ChatHLS-HLSFixer

Model Details

• Finetuned from model: Qwen/Qwen2.5-14B-Instruct
• Model source paper: ChatHLS: Towards Systematic Design Automation and Optimization for High-Level Synthesis
• Project repository: GEAR-SEU/ChatHLS-ACL-26

Description

ChatHLS-HLSFixer is a specialized Large Language Model fine-tuned for High-Level Synthesis (HLS) C/C++ error correction. It acts as the core analysis agent within the ChatHLS framework.

Quickstart

Prerequisites

pip install transformers torch accelerate

Example

Here is a simple example of how to format your prompt and generate debugging instructions for buggy HLS code:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "XXXiong/ChatHLS-HLSFixer"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

buggy_code = """
#include "ap_fixed.h"
void vector_mul(ap_fixed<32,16> a[8], ap_fixed<32,16> b[8], ap_fixed<32,16> c[8]) {
    #pragma HLS ARRAY_PARTITION variable=a complete dim=1
    #pragma HLS ARRAY_PARTITION variable=b complete dim=1
    #pragma HLS ARRAY_PARTITION variable=c complete dim=1
    for (int i = 0; i < 8; ++i) {
        #pragma HLS UNROLL
        #pragma HLS PIPELINE II=1
        c[i] = a[i] * b[i];
    }
}
"""

error_log = """
    "ERROR: [HLS 214-274] In 'VITIS_LOOP_7_1', Pragma conflict happens on 'UNROLL' and 'PIPELINE' pragmas: Complete unroll will break the target of other loop pragmas (vector_mul.cpp:7:18)",
    "ERROR: [HLS 200-1715] Encountered problem during source synthesis"
"""

prompt = f"""You are given a piece of HLS code that contains bugs, the compiler error log, and the description of the buggy code. 
Your task is to debug the buggy code by analyzing the differences between it and the original code, using the error log to guide you.
Generate the COT debugging strategy through the buggy code, the code's description and error log.

**Instructions:**

1. **Debug the Code**:
    - **Identify Errors**: List the erroneous code line in the buggy code.
    - **Analyze Reasons**: Explain why each bug occurred, referencing the error log and the differences from the original code.
    - **Propose Fixes**: Specify the exact changes needed to fix each bug, including code modifications.

2. **Construct a Chain of Thought (CoT)**:
    - **Sequential Reasoning**: Address one error at a time, starting with syntax errors from the compiler.
    - **Hypothesis Formation**: Hypothesize why logical errors occur based on the error log and code comparison.
    - **Testing and Verification**: Validate hypotheses through analysis.
    - **Reflection**: Reflect on the findings after each step and plan the next steps.

3. **Output**:
    - **Chain of Thought (CoT)**: Write a detailed explanation of your debugging process.
    - **Bug Analysis in JSON Format**: Provide a structured summary as per the format below.
    - **No other explanations or conclusions**

**Inputs:**

Buggy Code: 
```cpp
{buggy_code}
```

Vitis HLS Compiler Error Log: 
```text
{error_log}
```

**Expected Output Format:**

Bug Analysis:
```json
{{
    "bug_analysis": [
        {{
            "erroneous_code_line": "[code line]",
            "reason": "[Explanation of why the bug occurred]",
            "modification_actions": [
                {{
                    "action": "[Specific modification to be made]",
                    "reason": "[Reason for the modification]"
                }}
            ]
        }}
    ]
}}
```
Please ensure that the JSON is properly formatted and valid.
"""

messages = [
    {"role": "system", 
     "content": """You are an expert in High-Level Synthesis (HLS) programming. 
                Your primary task is to identify and debug HLS-incompatible errors in regular C/C++ programs. You will be provided with:
                - **Buggy Code**: A C/C++ program containing HLS-incompatible errors.
                - **Compiler Error Log**: The error messages generated by the Vitis HLS compiler."""},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=2048,
    temperature=0.7
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)