Trace Module Design Document

Overview

The Trace module is responsible for completely recording the inference traces of the large model during the AI Kernel generation process. It implements the tracking and saving functions for operations of components like Designer, Coder, and Verifier. Through standardized log formats and storage mechanisms, it supports the traceability and problem diagnosis of the generation process.

Class Method Descriptions

insert_designer_or_coder_record() and insert_verifier_record()

Function

• Records the inference process and generation results of the large model and saves them to a specified directory.

Parameter Description

Parameter Name	Type	Description
res	str	The raw response from the large model (including code and comments).
prompt	str	The complete prompt template for this operation.
reasoning	str	The reasoning process text from the large model.
action_type	ActionType	The type of action.

Stored Content

• result.txt: The parsed pure code + commented document.
• prompt.txt: The complete prompt template.
• reasoning.txt: The reasoning process text.

insert_verifier_record()

Function

• Records test results and performance analysis data.

Parameter Description

Parameter Name	Type	Description
verify_res	str	The verification result.
verify_log	str	Detailed verification log.
profile	str	Performance analysis data (JSON format).
action_type	ActionType	The type of action.

Stored Content

• error_log.txt: Detailed verification log.

File Naming Convention

I{TaskID}_S{StepNumber:02d}_{KernelName}_{ActionType}_{ParameterName}.txt

Example

I123_S03_exp_add_op_DO_DESIGNER_prompt.txt
I123_S04_exp_add_op_DO_CODER_result.txt

File Storage Specification

Directory Structure

log_dir/
└── op_name/
    ├── I{task_id}_S{step}_{op_name}_{action_type}_prompt.txt     # Complete prompt template
    ├── I{task_id}_S{step}_{op_name}_{action_type}_reasoning.txt  # Reasoning process text
    ├── I{task_id}_S{step}_{op_name}_{action_type}_result.txt     # Parsed pure code + commented document
    ├── I{task_id}_S{step}_{op_name}_{action_type}_error_log.txt  # Detailed verification log
    └── conductor/
        ├── I{task_id}_S{step}_{op_name}_CheckDesigner_prompt.txt  # designer self_check prompt
        ├── I{task_id}_S{step}_{op_name}_CheckDesigner_result.txt  # designer self_check result
        ├── I{task_id}_S{step}_{op_name}_CheckCoder_prompt.txt     # coder self_check prompt
        ├── I{task_id}_S{step}_{op_name}_CheckCoder_result.txt     # coder self_check result
        └── I{task_id}_S{step}_{op_name}_AnalyzeError_prompt.txt   # Error analysis prompt
        └── I{task_id}_S{step}_{op_name}_AnalyzeError_result.txt   # Error analysis result
    └── I{task_id}_S{step}_verify/
        ├── {op_name}_{framework}.py  # Original task mindspore/torch/numpy implementation
        ├── {op_name}_{impl_type}.py  # swft/triton implementation
        └── verify_{op_name}.py       # Error analysis

Usage Example

trace = Trace(op_name="swish", task_id="123", log_dir="~/logs")

trace.insert_designer_or_coder_record(
    res=designer_response,
    prompt=prompt_template,
    reasoning=llm_reasoning,
    action_type=ActionType.DO_DESIGNER
)