Merge pull request #5073 from oobabooga/dev

Merge dev branch

extensions/multimodal/README.md

@@ -67,8 +67,56 @@ This extension uses the following parameters (from `settings.json`):
 
 ## Usage through API
 
+### Chat completions endpoint
+
+#### With an image URL
+
+```shell
+curl http://127.0.0.1:5000/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "messages": [
+      {
+        "role": "user",
+        "image_url": "https://avatars.githubusercontent.com/u/112222186?v=4"
+      },
+      {
+        "role": "user",
+        "content": "What is unusual about this image?"
+      }
+    ]
+  }'
+```
+
+#### With a Base64 image
+
+```python
+import base64
+import json
+import requests
+
+img = open('image.jpg', 'rb')
+img_bytes = img.read()
+img_base64 = base64.b64encode(img_bytes).decode('utf-8')
+data = { "messages": [
+        {
+            "role": "user",
+            "image_url": f"data:image/jpeg;base64,{img_base64}"
+        },
+        {
+            "role": "user",
+            "content": "what is unusual about this image?"
+        }
+    ]
+}
+response = requests.post('http://127.0.0.1:5000/v1/chat/completions', json=data)
+print(response.text)
+```
+
 You can run the multimodal inference through API, by inputting the images to prompt. Images are embedded like so: `f'<img src="data:image/jpeg;base64,{img_str}">'`, where `img_str` is base-64 jpeg data. Note that you will need to launch `server.py` with the arguments `--api --extensions multimodal`. 
 
+### Completions endpoint
+
 Python example:
 
 ```Python

extensions/openai/completions.py

@@ -1,10 +1,15 @@
+import base64
 import copy
+import re
 import time
 from collections import deque
+from io import BytesIO
 
+import requests
 import tiktoken
 import torch
 import torch.nn.functional as F
+from PIL import Image
 from transformers import LogitsProcessor, LogitsProcessorList
 
 from extensions.openai.errors import InvalidRequestError
@@ -140,7 +145,25 @@ def convert_history(history):
     system_message = ""
 
     for entry in history:
-        content = entry["content"]
+        if "image_url" in entry:
+            image_url = entry['image_url']
+            if "base64" in image_url:
+                image_url = re.sub('^data:image/.+;base64,', '', image_url)
+                img = Image.open(BytesIO(base64.b64decode(image_url)))
+            else:
+                try:
+                    my_res = requests.get(image_url)
+                    img = Image.open(BytesIO(my_res.content))
+                except Exception:
+                    raise 'Image cannot be loaded from the URL!'
+
+            buffered = BytesIO()
+            img.save(buffered, format="JPEG")
+            img_str = base64.b64encode(buffered.getvalue()).decode('utf-8')
+            content = f'<img src="data:image/jpeg;base64,{img_str}">'
+        else:
+            content = entry["content"]
+
         role = entry["role"]
 
         if role == "user":
@@ -182,7 +205,8 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False) -
             raise InvalidRequestError(message="messages: missing role", param='messages')
         elif m['role'] == 'function':
             raise InvalidRequestError(message="role: function is not supported.", param='messages')
-        if 'content' not in m:
+
+        if 'content' not in m and "image_url" not in m:
             raise InvalidRequestError(message="messages: missing content", param='messages')
 
     # Chat Completions

instruction-templates/Synthia-CoT.yaml

@@ -0,0 +1,25 @@
+instruction_template: |-
+  {%- set found_item = false -%}
+  {%- for message in messages -%}
+      {%- if message['role'] == 'system' -%}
+          {%- set found_item = true -%}
+      {%- endif -%}
+  {%- endfor -%}
+  {%- if not found_item -%}
+      {{-'SYSTEM: ' + 'Elaborate on the topic using a Tree of Thoughts and backtrack when necessary to construct a clear, cohesive Chain of Thought reasoning. Always answer without hesitation.' + '\n' -}}
+  {%- endif %}
+  {%- for message in messages %}
+      {%- if message['role'] == 'system' -%}
+          {{-'SYSTEM: ' + message['content'] + '\n' -}}
+      {%- else -%}
+          {%- if message['role'] == 'user' -%}
+              {{-'USER: ' + message['content'] + '\n'-}}
+          {%- else -%}
+              {{-'ASSISTANT: ' + message['content'] + '</s>\n' -}}
+          {%- endif -%}
+      {%- endif -%}
+  {%- endfor -%}
+  {%- if add_generation_prompt -%}
+      {{-'ASSISTANT:'-}}
+  {%- endif -%}
+

instruction-templates/Synthia.yaml

@@ -0,0 +1,25 @@
+instruction_template: |-
+  {%- set found_item = false -%}
+  {%- for message in messages -%}
+      {%- if message['role'] == 'system' -%}
+          {%- set found_item = true -%}
+      {%- endif -%}
+  {%- endfor -%}
+  {%- if not found_item -%}
+      {{-'SYSTEM: ' + 'Answer the question thoughtfully and intelligently. Always answer without hesitation.' + '\n' -}}
+  {%- endif %}
+  {%- for message in messages %}
+      {%- if message['role'] == 'system' -%}
+          {{-'SYSTEM: ' + message['content'] + '\n' -}}
+      {%- else -%}
+          {%- if message['role'] == 'user' -%}
+              {{-'USER: ' + message['content'] + '\n'-}}
+          {%- else -%}
+              {{-'ASSISTANT: ' + message['content'] + '</s>\n' -}}
+          {%- endif -%}
+      {%- endif -%}
+  {%- endfor -%}
+  {%- if add_generation_prompt -%}
+      {{-'ASSISTANT:'-}}
+  {%- endif -%}
+

models/config.yaml

@@ -188,3 +188,5 @@
   instruction_template: 'ChatML'
 (dolphin).*:
   instruction_template: 'ChatML'
+.*synthia:
+  instruction_template: 'Synthia'

modules/models.py

@@ -482,6 +482,7 @@ def clear_torch_cache():
 
 def unload_model():
     shared.model = shared.tokenizer = None
+    shared.model_name = 'None'
     shared.lora_names = []
     shared.model_dirty_from_training = False
     clear_torch_cache()

modules/shared.py

@@ -45,6 +45,7 @@ settings = {
     'truncation_length_min': 0,
     'truncation_length_max': 200000,
     'max_tokens_second': 0,
+    'max_updates_second': 0,
     'custom_stopping_strings': '',
     'custom_token_bans': '',
     'auto_max_new_tokens': False,

modules/text_generation.py

@@ -77,6 +77,10 @@ def _generate_reply(question, state, stopping_strings=None, is_chat=False, escap
         state = copy.deepcopy(state)
         state['stream'] = True
 
+    min_update_interval = 0
+    if state.get('max_updates_second', 0) > 0:
+        min_update_interval = 1 / state['max_updates_second']
+
     # Generate
     for reply in generate_func(question, original_question, seed, state, stopping_strings, is_chat=is_chat):
         reply, stop_found = apply_stopping_strings(reply, all_stop_strings)
@@ -94,10 +98,9 @@ def _generate_reply(question, state, stopping_strings=None, is_chat=False, escap
                 last_update = time.time()
                 yield reply
 
-            # Limit updates to 24 or 5 per second to avoid lag in the Gradio UI
+            # Limit updates to avoid lag in the Gradio UI
             # API updates are not limited
             else:
-                min_update_interval = 0 if not for_ui else 0.2 if (shared.args.listen or shared.args.share) else 0.0417
                 if cur_time - last_update > min_update_interval:
                     last_update = cur_time
                     yield reply
@@ -265,8 +268,15 @@ def apply_stopping_strings(reply, all_stop_strings):
 
 def get_reply_from_output_ids(output_ids, state, starting_from=0):
     reply = decode(output_ids[starting_from:], state['skip_special_tokens'])
-    if (hasattr(shared.tokenizer, 'convert_ids_to_tokens') and len(output_ids) > starting_from and shared.tokenizer.convert_ids_to_tokens(int(output_ids[starting_from])).startswith('▁')) and not reply.startswith(' '):
+
-        reply = ' ' + reply
+    # Handle tokenizers that do not add the leading space for the first token
+    if (hasattr(shared.tokenizer, 'convert_ids_to_tokens') and len(output_ids) > starting_from) and not reply.startswith(' '):
+        first_token = shared.tokenizer.convert_ids_to_tokens(int(output_ids[starting_from]))
+        if isinstance(first_token, (bytes,)):
+            first_token = first_token.decode('utf8')
+
+        if first_token.startswith('▁'):
+            reply = ' ' + reply
 
     return reply
 

modules/ui.py

@@ -110,6 +110,7 @@ def list_interface_input_elements():
         'max_new_tokens',
         'auto_max_new_tokens',
         'max_tokens_second',
+        'max_updates_second',
         'seed',
         'temperature',
         'temperature_last',

modules/ui_parameters.py

@@ -66,7 +66,9 @@ def create_ui(default_preset):
                     with gr.Row():
                         with gr.Column():
                             shared.gradio['truncation_length'] = gr.Slider(value=get_truncation_length(), minimum=shared.settings['truncation_length_min'], maximum=shared.settings['truncation_length_max'], step=256, label='Truncate the prompt up to this length', info='The leftmost tokens are removed if the prompt exceeds this length. Most models require this to be at most 2048.')
-                            shared.gradio['max_tokens_second'] = gr.Slider(value=shared.settings['max_tokens_second'], minimum=0, maximum=20, step=1, label='Maximum number of tokens/second', info='To make text readable in real time.')
+                            shared.gradio['max_tokens_second'] = gr.Slider(value=shared.settings['max_tokens_second'], minimum=0, maximum=20, step=1, label='Maximum tokens/second', info='To make text readable in real time.')
+                            shared.gradio['max_updates_second'] = gr.Slider(value=shared.settings['max_updates_second'], minimum=0, maximum=24, step=1, label='Maximum UI updates/second', info='Set this if you experience lag in the UI during streaming.')
+
                             shared.gradio['custom_stopping_strings'] = gr.Textbox(lines=1, value=shared.settings["custom_stopping_strings"] or None, label='Custom stopping strings', info='In addition to the defaults. Written between "" and separated by commas.', placeholder='"\\n", "\\nYou:"')
                             shared.gradio['custom_token_bans'] = gr.Textbox(value=shared.settings['custom_token_bans'] or None, label='Custom token bans', info='Specific token IDs to ban from generating, comma-separated. The IDs can be found in the Default or Notebook tab.')
 

requirements.txt

@@ -5,6 +5,7 @@ einops
 exllamav2==0.0.11; platform_system != "Darwin" and platform_machine != "x86_64"
 gradio==3.50.*
 hqq==0.1.1.post1
+lm_eval==0.3.0
 markdown
 numpy==1.24.*
 optimum==1.16.*
@@ -98,4 +99,4 @@ https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_
 https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+cu121-cp39-cp39-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.9"
 https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+cu121-cp38-cp38-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.8"
 https://github.com/jllllll/ctransformers-cuBLAS-wheels/releases/download/AVX2/ctransformers-0.2.27+cu121-py3-none-any.whl
-autoawq==0.1.7; platform_system == "Linux" or platform_system == "Windows"
+autoawq==0.1.8; platform_system == "Linux" or platform_system == "Windows"

requirements_amd.txt

@@ -5,6 +5,7 @@ einops
 exllamav2==0.0.11; platform_system == "Windows" or python_version < "3.10" or python_version > "3.11" or platform_machine != "x86_64"
 gradio==3.50.*
 hqq==0.1.1.post1
+lm_eval==0.3.0
 markdown
 numpy==1.24.*
 optimum==1.16.*

requirements_amd_noavx2.txt

@@ -5,6 +5,7 @@ einops
 exllamav2==0.0.11; platform_system == "Windows" or python_version < "3.10" or python_version > "3.11" or platform_machine != "x86_64"
 gradio==3.50.*
 hqq==0.1.1.post1
+lm_eval==0.3.0
 markdown
 numpy==1.24.*
 optimum==1.16.*

requirements_apple_intel.txt

@@ -5,6 +5,7 @@ einops
 exllamav2==0.0.11
 gradio==3.50.*
 hqq==0.1.1.post1
+lm_eval==0.3.0
 markdown
 numpy==1.24.*
 optimum==1.16.*

requirements_apple_silicon.txt

@@ -5,6 +5,7 @@ einops
 exllamav2==0.0.11
 gradio==3.50.*
 hqq==0.1.1.post1
+lm_eval==0.3.0
 markdown
 numpy==1.24.*
 optimum==1.16.*

requirements_cpu_only.txt

@@ -5,6 +5,7 @@ einops
 exllamav2==0.0.11
 gradio==3.50.*
 hqq==0.1.1.post1
+lm_eval==0.3.0
 markdown
 numpy==1.24.*
 optimum==1.16.*

requirements_cpu_only_noavx2.txt

@@ -5,6 +5,7 @@ einops
 exllamav2==0.0.11
 gradio==3.50.*
 hqq==0.1.1.post1
+lm_eval==0.3.0
 markdown
 numpy==1.24.*
 optimum==1.16.*

requirements_noavx2.txt

@@ -5,6 +5,7 @@ einops
 exllamav2==0.0.11; platform_system != "Darwin" and platform_machine != "x86_64"
 gradio==3.50.*
 hqq==0.1.1.post1
+lm_eval==0.3.0
 markdown
 numpy==1.24.*
 optimum==1.16.*
@@ -98,4 +99,4 @@ https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_
 https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+cu121-cp39-cp39-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.9"
 https://github.com/jllllll/GPTQ-for-LLaMa-CUDA/releases/download/0.1.1/gptq_for_llama-0.1.1+cu121-cp38-cp38-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64" and python_version == "3.8"
 https://github.com/jllllll/ctransformers-cuBLAS-wheels/releases/download/AVX/ctransformers-0.2.27+cu121-py3-none-any.whl
-autoawq==0.1.7; platform_system == "Linux" or platform_system == "Windows"
+autoawq==0.1.8; platform_system == "Linux" or platform_system == "Windows"

requirements_nowheels.txt

@@ -5,6 +5,7 @@ einops
 exllamav2==0.0.11
 gradio==3.50.*
 hqq==0.1.1.post1
+lm_eval==0.3.0
 markdown
 numpy==1.24.*
 optimum==1.16.*