app.py

import gradio as gr
from gradio_client import Client
from gradio_client.exceptions import AppError
import frontmatter
import os
import spaces
import torch
import logging
from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers.utils import logging as transformers_logging

# Set up comprehensive logging
logging.basicConfig(
    level=logging.DEBUG,
    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s'
)
logger = logging.getLogger(__name__)

import huggingface_hub

import prep_decompiled

# Model configuration constants
MAX_CONTEXT_LENGTH = 8192
MAX_NEW_TOKENS = 1024

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",
)
print("Loaded vardecoder model successfully.")

logger.info("Loading fielddecoder model...")

fielddecoder_model = None
#fielddecoder_model = AutoModelForCausalLM.from_pretrained(
#    "ejschwartz/resym-fielddecoder", 
#    torch_dtype=torch.bfloat16,
#)
#logger.info("Successfully loaded fielddecoder model")

make_gradio_client = lambda: 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):
    try:
        field_helper_result = make_gradio_client().predict(
            decompiled_code=code,
            api_name="/predict",
        )
    except AppError as e:
        print(f"AppError: {e}")
        return None, [], None
    
    print(f"field helper result: {field_helper_result}")

    fields = sorted(list(set([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"
    if len(fields) > 0:
        prompt += f"{fields[0]}:"

    print(f"field prompt: {repr(prompt)}")

    return prompt, fields, field_helper_result

@spaces.GPU
def infer(code):

    splitcode = code.splitlines()
    #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])

    first_var = 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```{first_var}:"

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

    var_input_ids = tokenizer.encode(var_prompt, return_tensors="pt").cuda()[
        :, : MAX_CONTEXT_LENGTH - MAX_NEW_TOKENS
    ]
    var_output = vardecoder_model.generate(
        input_ids=var_input_ids,
        max_new_tokens=MAX_NEW_TOKENS,
        num_beams=4,
        num_return_sequences=1,
        do_sample=False,
        early_stopping=False,
        pad_token_id=0,
        eos_token_id=0,
    )
    print(f"Pre Var output: {var_output}")
    var_output = var_output[0]
    var_output = tokenizer.decode(
        var_output[var_input_ids.size(1) :],
        skip_special_tokens=True,
        clean_up_tokenization_spaces=True,
    )

    print(f"Var output: {repr(var_output)}")

    field_prompt_result, fields, field_helper_result = field_prompt(code)
    if len(fields) == 0:
        field_output = "Failed to parse fields" if field_prompt_result is None else "No fields"
    else:
        field_input_ids = tokenizer.encode(field_prompt_result, return_tensors="pt").cuda()[
            :, : MAX_CONTEXT_LENGTH - MAX_NEW_TOKENS
        ]

        if fielddecoder_model is None:
            field_output = "TEMPORARILY DISABLED"
        else:
            field_output = fielddecoder_model.generate(
                input_ids=field_input_ids,
                max_new_tokens=MAX_NEW_TOKENS,
                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[field_input_ids.size(1) :],
                skip_special_tokens=True,
                clean_up_tokenization_spaces=True,
            )

            field_output = fields[0] + ":" + field_output
    var_output = first_var + ":" + var_output
    fieldstring = ", ".join(fields)
    return var_output, field_output, varstring, fieldstring


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"),
        gr.Text(label="Generated Field Access List"),
    ],
    # description=frontmatter.load("README.md").content,
    description="""This is a test space of the models from the [ReSym
artifacts](https://github.com/lt-asset/resym).  For more information, please see
[the
README](https://huggingface.co/spaces/ejschwartz/resym/blob/main/README.md).  If
you get an error, please make sure the [ReSym field helper
space](https://huggingface.co/spaces/ejschwartz/resym-field-helper) is
running.

The field decoder model is currently **not working** due to a HuggingFace accelerate library problem. I am investigating the issue.
""",
    examples=examples,
)
demo.launch()