easydel.inference.tools.parsers.deepseekv3_tool_parser

easydel.inference.tools.parsers.deepseekv3_tool_parser#

class easydel.inference.tools.parsers.deepseekv3_tool_parser.DeepSeekV3ToolParser(tokenizer: AutoTokenizer)[source]#

Bases: ToolParser

Tool parser for DeepSeek V3 models.

This parser handles the specific tool call format used by DeepSeek V3 models, which uses custom delimiters for tool calls and arguments. The format includes: - Tool calls wrapped in <｜tool▁calls▁begin｜> and <｜tool▁calls▁end｜> - Individual calls wrapped in <｜tool▁call▁begin｜> and <｜tool▁call▁end｜> - Tool type and name separated by <｜tool▁sep｜> - Arguments in JSON format wrapped in ```json blocks

current_tool_name_sent#

Tracks if tool name has been sent in streaming

Type

bool

prev_tool_call_arr#

Previous tool calls for comparison in streaming

Type

list

current_tool_id#

Index of current tool being processed

Type

int

streamed_args_for_tool#

Arguments streamed so far for each tool

Type

list

tool_calls_start_token#

Token marking start of tool calls section

Type

str

tool_calls_end_token#

Token marking end of tool calls section

Type

str

tool_call_start_token#

Token marking start of individual tool call

Type

str

tool_call_end_token#

Token marking end of individual tool call

Type

str

extract_tool_calls(model_output: str, request: ChatCompletionRequest) → ExtractedToolCallInformation[source]#

Extract tool calls from complete model output.

Parses the DeepSeek V3 format to extract function names and arguments from the model’s response. Handles the specific token structure with tool type, separator, and JSON argument blocks.

Parameters

  • model_output – Complete text output from the model

  • request – Original chat completion request with tool definitions

Returns

  • tools_called: Whether any tools were invoked

  • tool_calls: List of ToolCall objects with function details

  • content: Text content before tool calls (if any)

Return type

ExtractedToolCallInformation containing

extract_tool_calls_streaming(previous_text: str, current_text: str, delta_text: str, previous_token_ids: Sequence[int], current_token_ids: Sequence[int], delta_token_ids: Sequence[int], request: ChatCompletionRequest) → easydel.inference.openai_api_modules.DeltaMessage | None[source]#

Extract tool calls from streaming model output.

Handles incremental parsing of DeepSeek V3 tool call format during streaming generation. Maintains state across chunks to properly identify tool boundaries and stream arguments progressively.

Parameters

  • previous_text – Text generated up to previous chunk

  • current_text – All text generated so far

  • delta_text – New text in this chunk

  • previous_token_ids – Token IDs up to previous chunk

  • current_token_ids – All token IDs so far

  • delta_token_ids – New token IDs in this chunk

  • request – Original request with tool definitions

Returns

DeltaMessage with incremental tool call information, or None if more data needed for parsing

Contents