LaSER-Qwen3-8B

import torch
import torch.nn.functional as F
from transformers import AutoModelForCausalLM, AutoTokenizer


def laser_encode(model, tokenizer, texts, max_length=512, num_thinking_steps=3):
    """Encode texts using LaSER's latent thinking mechanism."""
    device = next(model.parameters()).device
    batch = tokenizer(texts, padding=True, truncation=True, max_length=max_length, return_tensors="pt").to(device)
    input_ids, attention_mask = batch["input_ids"], batch["attention_mask"]

    batch_size = input_ids.size(0)
    thinking_slots = num_thinking_steps - 1
    eos_id = tokenizer.eos_token_id

    if thinking_slots > 0:
        eos_padding = torch.full((batch_size, thinking_slots), eos_id, dtype=input_ids.dtype, device=device)
        mask_padding = torch.ones((batch_size, thinking_slots), dtype=attention_mask.dtype, device=device)
        input_ids = torch.cat([input_ids, eos_padding], dim=1)
        attention_mask = torch.cat([attention_mask, mask_padding], dim=1)

    input_embeds = model.get_input_embeddings()(input_ids)
    embedding_table = model.get_input_embeddings().weight
    base_seq_len = input_embeds.size(1) - thinking_slots

    past_key_values = None
    hidden_steps = []

    for step_idx in range(thinking_slots):
        pos = base_seq_len + step_idx
        step_embeds = input_embeds[:, :pos, :] if past_key_values is None else input_embeds[:, pos-1:pos, :]
        step_mask = attention_mask[:, :pos]

        outputs = model(inputs_embeds=step_embeds, attention_mask=step_mask,
                       output_hidden_states=True, past_key_values=past_key_values,
                       use_cache=True, return_dict=True)
        hidden_steps.append(outputs.hidden_states[-1][:, -1, :])
        token_probs = torch.softmax(outputs.logits[:, -1, :], dim=-1)
        new_embed = token_probs @ embedding_table
        past_key_values = outputs.past_key_values
        pre = input_embeds[:, :pos, :]
        post = input_embeds[:, pos+1:, :]
        input_embeds = torch.cat([pre, new_embed.unsqueeze(1), post], dim=1)

    final_embeds = input_embeds[:, -1:, :] if past_key_values else input_embeds
    outputs = model(inputs_embeds=final_embeds, attention_mask=attention_mask,
                   output_hidden_states=True, past_key_values=past_key_values,
                   use_cache=True, return_dict=True)
    hidden_steps.append(outputs.hidden_states[-1][:, -1, :])

    embeddings = torch.stack(hidden_steps, dim=1).mean(dim=1)
    return F.normalize(embeddings, p=2, dim=-1)


# Load model
model_name = "Alibaba-NLP/LaSER-Qwen3-8B"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.padding_side = "left"
if tokenizer.pad_token_id is None:
    tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForCausalLM.from_pretrained(
    model_name, torch_dtype=torch.float16, trust_remote_code=True
).cuda().eval()

# Encode queries and documents
with torch.inference_mode():
    query_emb = laser_encode(model, tokenizer, ["why is the sky blue"], num_thinking_steps=3)
    doc_emb = laser_encode(model, tokenizer, ["Rayleigh scattering makes short wavelengths scatter more strongly"], num_thinking_steps=3)

# Compute similarity
similarity = (query_emb @ doc_emb.T).item()
print(f"Cosine similarity: {similarity:.4f}")

queries = [
    "What causes tides in the ocean?",
    "How does photosynthesis convert light to energy?",
    "Why do metals conduct electricity?",
]

with torch.inference_mode():
    query_embeddings = laser_encode(model, tokenizer, queries, num_thinking_steps=3)
    print(f"Batch embeddings shape: {query_embeddings.shape}")  # (3, 4096)