data2/step22/emb_qwen_func.py

# Requires vllm>=0.8.5
import torch
import vllm
from vllm import LLM
from transformers import AutoTokenizer
from pathlib import Path
import os
import jsonlines

os.environ["CUDA_VISIBLE_DEVICES"] = "1"

# 必须在设置 CUDA_VISIBLE_DEVICES 之后
import multiprocessing as mp
mp.set_start_method("spawn", force=True)


def get_detailed_instruct(task_description: str, query: str) -> str:
    return f'Instruct: {task_description}\nQuery:{query}'


keywords = ["Quantum mechanics",
    "Gene editing",
    "Folding",
    "System biology",
    "Antibody",
    "Heterogeneity",
    "Ligand",
    "Drug repurpose",
    "Kinetics",
    "Next-generation sequencing",
    "Pharmacogenetics",
    "Phase-field technique",
    "Human",
    "Potential",
    "Hartree-Fock",
    "Flow matching",
    "Lipid",
    "Biomedical",
    "Antigen",
    "Stochastic modeling",
    "Coupled cluster",
    "Quantum biology",
    "Spatial biology",
    "Antagonist",
    "Free energy perturbation",
    "Cycle",
    "Pharmacology",
    "Redox",
    "Physiology",
    "Protein-Protein Interactions",
    "Single-cell",
    "Screening",
    "Hydrophobic",
    "First-principles based DFT",
    "Molecular biology",
    "Mechanism",
    "Reproduction number",
    "Spatial Transcriptomics",
    "Ion",
    "Computational Materials",
    "Absorption",
    "Pharmacometrics",
    "GAN",
    "Compartmental model",
    "Diagnostics",
    "Lead discovery",
    "QAPR",
    "Rosettafold",
    "Autoregressive",
    "Pharmacokinetics",
    "Biotechnology",
    "Hydrophilic",
    "3D",
    "Protein",
    "QM/MM",
    "Activation",
    "AMR",
    "Networks",
    "Genotype",
    "Gene regulatory networks",
    "Biologics",
    "Phenotype",
    "Nowcasting",
    "DFT",
    "AlphaFold",
    "Pandemic",
    "Immunology",
    "Pathology",
    "Chemical space",
    "Transformer",
    "Homeostasis",
    "Score",
    "High-throughput",
    "Cheminformatics",
    "Hit-to-lead",
    "Sequencing",
    "Enzyme",
    "Antimicrobial resistance modeling",
    "Allosteric",
    "Inhibition",
    "Computational Biochemistry",
    "Bioinformatics",
    "Transcriptomics",
    "Diffusion",
    "Anomaly detection",
    "Multi-omics",
    "Biology",
    "Pathway",
    "Metabolomics",
    "Synthetic biology",
    "Microbial",
    "Proteomics",
    "Pharmaceutics",
    "Organoid",
    "Network pharmacology",
    "Imaging",
    "Generative adversarial networks",
    "Microbiology",
    "Organ-on-a-chip",
    "De novo",
    "Substrate",
    "Personalized",
    "Drug",
    "Transcription",
    "RNA",
    "Explainable AI",
    "Generate",
    "Docking",
    "Pathogens",
    "Bio foundation model",
    "Reinforcement learning",
    "Mechanism of action",
    "Generative",
    "Metabolic",
    "Metabolic Flux Analysis",
    "Computational Chemistry",
    "Vaccine",
    "Biophysics",
    "Integration",
    "Biochemistry",
    "Physiologically based pharmacokinetics model",
    "Medicine",
    "Crystal",
    "Conjugate",
    "Variational autoencoders",
    "In Silico",
    "Protein-protein",
    "CRISPR",
    "Spatial transcriptomics",
    "Gene",
    "Translation",
    "Glycomics",
    "Lead optimization",
    "Pharmacodynamics",
    "Ab initio",
    "System immunology",
    "Pseudotime analysis",
    "Generative AI",
    "RNN",
    "Regulatory networks",
    "PBPK model",
    "Beta-blocker",
    "Lipidomics",
    "Reaction",
    "Bio",
    "Genesis",
    "Evolution",
    "Computational Biology",
    "VAE",
    "Pharmacogenomics",
    "Assay",
    "Sensors",
    "Conformation",
    "Finite element method",
    "Human atlas",
    "Translational medicine",
    "Neurology",
    "Genomics",
    "Cell biology",
    "Porous",
    "Biomarker",
    "Bioengineering",
    "Allele",
    "Recurrent neural networks",
    "Carbohydrate",
    "Metamaterial",
    "Virtual human",
    "DNA",
    "Omics",
    "Agonist",
    "Receptor",
    "Cofactor",
    "Metabolic flux analysis",
    "Cell atlas",
    "Signaling",
    "Electronic structure",
    "Monte Carlo",
    "Genomic surveillance",
    "Agent-based model",
    "Biosensors",
    "2D",
    "QSAR",
    "Codon",
    "Coenzyme",
    "Nucleic acids",
    "Dynamics",
    "Ensemble",
    "Spectrometry",
    "Multi-scale modeling",
    "ADMET",
    "Marker",
    "Toxicology",
    "Profiling",
    "Design",
    "Viral",
    "Chemistry",
    "Epigenetics",
    "Homo-Lumo",
    "Modeling",
    "Prediction",
    "Quantum Chemistry",
    "Half-life",
    "Material",
    "Disease",
    "Phylodynamic model",
    "Metagenomics",
    "Digital twin",
    "Cancer biology",
    "Discovery",
    "Bioavailability",
    "Digital PCR"
    ]

# Each query must come with a one-sentence instruction that describes the task
task = 'Given a web search query, retrieve relevant passages that answer the query'

queries = [
    get_detailed_instruct(task, ' '.join(keywords))
]

model = LLM(model="Qwen/Qwen3-Embedding-0.6B",
            task="embed",
            tensor_parallel_size=1,
            data_parallel_size=1)


def get_functions_contents(dir):
    subdirs = sorted([d for d in os.listdir(dir) if os.path.isdir(os.path.join(dir, d))])
    for subdir in subdirs:
        if subdir == 'ElectronicStructureLibrary___libxc':
            continue
        print(subdir)
        json_path = os.path.join(dir, subdir, 'functions.jsonl')
        contents = []
        if os.path.exists(json_path):
            objs = []
            has_scored = False
            has_read = False
            with jsonlines.open(json_path) as reader:
                has_read = True
                for obj in reader:
                    if 'score' in obj:
                        has_scored = True
                        break
                    file_path = obj['file']
                    with open(file_path, 'r', encoding='utf-8', errors='ignore') as f:
                        func_content = ''.join(f.readlines()[obj['start_line']-1:obj['end_line']])[:32000]
                        # obj['content'] = func_content
                        contents.append(func_content)
                        objs.append(obj)
                if has_read and not has_scored:
                    scores = get_scores(contents)
                    for i, obj in enumerate(objs):
                        obj['score'] = scores[i]
            if has_read and not has_scored:
                with jsonlines.open(json_path, 'w', flush=True) as writer:
                    writer.write_all(objs)
                    print("finish ", subdir)
            

tokenizer = AutoTokenizer.from_pretrained(
    "Qwen/Qwen3-Embedding-0.6B",
    trust_remote_code=True
)

MAX_TOKENS = 30000   # 留点 buffer
  
def truncate_to_max_tokens(text, max_tokens=MAX_TOKENS):
    tokens = tokenizer(
        text,
        truncation=True,
        max_length=max_tokens,
        return_tensors=None
    )
    return tokenizer.decode(tokens["input_ids"], skip_special_tokens=True)

def get_scores(documents):
    safe_queries = [truncate_to_max_tokens(q) for q in queries]
    safe_docs = [truncate_to_max_tokens(d) for d in documents]

    input_texts = safe_queries + safe_docs
    outputs = model.embed(input_texts)
    embeddings = torch.tensor([o.outputs.embedding for o in outputs])
    scores = (embeddings[0] @ embeddings[1:].T)
    return scores.tolist()

get_functions_contents('/home/weifengsun/tangou1/step2/step22/dataset')