: int, micro_batch
     # == Start prepping the model itself ==
     if not hasattr(shared.model, 'lm_head') or hasattr(shared.model.lm_head, 'weight'):
         logger.info("Getting model ready...")
-        prepare_model_for_int8_training(shared.model)
+        prepare_model_for_kbit_training(shared.model)
 
     logger.info("Prepping for training...")
     config = LoraConfig(
@@ -406,12 +409,19 @@ def do_train(lora_name: str, always_override: bool, save_steps: int, micro_batch
                 control.should_training_stop = True
             elif state.global_step > 0 and actual_save_steps > 0 and state.global_step % actual_save_steps == 0:
                 lora_model.save_pretrained(f"{lora_file_path}/checkpoint-{tracked.current_steps}/")
+                 # Save log
+                with open(f"{lora_file_path}/checkpoint-{tracked.current_steps}/training_log.json", 'w', encoding='utf-8') as file:
+                    json.dump(train_log, file, indent=2)    
+
 
         def on_substep_end(self, args: transformers.TrainingArguments, state: transformers.TrainerState, control: transformers.TrainerControl, **kwargs):
             tracked.current_steps += 1
             if WANT_INTERRUPT:
                 control.should_epoch_stop = True
                 control.should_training_stop = True
+        
+        def on_log(self, args: transformers.TrainingArguments, state: transformers.TrainerState, control: transformers.TrainerControl, logs, **kwargs):
+            train_log.update(logs)
 
     trainer = transformers.Trainer(
         model=lora_model,
@@ -448,7 +458,7 @@ def do_train(lora_name: str, always_override: bool, save_steps: int, micro_batch
     # == Save parameters for reuse ==
     with open(f"{lora_file_path}/training_parameters.json", 'w', encoding='utf-8') as file:
         vars = locals()
-        json.dump({x: vars[x] for x in PARAMETERS}, file)
+        json.dump({x: vars[x] for x in PARAMETERS}, file, indent=2)
 
     # == Main run and monitor loop ==
     logger.info("Starting training...")
@@ -462,7 +472,9 @@ def do_train(lora_name: str, always_override: bool, save_steps: int, micro_batch
         # Note: save in the thread in case the gradio thread breaks (eg browser closed)
         lora_model.save_pretrained(lora_file_path)
         logger.info("LoRA training run is completed and saved.")
-        tracked.did_save = True
+        # Save log
+        with open(f"{lora_file_path}/training_log.json", 'w', encoding='utf-8') as file:
+            json.dump(train_log, file, indent=2)
 
     thread = threading.Thread(target=threaded_run)
     thread.start()

requirements.txt

@@ -17,10 +17,10 @@ tqdm
 scipy
 transformers==4.30.2
 git+https://github.com/huggingface/peft@03eb378eb914fbee709ff7c86ba5b1d033b89524
-bitsandbytes==0.39.0; platform_system != "Windows"
+bitsandbytes==0.39.1; platform_system != "Windows"
-https://github.com/jllllll/bitsandbytes-windows-webui/raw/main/bitsandbytes-0.39.0-py3-none-any.whl; platform_system == "Windows"
+https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/bitsandbytes-0.39.1-py3-none-win_amd64.whl; platform_system == "Windows"
-llama-cpp-python==0.1.62; platform_system != "Windows"
+llama-cpp-python==0.1.64; platform_system != "Windows"
-https://github.com/abetlen/llama-cpp-python/releases/download/v0.1.62/llama_cpp_python-0.1.62-cp310-cp310-win_amd64.whl; platform_system == "Windows"
+https://github.com/abetlen/llama-cpp-python/releases/download/v0.1.64/llama_cpp_python-0.1.64-cp310-cp310-win_amd64.whl; platform_system == "Windows"
 https://github.com/PanQiWei/AutoGPTQ/releases/download/v0.2.2/auto_gptq-0.2.2+cu117-cp310-cp310-win_amd64.whl; platform_system == "Windows"
 https://github.com/PanQiWei/AutoGPTQ/releases/download/v0.2.2/auto_gptq-0.2.2+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
 https://github.com/jllllll/exllama/releases/download/0.0.1/exllama-0.0.1+cu117-cp310-cp310-win_amd64.whl; platform_system == "Windows"

server.py

@@ -122,7 +122,7 @@ def count_tokens(text):
         return 'Couldn\'t count the number of tokens. Is a tokenizer loaded?'
 
 
-def download_model_wrapper(repo_id):
+def download_model_wrapper(repo_id, progress=gr.Progress()):
     try:
         downloader_module = importlib.import_module("download-model")
         downloader = downloader_module.ModelDownloader()
@@ -131,6 +131,7 @@ def download_model_wrapper(repo_id):
         branch = repo_id_parts[1] if len(repo_id_parts) > 1 else "main"
         check = False
 
+        progress(0.0)
         yield ("Cleaning up the model/branch names")
         model, branch = downloader.sanitize_model_and_branch_names(model, branch)
 
@@ -141,13 +142,16 @@ def download_model_wrapper(repo_id):
         output_folder = downloader.get_output_folder(model, branch, is_lora)
 
         if check:
+            progress(0.5)
             yield ("Checking previously downloaded files")
             downloader.check_model_files(model, branch, links, sha256, output_folder)
+            progress(1.0)
         else:
             yield (f"Downloading files to {output_folder}")
-            downloader.download_model_files(model, branch, links, sha256, output_folder, threads=1)
+            downloader.download_model_files(model, branch, links, sha256, output_folder, progress_bar=progress, threads=1)
             yield ("Done!")
     except:
+        progress(1.0)
         yield traceback.format_exc()
 
 
@@ -193,7 +197,7 @@ def create_model_menus():
 
     with gr.Row():
         with gr.Column():
-            shared.gradio['loader'] = gr.Dropdown(label="Model loader", choices=["Transformers", "AutoGPTQ", "GPTQ-for-LLaMa", "ExLlama", "llama.cpp"], value=None)
+            shared.gradio['loader'] = gr.Dropdown(label="Model loader", choices=["Transformers", "AutoGPTQ", "GPTQ-for-LLaMa", "ExLlama", "ExLlama_HF", "llama.cpp"], value=None)
             with gr.Box():
                 with gr.Row():
                     with gr.Column():
@@ -233,6 +237,7 @@ def create_model_menus():
                         shared.gradio['trust_remote_code'] = gr.Checkbox(label="trust-remote-code", value=shared.args.trust_remote_code, info='Make sure to inspect the .py files inside the model folder before loading it with this option enabled.')
                         shared.gradio['gptq_for_llama_info'] = gr.Markdown('GPTQ-for-LLaMa is currently 2x faster than AutoGPTQ on some systems. It is installed by default with the one-click installers. Otherwise, it has to be installed manually following the instructions here: [instructions](https://github.com/oobabooga/text-generation-webui/blob/main/docs/GPTQ-models-(4-bit-mode).md#installation-1).')
                         shared.gradio['exllama_info'] = gr.Markdown('ExLlama has to be installed manually. See the instructions here: [instructions](https://github.com/oobabooga/text-generation-webui/blob/main/docs/ExLlama.md).')
+                        shared.gradio['exllama_HF_info'] = gr.Markdown('ExLlama_HF is a wrapper that lets you use ExLlama like a Transformers model, which means it can use the Transformers samplers. It\'s still a bit buggy, so feel free to help out by fixing issues.\n\nCheck out PR [#2777](https://github.com/oobabooga/text-generation-webui/pull/2777) for more details.')
 
         with gr.Column():
             with gr.Row():
@@ -276,7 +281,7 @@ def create_model_menus():
         save_model_settings, [shared.gradio[k] for k in ['model_menu', 'interface_state']], shared.gradio['model_status'], show_progress=False)
 
     shared.gradio['lora_menu_apply'].click(load_lora_wrapper, shared.gradio['lora_menu'], shared.gradio['model_status'], show_progress=False)
-    shared.gradio['download_model_button'].click(download_model_wrapper, shared.gradio['custom_model_menu'], shared.gradio['model_status'], show_progress=False)
+    shared.gradio['download_model_button'].click(download_model_wrapper, shared.gradio['custom_model_menu'], shared.gradio['model_status'], show_progress=True)
     shared.gradio['autoload_model'].change(lambda x: gr.update(visible=not x), shared.gradio['autoload_model'], load)