app.py

import gradio as gr
from gradio_client import Client
import frontmatter
import os
import spaces
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

import huggingface_hub

import prep_decompiled

hf_key = os.environ["HF_TOKEN"]
huggingface_hub.login(token=hf_key)

tokenizer = AutoTokenizer.from_pretrained("bigcode/starcoderbase-3b")
vardecoder_model = AutoModelForCausalLM.from_pretrained(
    "ejschwartz/resym-vardecoder", torch_dtype=torch.bfloat16, device_map="auto"
)
# fielddecoder_model = AutoModelForCausalLM.from_pretrained(
#     "ejschwartz/resym-fielddecoder", torch_dtype=torch.bfloat16, device_map="auto"
# )

gradio_client = Client("https://ejschwartz-resym-field-helper.hf.space/")

examples = [
    ex.encode().decode("unicode_escape") for ex in open("examples.txt", "r").readlines()
]


# Example prompt
#   "input": "```\n_BOOL8 __fastcall sub_409B9A(_QWORD *a1, _QWORD *a2)\n{\nreturn *a1 < *a2 || *a1 == *a2 && a1[1] < a2[1];\n}\n```\nWhat are the variable name and type for the following memory accesses:a1, a1[1], a2, a2[1]?\n",
#  "output": "a1: a, os_reltime* -> sec, os_time_t\na1[1]: a, os_reltime* -> usec, os_time_t\na2: b, os_reltime* -> sec, os_time_t\na2[1]: b, os_reltime* -> usec, os_time_t",
def field_prompt(code):
    field_helper_result = gradio_client.predict(
        decompiled_code=code,
        api_name="/predict",
    )
    print(f"field helper result: {field_helper_result}")

    fields = sorted([e['expr'] for e in field_helper_result[0] if e['expr'] != ''])
    print(f"fields: {fields}")

    prompt = f"```\n{code}\n```\nWhat are the variable name and type for the following memory accesses:{', '.join(fields)}?\n"

    return prompt, field_helper_result

@spaces.GPU
def infer(code):

    splitcode = [s.strip() for s in code.splitlines()]
    code = "\n".join(splitcode)
    bodyvars = [
        v["name"] for v in prep_decompiled.extract_comments(splitcode) if "name" in v
    ]
    argvars = [
        v["name"] for v in prep_decompiled.parse_signature(splitcode) if "name" in v
    ]
    vars = argvars + bodyvars
    # comments = prep_decompiled.extract_comments(splitcode)
    # sig = prep_decompiled.parse_signature(splitcode)
    # print(f"vars {vars}")

    varstring = ", ".join([f"`{v}`" for v in vars])

    var_name = vars[0]

    # ejs: Yeah, this var_name thing is really bizarre. But look at https://github.com/lt-asset/resym/blob/main/training_src/fielddecoder_inf.py
    var_prompt = f"What are the original name and data types of variables {varstring}?\n```\n{code}\n```{var_name}"

    print(f"Prompt:\n{var_prompt}")

    input_ids = tokenizer.encode(var_prompt, return_tensors="pt").cuda()[
        :, : 8192 - 1024
    ]
    var_output = vardecoder_model.generate(
        input_ids=input_ids,
        max_new_tokens=1024,
        num_beams=4,
        num_return_sequences=1,
        do_sample=False,
        early_stopping=False,
        pad_token_id=0,
        eos_token_id=0,
    )[0]
    var_output = tokenizer.decode(
        var_output[input_ids.size(1) :],
        skip_special_tokens=True,
        clean_up_tokenization_spaces=True,
    )

    field_prompt_result, field_helper_result = field_prompt(code)

    # field_output = fielddecoder_model.generate(
    #     input_ids=input_ids,
    #     max_new_tokens=1024,
    #     num_beams=4,
    #     num_return_sequences=1,
    #     do_sample=False,
    #     early_stopping=False,
    #     pad_token_id=0,
    #     eos_token_id=0,
    # )[0]
    # field_output = tokenizer.decode(
    #     field_output[input_ids.size(1) :],
    #     skip_special_tokens=True,
    #     clean_up_tokenization_spaces=True,
    # )

    var_output = var_name + ":" + var_output
    # field_output = var_name + ":" + field_output
    return var_output, varstring


demo = gr.Interface(
    fn=infer,
    inputs=[
        gr.Textbox(lines=10, value=examples[0], label="Hex-Rays Decompilation"),
    ],
    outputs=[
        gr.Text(label="Var Decoder Output"),
        # gr.Text(label="Field Decoder Output"),
        gr.Text(label="Generated Variable List"),
    ],
    description=frontmatter.load("README.md").content,
    examples=examples,
)
demo.launch()