generate-responses.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "accelerate",
#     "datasets",
#     "huggingface-hub",
#     "hf-xet",
#     "torch",
#     "tqdm",
#     "transformers>=5.1",
# ]
# ///
"""
Generate responses for prompts in a dataset using transformers continuous batching.

Uses transformers' native continuous batching (CB) for efficient GPU inference.
No vLLM dependency required - works with any model supported by transformers,
including newly released architectures.

Example usage:
    # Local execution
    uv run generate-responses.py \\
        username/input-dataset \\
        username/output-dataset \\
        --prompt-column question

    # With custom model and sampling parameters
    uv run generate-responses.py \\
        username/input-dataset \\
        username/output-dataset \\
        --model-id meta-llama/Llama-3.1-8B-Instruct \\
        --messages-column messages \\
        --temperature 0.9 \\
        --max-tokens 2048

    # HF Jobs execution (see script output for full command)
    hf jobs uv run --flavor l4x1 ...
"""

import argparse
import logging
import os
import sys
from datetime import datetime
from typing import Optional

import torch
from datasets import load_dataset
from huggingface_hub import DatasetCard, get_token, login
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers.generation import GenerationConfig

# Enable HF Transfer for faster downloads
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"

logging.basicConfig(
    level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s"
)
logger = logging.getLogger(__name__)


def check_gpu_availability() -> int:
    """Check if CUDA is available and return the number of GPUs."""
    if not torch.cuda.is_available():
        logger.error("CUDA is not available. This script requires a GPU.")
        logger.error(
            "Please run on a machine with NVIDIA GPU or use HF Jobs with GPU flavor."
        )
        sys.exit(1)

    num_gpus = torch.cuda.device_count()
    for i in range(num_gpus):
        gpu_name = torch.cuda.get_device_name(i)
        gpu_memory = torch.cuda.get_device_properties(i).total_memory / 1024**3
        logger.info(f"GPU {i}: {gpu_name} with {gpu_memory:.1f} GB memory")

    return num_gpus


def create_dataset_card(
    source_dataset: str,
    model_id: str,
    messages_column: str,
    prompt_column: Optional[str],
    generation_config: GenerationConfig,
    num_gpus: int,
    num_examples: int,
    generation_time: str,
    num_skipped: int = 0,
    attn_implementation: str = "paged|sdpa",
) -> str:
    """Create a dataset card documenting the generation process."""
    filtering_section = ""
    if num_skipped > 0:
        skip_percentage = (num_skipped / num_examples) * 100
        processed = num_examples - num_skipped
        filtering_section = f"""

### Filtering Statistics

- **Total Examples**: {num_examples:,}
- **Processed**: {processed:,} ({100 - skip_percentage:.1f}%)
- **Skipped (too long)**: {num_skipped:,} ({skip_percentage:.1f}%)

Note: Prompts exceeding the model's maximum context length were skipped and have empty responses."""

    input_col = prompt_column if prompt_column else messages_column
    input_type = "plain text prompts" if prompt_column else "chat messages"

    return f"""---
tags:
- generated
- transformers
- continuous-batching
- uv-script
---

# Generated Responses Dataset

This dataset contains generated responses for prompts from [{source_dataset}](https://huggingface.co/datasets/{source_dataset}).

## Generation Details

- **Source Dataset**: [{source_dataset}](https://huggingface.co/datasets/{source_dataset})
- **Input Column**: `{input_col}` ({input_type})
- **Model**: [{model_id}](https://huggingface.co/{model_id})
- **Backend**: transformers continuous batching
- **Number of Examples**: {num_examples:,}
- **Generation Date**: {generation_time}{filtering_section}

### Generation Parameters

- **Temperature**: {generation_config.temperature}
- **Top P**: {generation_config.top_p}
- **Top K**: {generation_config.top_k}
- **Max New Tokens**: {generation_config.max_new_tokens}
- **Max Batch Tokens**: {generation_config.max_batch_tokens}
- **Repetition Penalty**: {generation_config.repetition_penalty}

### Hardware Configuration

- **GPUs**: {num_gpus}
- **Attention Implementation**: {attn_implementation}

## Dataset Structure

The dataset contains all columns from the source dataset plus:
- `response`: The generated response from the model

## Generation Script

Generated using the transformers continuous batching script from [uv-scripts/transformers-inference](https://huggingface.co/datasets/uv-scripts/transformers-inference).

To reproduce this generation:

```bash
uv run https://huggingface.co/datasets/uv-scripts/transformers-inference/raw/main/generate-responses.py \\
    {source_dataset} \\
    <output-dataset> \\
    --model-id {model_id} \\
    {"--prompt-column " + prompt_column if prompt_column else "--messages-column " + messages_column} \\
    --temperature {generation_config.temperature} \\
    --top-p {generation_config.top_p} \\
    --top-k {generation_config.top_k} \\
    --max-tokens {generation_config.max_new_tokens}
```
"""


def main(
    src_dataset_hub_id: str,
    output_dataset_hub_id: str,
    model_id: str = "Qwen/Qwen3-4B-Instruct-2507",
    messages_column: str = "messages",
    prompt_column: Optional[str] = None,
    output_column: str = "response",
    temperature: float = 0.7,
    top_p: float = 0.8,
    top_k: int = 20,
    max_tokens: int = 4096,
    repetition_penalty: float = 1.0,
    max_batch_tokens: int = 512,
    dtype: str = "bfloat16",
    attn_implementation: str = "paged|sdpa",
    subset: Optional[str] = None,
    split: str = "train",
    messages_until_role: Optional[str] = None,
    skip_long_prompts: bool = True,
    max_samples: Optional[int] = None,
    hf_token: Optional[str] = None,
):
    generation_start_time = datetime.now().isoformat()

    # GPU check
    num_gpus = check_gpu_availability()

    # Authentication
    HF_TOKEN = hf_token or os.environ.get("HF_TOKEN") or get_token()
    if not HF_TOKEN:
        logger.error("No HuggingFace token found. Please provide token via:")
        logger.error("  1. --hf-token argument")
        logger.error("  2. HF_TOKEN environment variable")
        logger.error("  3. Run 'huggingface-cli login' or use login() in Python")
        sys.exit(1)

    logger.info("HuggingFace token found, authenticating...")
    login(token=HF_TOKEN)

    # Resolve dtype
    torch_dtype = getattr(torch, dtype, None)
    if torch_dtype is None:
        logger.error(f"Unknown dtype: {dtype}. Use 'bfloat16', 'float16', or 'float32'.")
        sys.exit(1)

    # Load model with continuous batching support
    # Note: CB currently requires single-GPU. Multi-GPU device_map="auto" causes
    # device mismatch errors with PagedAttention cache. Use a model that fits on one GPU.
    if num_gpus > 1:
        logger.warning(
            "Multiple GPUs detected but transformers CB currently works on single GPU only. "
            "Using cuda:0. Choose a model that fits on one GPU."
        )
    device_map = "cuda"
    logger.info(f"Loading model: {model_id} (dtype={dtype}, attn={attn_implementation}, device_map={device_map})")

    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        attn_implementation=attn_implementation,
        device_map=device_map,
        dtype=torch_dtype,
    )

    # Load tokenizer
    logger.info("Loading tokenizer...")
    tokenizer = AutoTokenizer.from_pretrained(model_id, padding_side="left")

    # Create generation config
    do_sample = temperature != 1.0 or top_p != 1.0 or top_k != 0
    generation_config = GenerationConfig(
        max_new_tokens=max_tokens,
        max_batch_tokens=max_batch_tokens,
        do_sample=do_sample,
        temperature=temperature,
        top_p=top_p,
        top_k=top_k,
        repetition_penalty=repetition_penalty,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.pad_token_id,
    )

    # Load dataset
    logger.info(f"Loading dataset: {src_dataset_hub_id}" + (f" ({subset})" if subset else ""))
    dataset = load_dataset(src_dataset_hub_id, subset, split=split)

    if max_samples is not None and max_samples < len(dataset):
        logger.info(f"Limiting dataset to {max_samples} samples")
        dataset = dataset.select(range(max_samples))

    total_examples = len(dataset)
    logger.info(f"Dataset loaded with {total_examples:,} examples")

    # Validate column
    if prompt_column:
        if prompt_column not in dataset.column_names:
            logger.error(
                f"Column '{prompt_column}' not found. Available columns: {dataset.column_names}"
            )
            sys.exit(1)
        logger.info(f"Using prompt column: '{prompt_column}'")
        use_messages = False
    else:
        if messages_column not in dataset.column_names:
            logger.error(
                f"Column '{messages_column}' not found. Available columns: {dataset.column_names}"
            )
            sys.exit(1)
        logger.info(f"Using messages column: '{messages_column}'")
        use_messages = True

    # Get model's max context length for filtering
    effective_max_len = model.config.max_position_embeddings
    logger.info(f"Model max context length: {effective_max_len}")

    # Prepare prompts and tokenize
    logger.info("Preparing and tokenizing prompts...")
    valid_input_ids = []
    valid_indices = []
    skipped_info = []

    for i, example in enumerate(dataset):
        if use_messages:
            messages = example[messages_column]
            # Optionally truncate messages up to last occurrence of a given role
            if messages_until_role:
                truncated = []
                for msg in messages:
                    truncated.append(msg)
                    if msg["role"] == messages_until_role:
                        last_truncated = list(truncated)
                messages = last_truncated
            prompt = tokenizer.apply_chat_template(
                messages, tokenize=False, add_generation_prompt=True
            )
        else:
            user_prompt = example[prompt_column]
            messages = [{"role": "user", "content": user_prompt}]
            prompt = tokenizer.apply_chat_template(
                messages, tokenize=False, add_generation_prompt=True
            )

        input_ids = tokenizer.encode(prompt)

        if skip_long_prompts:
            if len(input_ids) <= effective_max_len:
                valid_input_ids.append(input_ids)
                valid_indices.append(i)
            else:
                skipped_info.append((i, len(input_ids)))
        else:
            valid_input_ids.append(input_ids)
            valid_indices.append(i)

    # Log filtering results
    if skip_long_prompts and skipped_info:
        logger.warning(
            f"Skipped {len(skipped_info)} prompts exceeding context length ({effective_max_len} tokens)"
        )
        for idx, (prompt_idx, token_count) in enumerate(skipped_info[:10]):
            logger.info(
                f"  - Example {prompt_idx}: {token_count} tokens (exceeds by {token_count - effective_max_len})"
            )
        if len(skipped_info) > 10:
            logger.info(f"  ... and {len(skipped_info) - 10} more")

        skip_percentage = (len(skipped_info) / total_examples) * 100
        if skip_percentage > 10:
            logger.warning(f"WARNING: {skip_percentage:.1f}% of prompts were skipped!")

    if not valid_input_ids:
        logger.error("No valid prompts to process after filtering!")
        sys.exit(1)

    # Generate responses using continuous batching
    logger.info(f"Starting generation for {len(valid_input_ids):,} prompts using continuous batching...")
    logger.info(f"max_batch_tokens={max_batch_tokens}, max_new_tokens={max_tokens}")

    batch_outputs = model.generate_batch(
        inputs=valid_input_ids,
        generation_config=generation_config,
        progress_bar=True,
    )

    # Extract generated text
    logger.info("Extracting generated responses...")
    responses = [""] * total_examples

    for request_id, output in batch_outputs.items():
        # request_id is formatted as "req_0", "req_1", etc.
        idx = int(request_id.split("_", 1)[1])
        original_idx = valid_indices[idx]
        generated_text = tokenizer.decode(
            output.generated_tokens, skip_special_tokens=True
        )
        responses[original_idx] = generated_text.strip()

    # Count non-empty responses
    non_empty = sum(1 for r in responses if r)
    logger.info(f"Generated {non_empty:,} non-empty responses out of {total_examples:,} total")

    # Add responses to dataset
    logger.info("Adding responses to dataset...")
    dataset = dataset.add_column(output_column, responses)

    # Create dataset card
    logger.info("Creating dataset card...")
    card_content = create_dataset_card(
        source_dataset=src_dataset_hub_id,
        model_id=model_id,
        messages_column=messages_column,
        prompt_column=prompt_column,
        generation_config=generation_config,
        num_gpus=num_gpus,
        num_examples=total_examples,
        generation_time=generation_start_time,
        num_skipped=len(skipped_info) if skip_long_prompts else 0,
        attn_implementation=attn_implementation,
    )

    # Push to Hub
    logger.info(f"Pushing dataset to: {output_dataset_hub_id}")
    dataset.push_to_hub(output_dataset_hub_id, token=HF_TOKEN)

    card = DatasetCard(card_content)
    card.push_to_hub(output_dataset_hub_id, token=HF_TOKEN)

    logger.info("Generation complete!")
    logger.info(
        f"Dataset available at: https://huggingface.co/datasets/{output_dataset_hub_id}"
    )


if __name__ == "__main__":
    if len(sys.argv) > 1:
        parser = argparse.ArgumentParser(
            description="Generate responses using transformers continuous batching",
            formatter_class=argparse.RawDescriptionHelpFormatter,
            epilog="""
Examples:
  # Basic usage with default Qwen model
  uv run generate-responses.py input-dataset output-dataset \\
    --prompt-column question

  # With custom model and parameters
  uv run generate-responses.py input-dataset output-dataset \\
    --model-id meta-llama/Llama-3.1-8B-Instruct \\
    --messages-column messages \\
    --temperature 0.9 \\
    --max-tokens 2048

  # Increase batch token budget for better GPU utilization
  uv run generate-responses.py input-dataset output-dataset \\
    --prompt-column text \\
    --max-batch-tokens 2048
            """,
        )

        parser.add_argument(
            "src_dataset_hub_id",
            help="Input dataset on Hugging Face Hub (e.g., username/dataset-name)",
        )
        parser.add_argument(
            "output_dataset_hub_id",
            help="Output dataset name on Hugging Face Hub",
        )
        parser.add_argument(
            "--model-id",
            type=str,
            default="Qwen/Qwen3-4B-Instruct-2507",
            help="Model to use for generation (default: Qwen3-4B-Instruct-2507)",
        )
        parser.add_argument(
            "--subset",
            type=str,
            help="Dataset subset/config name (e.g., 'SFT' for HuggingFaceTB/smoltalk2)",
        )
        parser.add_argument(
            "--split",
            type=str,
            default="train",
            help="Dataset split to use (default: train)",
        )
        parser.add_argument(
            "--messages-column",
            type=str,
            default="messages",
            help="Column containing chat messages (default: messages)",
        )
        parser.add_argument(
            "--prompt-column",
            type=str,
            help="Column containing plain text prompts (alternative to --messages-column)",
        )
        parser.add_argument(
            "--messages-until-role",
            type=str,
            help="Truncate messages up to the last message with this role. "
            "Use 'user' to strip existing assistant responses and regenerate them.",
        )
        parser.add_argument(
            "--output-column",
            type=str,
            default="response",
            help="Column name for generated responses (default: response)",
        )
        parser.add_argument(
            "--max-samples",
            type=int,
            help="Maximum number of samples to process (default: all)",
        )
        parser.add_argument(
            "--temperature",
            type=float,
            default=0.7,
            help="Sampling temperature (default: 0.7)",
        )
        parser.add_argument(
            "--top-p",
            type=float,
            default=0.8,
            help="Top-p sampling parameter (default: 0.8)",
        )
        parser.add_argument(
            "--top-k",
            type=int,
            default=20,
            help="Top-k sampling parameter (default: 20)",
        )
        parser.add_argument(
            "--max-tokens",
            type=int,
            default=4096,
            help="Maximum tokens to generate (default: 4096)",
        )
        parser.add_argument(
            "--repetition-penalty",
            type=float,
            default=1.0,
            help="Repetition penalty (default: 1.0)",
        )
        parser.add_argument(
            "--max-batch-tokens",
            type=int,
            default=512,
            help="Token budget per batch for continuous batching scheduler (default: 512). "
            "Increase for better GPU utilization on large GPUs (e.g., 2048-4096 for A100/H100).",
        )
        parser.add_argument(
            "--dtype",
            type=str,
            default="bfloat16",
            choices=["bfloat16", "float16", "float32"],
            help="Model dtype (default: bfloat16)",
        )
        parser.add_argument(
            "--attn-implementation",
            type=str,
            default="paged|sdpa",
            help="Attention implementation (default: paged|sdpa). "
            "Use 'paged|flash_attention_2' if flash-attn is installed.",
        )
        parser.add_argument(
            "--hf-token",
            type=str,
            help="Hugging Face token (can also use HF_TOKEN env var)",
        )
        parser.add_argument(
            "--skip-long-prompts",
            action="store_true",
            default=True,
            help="Skip prompts exceeding context length (default: True)",
        )
        parser.add_argument(
            "--no-skip-long-prompts",
            dest="skip_long_prompts",
            action="store_false",
            help="Fail on prompts that exceed context length",
        )

        args = parser.parse_args()

        main(
            src_dataset_hub_id=args.src_dataset_hub_id,
            output_dataset_hub_id=args.output_dataset_hub_id,
            model_id=args.model_id,
            messages_column=args.messages_column,
            prompt_column=args.prompt_column,
            messages_until_role=args.messages_until_role,
            output_column=args.output_column,
            temperature=args.temperature,
            top_p=args.top_p,
            top_k=args.top_k,
            max_tokens=args.max_tokens,
            repetition_penalty=args.repetition_penalty,
            max_batch_tokens=args.max_batch_tokens,
            dtype=args.dtype,
            attn_implementation=args.attn_implementation,
            subset=args.subset,
            split=args.split,
            skip_long_prompts=args.skip_long_prompts,
            max_samples=args.max_samples,
            hf_token=args.hf_token,
        )
    else:
        print("""
Transformers Continuous Batching - Response Generation
======================================================

This script requires arguments. For usage information:
    uv run generate-responses.py --help

Why transformers CB instead of vLLM?
  - Works with ANY model supported by transformers (new models immediately!)
  - No vLLM/flashinfer dependency issues
  - Simpler setup - no custom Docker images or wheel indexes needed
  - ~95% of vLLM throughput with PagedAttention and continuous scheduling

Example HF Jobs command:
    hf jobs uv run \\
        --flavor l4x1 \\
        -s HF_TOKEN \\
        https://huggingface.co/datasets/uv-scripts/transformers-inference/raw/main/generate-responses.py \\
        username/input-dataset \\
        username/output-dataset \\
        --prompt-column question \\
        --model-id Qwen/Qwen3-4B-Instruct-2507 \\
        --temperature 0.7 \\
        --max-tokens 4096
        """)