app.py

import os
import re

os.environ["TOKENIZERS_PARALLELISM"] = "true"

import frontmatter
import gradio as gr
import json
import spaces
import torch
from normalize import normalize

from transformers import AutoTokenizer
from modeling_nova import NovaTokenizer, NovaForCausalLM


def fix_assembly_tabs(asm_text):
    """
    Fix assembly code formatting by ensuring proper tab placement.
    Expected format: address:TABhex_bytesWHITESPACEinstructionWHITESPACEoperands
    """
    lines = asm_text.split("\n")
    fixed_lines = []

    for line in lines:
        line = line.rstrip()  # Remove trailing whitespace
        if not line.strip():  # Skip empty lines
            fixed_lines.append(line)
            continue

        # Check if this looks like an assembly instruction line
        # Pattern: optional_spaces + hex_address + colon + hex_bytes + instruction + operands
        asm_pattern = r"^(\s*)([0-9a-f]+):\s*([0-9a-f\s]+?)\s+(\w+)(\s+.*)?$"
        match = re.match(asm_pattern, line, re.IGNORECASE)

        if match:
            indent, address, hex_bytes, instruction, operands = match.groups()
            operands = operands or ""

            # Clean up hex bytes (remove extra spaces)
            hex_bytes = re.sub(r"\s+", " ", hex_bytes.strip())

            # Reconstruct with proper tab formatting
            # Format: indent + address + ":" + TAB + hex_bytes + TAB + instruction + operands
            fixed_line = f"{indent}{address}:\t{hex_bytes}\t{instruction}{operands}"
            fixed_lines.append(fixed_line)
        else:
            # Not an assembly instruction line, keep as is
            fixed_lines.append(line)

    return "\n".join(fixed_lines)


print("Downloading model")

tokenizer = AutoTokenizer.from_pretrained(
    "lt-asset/nova-6.7b-bcr", trust_remote_code=True
)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.pad_token_id = tokenizer.eos_token_id
nova_tokenizer = NovaTokenizer(tokenizer)

model = NovaForCausalLM.from_pretrained(
    "lt-asset/nova-6.7b-bcr", torch_dtype=torch.bfloat16, device_map="auto"
).eval()

examples = json.load(open("humaneval_decompile_nova_6.7b.json", "r"))


@spaces.GPU
def predict(type, input_asm, _c_source):

    if "<func0>:" not in input_asm:
        # Needs normalizing

        # Add a bogus function header if needed.
        first_line = input_asm.split("\n")[0]
        if "<" not in first_line or ">" not in first_line:
            print("Adding synthetic function header")
            input_asm = "<func0>:\n" + input_asm

        # Fix tab formatting in assembly code
        input_asm = fix_assembly_tabs(input_asm)

        # Normalizing
        normalized_asm = normalize(input_asm)
        print(f"Normalized asm: {normalized_asm}")
    else:
        normalized_asm = input_asm

    prompt_before = f"# This is the assembly code with {type} optimization:\n<func0>:"
    asm = normalized_asm.strip()
    assert asm.startswith("<func0>:")
    asm = asm[len("<func0>:") :]
    prompt_after = "\nWhat is the source code?\n"

    inputs = prompt_before + asm + prompt_after
    print("Inputs:", inputs)

    # 0 for non-assembly code characters and 1 for assembly characters, required by nova tokenizer
    char_types = "0" * len(prompt_before) + "1" * len(asm) + "0" * len(prompt_after)

    tokenizer_output = nova_tokenizer.encode(inputs, "", char_types)
    input_ids = torch.LongTensor(tokenizer_output["input_ids"].tolist()).unsqueeze(0)
    print("Input IDs:", input_ids.shape)
    nova_attention_mask = torch.LongTensor(
        tokenizer_output["nova_attention_mask"]
    ).unsqueeze(0)

    output = model.generate(
        inputs=input_ids.cuda(),
        max_new_tokens=512,
        temperature=0.2,
        top_p=0.95,
        num_return_sequences=1,
        do_sample=True,
        nova_attention_mask=nova_attention_mask.cuda(),
        no_mask_idx=torch.LongTensor([tokenizer_output["no_mask_idx"]]).cuda(),
        pad_token_id=tokenizer.pad_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )
    print("Output 1:", output)

    output = tokenizer.decode(
        output[0][input_ids.size(1) :],
        skip_special_tokens=True,
        clean_up_tokenization_spaces=True,
    )
    print("Output 2:", output)

    return output

example = """   0:	f3 0f 1e fa          	endbr64 
   4:	55                   	push   %rbp
   5:	48 89 e5             	mov    %rsp,%rbp
   8:	89 7d fc             	mov    %edi,-0x4(%rbp)
   b:	8b 45 fc             	mov    -0x4(%rbp),%eax
   e:	83 c0 2a             	add    $0x2a,%eax
  11:	5d                   	pop    %rbp
  12:	c3                   	ret    
"""

demo = gr.Interface(
    fn=predict,
    inputs=[
        gr.Text(label="Optimization Type", value="O0"),
        gr.Text(label="Assembly Code (Normalized or not)", value=example),
        gr.Text(label="Original C Code"),
    ],
    outputs=gr.Text(label="Raw Nova Output"),
    description=frontmatter.load("README.md").content,
    examples=[[ex["type"], ex["normalized_asm"], ex["c_func"]] for ex in examples],
)
demo.launch(show_error=True)