## How to use TriStageHLA-BIND
```python
import torch
import torch.nn.functional as F
from transformers import AutoModel, AutoTokenizer

# Device
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load model (replace with your model id if different)
model_id = "SkywalkerLu/TriStageHLA-BIND"
model = AutoModel.from_pretrained(model_id, trust_remote_code=True).to(device).eval()

# Load tokenizer used in training (ESM2 650M)
tok = AutoTokenizer.from_pretrained("facebook/esm2_t33_650M_UR50D")

# Example inputs
peptide = "GILGFVFTL"  # 9-mer example
# Fake placeholder pseudosequence for demo; replace with a real one from your mapping/data
hla_pseudoseq = (
    "YYSEYRNIYAQTDESNLYLSYDYYTWAERAYEWY"
)

# Fixed lengths (must match training)
PEP_LEN = 16
HLA_LEN = 36
PAD_ID = tok.pad_token_id if tok.pad_token_id is not None else 1

def pad_to_len(ids_list, target_len, pad_id):
    return ids_list + [pad_id] * (target_len - len(ids_list)) if len(ids_list) < target_len else ids_list[:target_len]

# Tokenize
pep_ids = tok(peptide, add_special_tokens=True)["input_ids"]
hla_ids = tok(hla_pseudoseq, add_special_tokens=True)["input_ids"]

# Pad/truncate
pep_ids = pad_to_len(pep_ids, PEP_LEN, PAD_ID)
hla_ids = pad_to_len(hla_ids, HLA_LEN, PAD_ID)

# Tensors (batch=1)
pep_tensor = torch.tensor([pep_ids], dtype=torch.long, device=device)
hla_tensor = torch.tensor([hla_ids], dtype=torch.long, device=device)

# Forward + probability
with torch.no_grad():
    logits, features = model(pep_tensor, hla_tensor)
    prob_bind = F.softmax(logits, dim=1)[0, 1].item()
    pred = int(prob_bind >= 0.5)

print({"peptide": peptide, "bind_prob": round(prob_bind, 6), "label": pred})