Add code to push spaces chatbot

create_data.py

@@ -16,6 +16,8 @@
 import numpy as np
 from tqdm import tqdm
 
+from utils import flatten_list
+
 
 def parse_rst_file(filepath):
     with open(filepath, 'r') as f:
@@ -525,17 +527,6 @@ def test_show_prompts():
             print(generate_prompt(data_point, 'plain', False, False)[0])
 
 
-def flatten_list(lis):
-    """Given a list, possibly nested to any level, return it flattened."""
-    new_lis = []
-    for item in lis:
-        if type(item) == type([]):
-            new_lis.extend(flatten_list(item))
-        else:
-            new_lis.append(item)
-    return new_lis
-
-
 def test_get_open_datasets():
     # HF changed things so don't get raw list of all datasets, so not have to filter, but can't do negative filter
     open_tags = ['license:Apache License 2.0',

finetune.py

@@ -184,19 +184,24 @@ def train(
         ddp: bool = True,  # set to False if OOM with True, for multi-GPU model parallelism
         local_files_only: bool = False,  # else will download new versions, normally unwanted
         resume_download: bool = True,
-        use_auth_token: bool = False,  # True requires CLI did huggingface-cli login before running
+        use_auth_token: Union[str, bool] = False,  # True requires CLI did huggingface-cli login before running
         warmup_steps: int = 100,
         logging_steps: int = 1,
         save_steps: int = None,  # must be round multiple of eval_steps
         add_eos_token: bool = False,
 ):
+    # allow set token directly
+    use_auth_token = os.environ.get("HUGGINGFACE_API_TOKEN", use_auth_token)
+
     prompt_type = str(prompt_type)  # migration from integers
     assert prompt_type in prompt_types
+
     world_size = int(os.getenv("WORLD_SIZE", 1))
     local_rank = int(os.getenv("LOCAL_RANK", 0))
     rank = int(os.getenv("RANK", 0))
     print(f"local_rank: {local_rank}")
     print(f"global rank: {rank}")
+
     gpus = max(world_size, torch.cuda.device_count())
     run_id = run_id or 0
     if not data_path:

generate.py

@@ -5,8 +5,8 @@
 import os
 import typing
 
+from utils import set_seed, flatten_list
 SEED = 1236
-from utils import set_seed
 set_seed(SEED)
 
 os.environ['HF_HUB_DISABLE_TELEMETRY'] = '1'
@@ -55,7 +55,7 @@ def main(
         share: bool = True,
         local_files_only: bool = False,
         resume_download: bool = True,
-        use_auth_token: bool = False,  # True requires CLI did huggingface-cli login before running
+        use_auth_token: Union[str, bool] = False,  # True requires CLI did huggingface-cli login before running
 
         src_lang: str = "English",
         tgt_lang: str = "Russian",
@@ -86,6 +86,13 @@ def main(
         eval_sharegpt_prompts_only_seed: int = 1234,
         eval_sharegpt_as_output: bool = False,
 ):
+    # allow set token directly
+    use_auth_token = os.environ.get("HUGGINGFACE_API_TOKEN", use_auth_token)
+    # override share if in spaces
+    if os.environ.get("HUGGINGFACE_SPACES"):
+        share = False
+        base_model = 'h2oai/h2ogpt-oasst1-512-12b'
+        load_8bit = True
 
     # get defaults
     model_lower = base_model.lower()
@@ -246,7 +253,7 @@ def get_device():
     return device
 
 
-def get_non_lora_model(base_model, model_loader, load_half, model_kwargs, reward_type, force_1_gpu=True, use_auth_token=True):
+def get_non_lora_model(base_model, model_loader, load_half, model_kwargs, reward_type, force_1_gpu=True, use_auth_token=False):
     """
     Ensure model gets on correct device
     :param base_model:
@@ -316,7 +323,7 @@ def get_model(
         reward_type: bool = None,
         local_files_only: bool = False,
         resume_download: bool = True,
-        use_auth_token: bool = True,
+        use_auth_token: Union[str, bool] = False,
         compile: bool = True,
         **kwargs,
 ):
@@ -499,7 +506,12 @@ def go_gradio(**kwargs):
                       Hash: {get_githash()}
                       """
     else:
-        description = ""
+        description = "For more information, visit [the project's website](https://github.com/h2oai/h2ogpt).<br>"
+    if os.environ.get("HUGGINGFACE_SPACES"):
+        description += """<p><b> DISCLAIMERS: </b><ul><i><li>The data used to train this model include The Pile and other sources. These may contain objectionable content, so the model may reproduce that material. Use application and responses at own risk.</i></li>"""
+        if kwargs['load_8bit']:
+            description += """<i><li> Model is loaded in 8-bit to fit on HF GPUs, so model may perform worse than 16-bit.</i></li>"""
+        description += """<i><li>Model loading and unloading disabled on HF SPACES to avoid GPU OOM for multi-user environment.</i></li></ul></p>"""
 
     if kwargs['verbose']:
         task_info_md = f"""
@@ -534,7 +546,6 @@ def go_gradio(**kwargs):
     # css_code = 'body{background-image:url("https://h2o.ai/content/experience-fragments/h2o/us/en/site/header/master/_jcr_content/root/container/header_copy/logo.coreimg.svg/1678976605175/h2o-logo.svg");}'
     # demo = gr.Blocks(theme='gstaff/xkcd', css=css_code)
 
-    from create_data import flatten_list
     model_options = flatten_list(list(prompt_type_to_model_name.values())) + kwargs['extra_model_options']
     if kwargs['base_model'].strip() not in model_options:
         lora_options = [kwargs['base_model'].strip()] + model_options
@@ -676,7 +687,9 @@ def go_gradio(**kwargs):
                                     model_choice = gr.Dropdown(model_options_state.value[0], label="Choose Model", value=kwargs['base_model'])
                                     lora_choice = gr.Dropdown(lora_options_state.value[0], label="Choose LORA", value=kwargs['lora_weights'], visible=kwargs['show_lora'])
                                 with gr.Column(scale=1):
-                                    load_model_button = gr.Button("Load Model/LORA")
+                                    load_msg = "Load Model/LORA" if not os.environ.get("HUGGINGFACE_SPACES") \
+                                        else "LOAD DISABLED ON HF SPACES"
+                                    load_model_button = gr.Button(load_msg)
                                     model_used = gr.Textbox(label="Current Model", value=kwargs['base_model'])
                                     lora_used = gr.Textbox(label="Current LORA", value=kwargs['lora_weights'], visible=kwargs['show_lora'])
                             with gr.Row(scale=1):
@@ -941,7 +954,8 @@ def chatbot_list(x, model_used_in):
                                outputs=[model_state, model_used, lora_used, prompt_type])
         prompt_update_args = dict(fn=dropdown_prompt_type_list, inputs=prompt_type, outputs=prompt_type)
         chatbot_update_args = dict(fn=chatbot_list, inputs=[text_output, model_used], outputs=text_output)
-        load_model_event = load_model_button.click(**load_model_args).then(**prompt_update_args).then(**chatbot_update_args)
+        if not os.environ.get("HUGGINGFACE_SPACES"):
+            load_model_event = load_model_button.click(**load_model_args).then(**prompt_update_args).then(**chatbot_update_args)
 
         def dropdown_model_list(list0, x):
             new_state = [list0[0] + [x]]

spaces/chatbot/LICENSE

@@ -0,0 +1 @@
+../../LICENSE

spaces/chatbot/client_test.py

@@ -0,0 +1 @@
+../../client_test.py

spaces/chatbot/finetune.py

@@ -0,0 +1 @@
+../../finetune.py

spaces/chatbot/generate.py

@@ -0,0 +1 @@
+../../generate.py

spaces/chatbot/h2o-logo.svg

@@ -0,0 +1 @@
+../../h2o-logo.svg

spaces/chatbot/prompter.py

@@ -0,0 +1 @@
+../../prompter.py

spaces/chatbot/repo_to_spaces.sh

@@ -0,0 +1,19 @@
+#!/bin/bash
+
+# start in h2ogpt repo
+ln -sr generate.py h2o-logo.svg LICENSE stopping.py prompter.py finetune.py utils.py client_test.py requirements.txt spaces/chatbot/
+cd ..
+
+git clone https://huggingface.co/spaces/h2oai/h2ogpt-chatbot
+cd h2ogpt-chatbot
+rm -rf app.py h2o-logo.svg LICENSE stopping.py prompter.py finetune.py utils.py client_test.py requirements.txt
+cd ../h2ogpt/spaces/chatbot/
+cp generate.py h2o-logo.svg LICENSE stopping.py prompter.py finetune.py utils.py client_test.py requirements.txt ../../../h2ogpt-chatbot/
+cd ../../../h2ogpt-chatbot/
+
+mv generate.py app.py
+
+git add app.py h2o-logo.svg LICENSE stopping.py prompter.py finetune.py utils.py client_test.py requirements.txt
+git commit -m "Add application file and dependencies"
+# ensure write token used and login with git control: huggingface-cli login --token <HUGGINGFACE_API_TOKEN> --add-to-git-credential
+git push

spaces/chatbot/requirements.txt

@@ -0,0 +1 @@
+../../requirements.txt

spaces/chatbot/stopping.py

@@ -0,0 +1 @@
+../../stopping.py

spaces/chatbot/utils.py

@@ -0,0 +1 @@
+../../utils.py

spaces/demo/h2oai_pipeline.py

@@ -0,0 +1 @@
+../../h2oai_pipeline.py

utils.py

@@ -18,3 +18,14 @@ def set_seed(seed: int):
     torch.backends.cudnn.benchmark = False
     os.environ['PYTHONHASHSEED'] = str(seed)
     return random_state
+
+
+def flatten_list(lis):
+    """Given a list, possibly nested to any level, return it flattened."""
+    new_lis = []
+    for item in lis:
+        if type(item) == type([]):
+            new_lis.extend(flatten_list(item))
+        else:
+            new_lis.append(item)
+    return new_lis