lib/graph_extract.py

import json
import re
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
import torch
import warnings
import spaces

flash_attn_installed = False
try:
    import subprocess
    print("Installing flash-attn...")
    subprocess.run(
    "pip install flash-attn --no-build-isolation",
    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
    shell=True,
    )
    flash_attn_installed = True
except Exception as e:
    print(f"Error installing flash-attn: {e}")


# Suppress specific warnings
warnings.filterwarnings(
    "ignore",
    message="You have modified the pretrained model configuration to control generation.",
)
warnings.filterwarnings(
    "ignore",
    message="You are not running the flash-attention implementation, expect numerical differences.",
)

print("Initializing application...")

model = AutoModelForCausalLM.from_pretrained(
    "sciphi/triplex", 
    trust_remote_code=True,
    attn_implementation="flash_attention_2" if flash_attn_installed else None,
    torch_dtype=torch.bfloat16,
    device_map="auto", 
    low_cpu_mem_usage=True,#advised if any device map given
).eval()

tokenizer = AutoTokenizer.from_pretrained(
    "sciphi/triplex", 
    trust_remote_code=True,
    attn_implementation="flash_attention_2" if flash_attn_installed else None,
        torch_dtype=torch.bfloat16,
    )


print("Model and tokenizer loaded successfully.")

# Set up generation config
generation_config = GenerationConfig.from_pretrained("sciphi/triplex")
generation_config.max_length = 2048
generation_config.pad_token_id = tokenizer.eos_token_id


@spaces.GPU
def triplextract(text, entity_types, predicates):
    input_format = """Perform Named Entity Recognition (NER) and extract knowledge graph triplets from the text. NER identifies named entities of given entity types, and triple extraction identifies relationships between entities using specified predicates. Return the result as a JSON object with an "entities_and_triples" key containing an array of entities and triples.
        
        **Entity Types:**
        {entity_types}
        
        **Predicates:**
        {predicates}
        
        **Text:**
        {text}
        """
    message = input_format.format(
                entity_types = json.dumps({"entity_types": entity_types}),
                predicates = json.dumps({"predicates": predicates}),
                text = text)
    
    # message = input_format.format(
    #     entity_types=entity_types, predicates=predicates, text=text
    # )

    messages = [{"role": "user", "content": message}]

    print("Tokenizing input...")
    input_ids = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True, return_tensors="pt"
    ).to(model.device)

    attention_mask = input_ids.ne(tokenizer.pad_token_id)

    print("Generating output...")
    try:
        with torch.no_grad():
            output = model.generate(
                input_ids=input_ids,
                attention_mask=attention_mask,
                generation_config=generation_config,
                do_sample=True,
            )

        decoded_output = tokenizer.decode(output[0], skip_special_tokens=True)
        print("Decoding output completed.")

        return decoded_output
    except torch.cuda.OutOfMemoryError as e:
        print(f"CUDA out of memory error: {e}")
        return "Error: CUDA out of memory."
    except Exception as e:
        print(f"Error in generation: {e}")
        return f"Error in generation, please try again: {str(e)}"

def parse_triples(prediction):
    entities = {}
    relationships = []

    try:
        data = json.loads(prediction)
        items = data.get("entities_and_triples", [])
    except json.JSONDecodeError:
        json_match = re.search(r"```json\s*(.*?)\s*```", prediction, re.DOTALL)
        if json_match:
            try:
                data = json.loads(json_match.group(1))
                items = data.get("entities_and_triples", [])
            except json.JSONDecodeError:
                items = re.findall(r"\[(.*?)\]", prediction)
        else:
            items = re.findall(r"\[(.*?)\]", prediction)

    for item in items:
        if isinstance(item, str):
            try:
                if ":" in item:
                    id, entity = item.split(",", 1)
                    id = id.strip("[]").strip()
                    entity_type, entity_value = entity.split(":", 1)
                    entities[id] = {
                        "type": entity_type.strip(),
                        "value": entity_value.strip(),
                    }
                else:
                    parts = item.split()
                    if len(parts) >= 3:
                        source = parts[0].strip("[]")
                        relation = " ".join(parts[1:-1])
                        target = parts[-1].strip("[]")
                        relationships.append((source, relation.strip(), target))
            except Exception as e:
                # TODO: Handle gracefully
                print(f"Error in processing: {item}: {e}")
                
    return entities, relationships