app.py

import gradio as gr
import json
import os
import spaces
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

import huggingface_hub

import prep_decompiled

description = """# ReSym Test Space

This is a test space of the models from the [ReSym
artifacts](https://github.com/lt-asset/resym).  Sadly, at the time I am writing
this, not all of ReSym is publicly available; specifically, the Prolog component
is [not available](https://github.com/lt-asset/resym/issues/2).

This space simply performs inference on the two pretrained models available as
part of the ReSym artifacts. It takes a variable name and some decompiled code
as input, and outputs the variable type and other information.

## Disclaimer

I'm not a ReSym developer and I may have messed something up.  In particular,
you must prompt the variable names in the decompiled code as part of the prompt,
and I reused some of their own code to do this.

## Todo

* Add support for FieldDecoder model

"""

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"
)

example = r"""{
  "input": "What are the original name and data type of variables `a1`, `a2`, `a3`, `v4`, `v5`?\n```\n__int64 __fastcall sub_410D81(__int64 a1, __int64 a2, __int64 a3)\n{\nint v4; // [rsp+20h] [rbp-20h] BYREF\n__int64 v5; // [rsp+28h] [rbp-18h]\n\nif ( !a1 || !a2 || !a3 )\nreturn 0LL;\nv4 = 5;\nv5 = a3;\nreturn sub_411142(a1, a2, &v4);\n}\n```",
  "output": "a1: dict, pdfio_dict_t*\na2: key, const char*\na3: value, pdfio_dict_t*\nv4: temp, struct _pdfio_value_t\nv5: -, -",
  "funname": "pdfioDictSetDict",
  "bin": "d0ebaa77558783765df381826c961e95c83338e285011d501a62bea474e93451",
  "proj": "michaelrsweet/pdfio",
  "cluster_var": {
    "struct": [
      [
        "v4",
        "v5"
      ]
    ]
  }
}"""


@spaces.GPU
def infer(var_name, code):

    splitcode = code.splitlines()
    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}")

    #line = json.loads(input)
    #first_token = line["output"].split(":")[0]
    prompt = code + var_name + ":"

    input_ids = tokenizer.encode(prompt, return_tensors="pt").cuda()[:, : 8192 - 1024]
    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]
    output = tokenizer.decode(
        output[input_ids.size(1) :],
        skip_special_tokens=True,
        clean_up_tokenization_spaces=True,
    )

    output = var_name + ":" + output
    return output


demo = gr.Interface(
    fn=infer,
    inputs=[
        gr.Text(label="First Token", value="a1"),
        gr.Textbox(lines=10, value=json.loads(example)['input']),
    ],
    outputs=gr.Text(label="Var Decoder Output"),
    description=description
)
demo.launch()