README.md

---
{}
---

Deployment:
```yaml
build_commands: []
external_package_dirs: []
model_metadata: {}
model_name: fp8-baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification
python_version: py39
requirements: []
resources:
  accelerator: H100:1
  cpu: "1"
  memory: 64Gi
  use_gpu: true
secrets:
  hf_access_token: set token in baseten workspace
system_packages: []
trt_llm:
  build:
    base_model: encoder
    # automatically infered from config[max_position_embeddings]
    max_seq_len: 42 
    # max_batch_size per dynamic batch, recommended to stay at 32
    max_batch_size: 32
    # max num tokens per dynamic batch, strongly recommended to keep this number
    max_num_tokens: 16384 
    checkpoint_repository:
      source: HF
      repo: "baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification"
      revision: "main" # hf revision hash
    # `fp8` or `no_quant` (=fp16) are allowed.
    quantization_type: fp8
    num_builder_gpus: 4
```

Usage:
```python
import requests
import os
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Skywork/Skywork-Reward-Llama-3.1-8B-v0.2")

prompt = "Jane has 12 apples. She gives 4 apples to her friend Mark, then buys 1 more apple, and finally splits all her apples equally among herself and her 2 siblings. How many apples does each person get?"
# Positive example, gets high score 0.999 or raw around inv_sig(0.999) ~ 13
response1 = "1. Jane starts with 12 apples and gives 4 to Mark. 12 - 4 = 8. Jane now has 8 apples.\n2. Jane buys 1 more apple. 8 + 1 = 9. Jane now has 9 apples.\n3. Jane splits the 9 apples equally among herself and her 2 siblings (3 people in total). 9 ÷ 3 = 3 apples each. Each person gets 3 apples."
# negative example, gets low score ~0.001 or raw around inv_sig(0.001) ~ -9
response2 = "1. Jane starts with 12 apples and gives 4 to Mark. 12 - 4 = 8. Jane now has 8 apples.\n2. Jane buys 1 more apple. 8 + 1 = 9. Jane now has 9 apples.\n3. Jane splits the 9 apples equally among her 2 siblings (2 people in total). 9 ÷ 2 = 4.5 apples each. Each person gets 4 apples."

# predict api: {
#   "inputs": "What is Deep Learning?", # str, may be formatted with chat template.
#   "raw_scores": false, # with or without sigmoid activation
#   "truncate": false,
#   "truncation_direction": "right"
# }

for assistant_response in [response1, response2]:
    # Feel free to parallelize this, requests will be batched in the backend.
    
    conv = [{"role": "user", "content": prompt}, {"role": "assistant", "content": assistant_response}]
    conv_formatted = tokenizer.apply_chat_template(conv, tokenize=False)
    input_json = dict(inputs=conv_formatted, raw_scores=True)
    resp = requests.post(
        "https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
        headers={"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"},
        json=input_json,
    )

    print(resp.json())
    # prints
    # [{'score': 13.714337, 'label': 'LABEL_0'}]
    # [{'score': -9.353895, 'label': 'LABEL_0'}]
```

Reproduce this model:
```python
#!/usr/bin/env python
import torch
from transformers import (
    AutoConfig,
    AutoTokenizer,
    AutoModelForCausalLM,
    LlamaForSequenceClassification,
)
# install torch, transformers, accelerate

def main():
    # Define the input and output repository names.
    input_model_id = "meta-llama/Meta-Llama-3-70B-Instruct"
    split_2 = input_model_id.split("/")[1]
    output_model_id = f"baseten/example-{split_2}ForSequenceClassification"  

    # Load the original configuration.
    # (If needed, add trust_remote_code=True for custom implementations.)
    config = AutoConfig.from_pretrained(input_model_id)
    
    # Update the config for a sequence classification task with 10 labels.
    num_labels = 30
    config.num_labels = num_labels
    config.id2label = {i: f"token activation {i}" for i in range(num_labels)}
    config.label2id = {f"token activation {i}": i for i in range(num_labels)}

    # Download the tokenizer from the original model.
    tokenizer = AutoTokenizer.from_pretrained(input_model_id)

    # Load the original causal LM model.
    lm_model = AutoModelForCausalLM.from_pretrained(input_model_id, config=config, device_map="auto", low_cpu_mem_usage=True)
    config.architectures = ["LlamaForSequenceClassification"]
    del lm_model.model
    print("loaded lm model")
    # Initialize the sequence classification model.
    # NOTE: We are using the built-in LlamaForSequenceClassification,
    # which uses a `.score` attribute as the output head.
    seq_cls_model = LlamaForSequenceClassification.from_pretrained(input_model_id, config=config, device_map="auto", low_cpu_mem_usage=True)

    # --- Initialize the Classification Head ---
    # Here we re-use the first 10 rows from the original LM head
    # (i.e. rows 0 to 9) to initialize the new classification head.
    with torch.no_grad():
        # lm_model.lm_head.weight has shape [vocab_size, hidden_size]
        # We take the first 10 rows to form a [10, hidden_size] weight matrix.
        seq_cls_model.score.weight.copy_(lm_model.lm_head.weight.data[:num_labels, :])
        if lm_model.lm_head.bias is not None:
            seq_cls_model.score.bias.copy_(lm_model.lm_head.bias.data[:num_labels])

    # Optionally, save the new model locally.
    # save_directory = f"./{output_model_id.replace('/','_')}"
    # seq_cls_model.save_pretrained(save_directory)
    # tokenizer.save_pretrained(save_directory)

    # Push the new model and tokenizer to the Hub.
    # (Ensure you are authenticated with Hugging Face Hub via `huggingface-cli login`.)
    tokenizer.push_to_hub(output_model_id)
    seq_cls_model.push_to_hub(output_model_id)
    

    print(f"New model pushed to the Hub: {output_model_id}")

if __name__ == "__main__":
    main()
```