Merge pull request #74 from pabl-o-ce/deepseek

Add: DeepseekCoder v2 chat format

Author: Maximilian-Winter

ReadMe.md

@@ -1,7 +1,7 @@
 # llama-cpp-agent
 
 [![PyPI - Version](https://img.shields.io/pypi/v/llama-cpp-agent?logo=pypi&color=%2341bb13)](https://pypi.org/project/llama-cpp-agent/)
-[![Discord](https://img.shields.io/discord/1237393014154985582?logo=Discord&logoColor=%23ffffff&label=Discord&link=https%3A%2F%2Fdiscord.gg%2FsRMvWKrh)](https://discord.gg/sRMvWKrh)
+[![Discord](https://img.shields.io/discord/1237393014154985582?logo=Discord&logoColor=%23ffffff&label=Discord&link=https%3A%2F%2Fdiscord.gg%2FsRMvWKrh)](https://discord.gg/fgr5RycPFP)
 
 ![llama-cpp-agent logo](https://raw.githubusercontent.com/Maximilian-Winter/llama-cpp-agent/master/logo/logo_orange.png)
 
@@ -126,6 +126,8 @@ The llama-cpp-agent framework provides predefined message formatters to format m
 - `MessagesFormatterType.B22`: Formats messages using the B22 format.
 - `MessagesFormatterType.LLAMA_3`: Formats messages using the LLAMA 3 format.
 - `MessagesFormatterType.PHI_3`: Formats messages using the PHI 3 format.
+- `MessagesFormatterType.AUTOCODER`: Formats messages using the Autocoder format.
+- `MessagesFormatterType.DEEP_SEEK_CODER_2`: Formats messages using the DeepSeek Coder v2 format.
 
 ### Creating Custom Messages Formatter
 

src/llama_cpp_agent/messages_formatter.py

@@ -27,6 +27,7 @@ class MessagesFormatterType(Enum):
     OPEN_INTERPRETER = 14
     AUTOCODER = 15
     GEMMA_2 = 16
+    DEEP_SEEK_CODER_2 = 17
 
 @dataclass
 class PromptMarkers:
@@ -236,6 +237,12 @@ def _format_response(
     Roles.assistant: PromptMarkers("Assistant: ", "<|EOT|>\n"),
     Roles.tool: PromptMarkers("", ""),
 }
+deep_seek_coder_2_chat_prompt_markers = {
+    Roles.system: PromptMarkers("""<｜begin▁of▁sentence｜>""", """\n\n"""),
+    Roles.user: PromptMarkers("""User: """, """ \n\n"""),
+    Roles.assistant: PromptMarkers("""Assistant: """, """<｜end▁of▁sentence｜>"""),
+    Roles.tool: PromptMarkers("", ""),
+}
 
 """
 ### Instruction:
@@ -358,7 +365,22 @@ def _format_response(
     eos_token="<|EOT|>",
 )
 
-gemma_2_chat_formatter = MessagesFormatter("", gemma_2_prompt_markers, True, ["<end_of_turn>", "<start_of_turn>"])
+gemma_2_chat_formatter = MessagesFormatter(
+    "",
+    gemma_2_prompt_markers,
+    True,
+    ["<end_of_turn>", "<start_of_turn>"]
+)
+
+deep_seek_coder_2_chat_formatter = MessagesFormatter(
+    "",
+    deep_seek_coder_2_chat_prompt_markers,
+    True,
+    ["<｜end▁of▁sentence｜>"],
+    bos_token="<｜begin▁of▁sentence｜>",
+    eos_token="<｜end▁of▁sentence｜>",
+)
+
 predefined_formatter = {
     MessagesFormatterType.MISTRAL: mixtral_formatter,
     MessagesFormatterType.CHATML: chatml_formatter,
@@ -375,7 +397,8 @@ def _format_response(
     MessagesFormatterType.PHI_3: phi_3_chat_formatter,
     MessagesFormatterType.OPEN_INTERPRETER: open_interpreter_chat_formatter,
     MessagesFormatterType.AUTOCODER: autocoder_chat_formatter,
-    MessagesFormatterType.GEMMA_2: gemma_2_chat_formatter
+    MessagesFormatterType.GEMMA_2: gemma_2_chat_formatter,
+    MessagesFormatterType.DEEP_SEEK_CODER_2: deep_seek_coder_2_chat_formatter
 }