diff --git a/demo_function_calling.py b/demo_function_calling.py
new file mode 100644
index 0000000..f36fc64
--- /dev/null
+++ b/demo_function_calling.py
@@ -0,0 +1,107 @@
+"""
+DeepSeek-V3 Function Calling Demo
+
+This script demonstrates how to use function calling with the DeepSeek-V3 model
+after applying the fix for the chat template.
+"""
+
+from transformers import AutoTokenizer
+import json
+import argparse
+
+
+def get_current_temperature(location: str) -> float:
+ """
+ Get the current temperature at a location.
+
+ Args:
+ location: Get the location of the temperature in the format of "city, country"
+ Returns:
+ Displays the current temperature at the specified location as a floating point number.
+ """
+ # This is a mock function that would normally call a weather API
+ print(f"Getting temperature for {location}")
+ return 22.0
+
+
+def get_current_time(timezone: str) -> str:
+ """
+ Get the current time in a specific timezone.
+
+ Args:
+ timezone: The timezone to get the current time for (e.g., "UTC", "America/New_York")
+ Returns:
+ The current time as a string.
+ """
+ # This is a mock function that would normally get the current time
+ print(f"Getting time for timezone {timezone}")
+ return "12:30 PM"
+
+
+def main():
+ parser = argparse.ArgumentParser(
+ description="Test DeepSeek-V3 function calling")
+ parser.add_argument("--model_path", type=str, required=True,
+ help="Path to the DeepSeek-V3 model")
+ args = parser.parse_args()
+
+ print(f"Loading tokenizer from {args.model_path}")
+ tokenizer = AutoTokenizer.from_pretrained(
+ args.model_path, trust_remote_code=True)
+
+ # Example 1: Simple weather query
+ print("\nExample 1: Weather query")
+ tool_call = {"name": "get_current_temperature",
+ "arguments": {"location": "Paris, France"}}
+ messages = [
+ {"role": "system", "content": "You are a helpful assistant that provides weather information."},
+ {"role": "user", "content": "What is the temperature in Paris now?"},
+ {"role": "assistant", "tool_calls": [
+ {"type": "function", "function": tool_call}]},
+ {"role": "user", "content": "Thanks for checking! Is that warm for this time of year?"}
+ ]
+
+ print("Processing chat template...")
+ chat_input = tokenizer.apply_chat_template(
+ messages,
+ tools=[get_current_temperature],
+ add_generation_prompt=True,
+ tokenize=False,
+ tools_in_user_message=False
+ )
+
+ print("\nGenerated chat template:")
+ print("-" * 50)
+ print(chat_input)
+ print("-" * 50)
+
+ # Example 2: Multiple tool calls
+ print("\nExample 2: Multiple function calls")
+ messages = [
+ {"role": "system", "content": "You are a helpful assistant."},
+ {"role": "user", "content": "What's the time in New York and temperature in Tokyo?"},
+ {"role": "assistant", "tool_calls": [
+ {"type": "function", "function": {"name": "get_current_time",
+ "arguments": {"timezone": "America/New_York"}}},
+ {"type": "function", "function": {
+ "name": "get_current_temperature", "arguments": {"location": "Tokyo, Japan"}}}
+ ]},
+ ]
+
+ print("Processing chat template...")
+ chat_input = tokenizer.apply_chat_template(
+ messages,
+ tools=[get_current_time, get_current_temperature],
+ add_generation_prompt=True,
+ tokenize=False,
+ tools_in_user_message=False
+ )
+
+ print("\nGenerated chat template:")
+ print("-" * 50)
+ print(chat_input)
+ print("-" * 50)
+
+
+if __name__ == "__main__":
+ main()
diff --git a/docs/function_calling.md b/docs/function_calling.md
new file mode 100644
index 0000000..e4ab3c0
--- /dev/null
+++ b/docs/function_calling.md
@@ -0,0 +1,80 @@
+# Function Calling with DeepSeek-V3
+
+This document provides guidance on using function calling with DeepSeek-V3 models.
+
+## Overview
+
+Function calling allows the model to call external functions through a structured interface. It's particularly useful for:
+
+- Retrieving real-time information (weather, time, data from APIs)
+- Performing calculations
+- Executing actions based on user requests
+
+## Usage with Transformers
+
+DeepSeek-V3 supports function calling through the Hugging Face Transformers library. The example below demonstrates how to use this feature:
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+# Define your function
+def get_weather(location: str) -> str:
+ """Get the weather for a location."""
+ # In a real application, this would call a weather API
+ return f"Sunny, 22°C in {location}"
+
+# Load the model and tokenizer
+tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V3", trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-V3", trust_remote_code=True)
+
+# Create a conversation with function calling
+messages = [
+ {"role": "system", "content": "You are a helpful assistant."},
+ {"role": "user", "content": "What's the weather like in Tokyo?"},
+ {"role": "assistant", "tool_calls": [
+ {"type": "function", "function": {"name": "get_weather", "arguments": {"location": "Tokyo, Japan"}}}
+ ]},
+ {"role": "user", "content": "Thanks! And what about New York?"}
+]
+
+# Apply the chat template
+inputs = tokenizer.apply_chat_template(
+ messages,
+ tools=[get_weather],
+ add_generation_prompt=True,
+ tokenize=True,
+ tools_in_user_message=False
+)
+
+# Generate a response
+output_ids = model.generate(inputs, max_new_tokens=100)
+response = tokenizer.decode(output_ids[0], skip_special_tokens=True)
+print(response)
+```
+
+## Function Definitions
+
+Functions must have type annotations and docstrings following the OpenAI format:
+
+```python
+def function_name(param1: type, param2: type) -> return_type:
+ """
+ Brief description of what the function does.
+
+ Args:
+ param1: Description of parameter 1
+ param2: Description of parameter 2
+ Returns:
+ Description of what is returned
+ """
+ # Function implementation
+ pass
+```
+
+## Limitations
+
+- Function parameters must be JSON-serializable types
+- Function return values should also be JSON-serializable
+- Complex object types are not directly supported
+
+For more advanced use cases, please refer to the Hugging Face documentation on function calling.
\ No newline at end of file
diff --git a/test_fix.py b/test_fix.py
new file mode 100644
index 0000000..f3ef08b
--- /dev/null
+++ b/test_fix.py
@@ -0,0 +1,126 @@
+from transformers import AutoTokenizer
+import json
+import os
+import sys
+
+# Function to test
+
+
+def get_current_temperature(location: str) -> float:
+ """
+ Get the current temperature at a location.
+
+ Args:
+ location: Get the location of the temperature in the format of "city, country"
+ Returns:
+ Displays the current temperature at the specified location as a floating point number (in the specified unit).
+ """
+ return 22.0
+
+
+def test_with_original_tokenizer(model_path):
+ print("Testing with original tokenizer...")
+ try:
+ tokenizer = AutoTokenizer.from_pretrained(
+ model_path, trust_remote_code=True)
+
+ tool_call = {"name": "get_current_temperature",
+ "arguments": {"location": "Paris, France"}}
+ messages = [
+ {"role": "system", "content": "You are a robot that responds to weather queries."},
+ {"role": "user", "content": "What is the temperature in Paris now?"},
+ {"role": "assistant", "tool_calls": [
+ {"type": "function", "function": tool_call}]},
+ ]
+
+ inputs = tokenizer.apply_chat_template(
+ messages,
+ tools=[get_current_temperature],
+ add_generation_prompt=False,
+ tokenize=False,
+ tools_in_user_message=False
+ )
+ print("Success with original tokenizer!")
+ print(inputs)
+ return True
+ except Exception as e:
+ print(f"Error with original tokenizer: {e}")
+ return False
+
+
+def test_with_fixed_tokenizer(model_path, fixed_config_path):
+ print("Testing with fixed tokenizer config...")
+ try:
+ # Read the original tokenizer files
+ tokenizer_json_path = os.path.join(model_path, "tokenizer.json")
+ if not os.path.exists(tokenizer_json_path):
+ print(f"Error: tokenizer.json not found at {tokenizer_json_path}")
+ return False
+
+ # Copy the tokenizer.json and use our fixed config
+ fixed_dir = "fixed_tokenizer"
+ os.makedirs(fixed_dir, exist_ok=True)
+
+ # Copy tokenizer.json
+ import shutil
+ shutil.copy(tokenizer_json_path, os.path.join(
+ fixed_dir, "tokenizer.json"))
+
+ # Create fixed tokenizer_config.json
+ with open(fixed_config_path, 'r') as f:
+ fixed_config = f.read()
+
+ with open(os.path.join(fixed_dir, "tokenizer_config.json"), 'w') as f:
+ f.write(fixed_config)
+
+ # Load the fixed tokenizer
+ tokenizer = AutoTokenizer.from_pretrained(
+ fixed_dir, trust_remote_code=True)
+
+ tool_call = {"name": "get_current_temperature",
+ "arguments": {"location": "Paris, France"}}
+ messages = [
+ {"role": "system", "content": "You are a robot that responds to weather queries."},
+ {"role": "user", "content": "What is the temperature in Paris now?"},
+ {"role": "assistant", "tool_calls": [
+ {"type": "function", "function": tool_call}]},
+ ]
+
+ inputs = tokenizer.apply_chat_template(
+ messages,
+ tools=[get_current_temperature],
+ add_generation_prompt=False,
+ tokenize=False,
+ tools_in_user_message=False
+ )
+ print("Success with fixed tokenizer!")
+ print(inputs)
+ return True
+ except Exception as e:
+ print(f"Error with fixed tokenizer: {e}")
+ return False
+
+
+def main():
+ if len(sys.argv) < 2:
+ print("Usage: python test_fix.py /path/to/DeepSeek-V3 /path/to/fixed_config.json")
+ return
+
+ model_path = sys.argv[1]
+ fixed_config_path = sys.argv[2] if len(
+ sys.argv) > 2 else "tokenizer_config.json"
+
+ # Test with original tokenizer (should fail)
+ original_success = test_with_original_tokenizer(model_path)
+
+ # Test with fixed tokenizer (should succeed)
+ fixed_success = test_with_fixed_tokenizer(model_path, fixed_config_path)
+
+ if not original_success and fixed_success:
+ print("\n✅ Fix was successful! The issue has been resolved.")
+ else:
+ print("\n❌ Testing did not confirm the fix. Please check the logs above.")
+
+
+if __name__ == "__main__":
+ main()
diff --git a/tokenizer_config.json b/tokenizer_config.json
new file mode 100644
index 0000000..a27583b
--- /dev/null
+++ b/tokenizer_config.json
@@ -0,0 +1,35 @@
+{
+ "add_bos_token": false,
+ "add_eos_token": false,
+ "bos_token": {
+ "__type": "AddedToken",
+ "content": "<｜begin of sentence｜>",
+ "lstrip": false,
+ "normalized": true,
+ "rstrip": false,
+ "single_word": false
+ },
+ "chat_template": "{% set ns = namespace(found=false, is_first=false, is_tool=false) %}\n{%- for message in messages %}\n {%- if message['role'] == 'system' %}\n {{ '<｜begin of sentence｜><｜system｜>' + message['content'] + '<｜end of sentence｜>' }}\n {%- elif message['role'] == 'user' %}\n {%- if message.get('content') and message.get('tool_results') %}\n {{ '<｜begin of sentence｜><｜User｜>' + message['content'] }}\n {%- for result in message['tool_results'] %}\n {{ '\n<｜tool result begin｜>' + result['name'] + '\n```json\n' + result['result'] + '\n```<｜tool result end｜>' }}\n {%- endfor %}\n {{ '<｜end of sentence｜>' }}\n {%- elif message.get('content') %}\n {{ '<｜begin of sentence｜><｜User｜>' + message['content'] + '<｜end of sentence｜>' }}\n {%- endif %}\n {%- elif message['role'] == 'assistant' %}\n {%- if message.get('content') is none or message.get('content') == '' %}\n {%- set ns.is_tool = false -%}\n {%- for tool in message.get('tool_calls', []) %}\n {%- if not ns.is_first %}\n {{ '<｜Assistant｜><｜tool calls begin｜><｜tool call begin｜>' + tool['type'] + '<｜tool sep｜>' + tool['function']['name'] + '\n```json\n' + (tool['function']['arguments']|tojson) + '\n```<｜tool call end｜>' }}\n {%- set ns.is_first = true -%}\n {%- else %}\n {{ '\n<｜tool call begin｜>' + tool['type'] + '<｜tool sep｜>' + tool['function']['name'] + '\n```json\n' + (tool['function']['arguments']|tojson) + '\n```<｜tool call end｜>' }}\n {{ '<｜tool calls end｜><｜end of sentence｜>' }}\n {%- endif %}\n {%- endfor %}\n {%- else %}\n {{ '<｜begin of sentence｜><｜Assistant｜>' + message['content'] + '<｜end of sentence｜>' }}\n {%- endif %}\n {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n {{ '<｜begin of sentence｜><｜Assistant｜>' }}\n{%- endif %}\n",
+ "clean_up_tokenization_spaces": true,
+ "model_max_length": 131072,
+ "eos_token": {
+ "__type": "AddedToken",
+ "content": "<｜end of sentence｜>",
+ "lstrip": false,
+ "normalized": true,
+ "rstrip": false,
+ "single_word": false
+ },
+ "tokenizer_class": "CodeLlamaTokenizer",
+ "pad_token": null,
+ "special_tokens_map_file": null,
+ "tokenizer_file": "tokenizer.json",
+ "unk_token": {
+ "__type": "AddedToken",
+ "content": "<｜unknown｜>",
+ "lstrip": false,
+ "normalized": true,
+ "rstrip": false,
+ "single_word": false
+ }
+}
\ No newline at end of file