tp_gpt_oss.py

import os
import torch
from torch import distributed as dist
from transformers import GptOssForCausalLM, PreTrainedTokenizerFast

def initialize_process():
    # torchrun exports: LOCAL_RANK
    local_rank = int(os.environ["LOCAL_RANK"])
    torch.cuda.set_device(local_rank)
    dist.init_process_group(backend="nccl")

def run_inference():
    model_id = "openai/gpt-oss-120b"
    tok = PreTrainedTokenizerFast.from_pretrained(model_id)

    # built in TP
    model = GptOssForCausalLM.from_pretrained(
        model_id,
        tp_plan="auto",
        torch_dtype="auto",
    ).eval()

    messages = [
        {"role": "system", "content": "Be concise."},
        {"role": "user", "content": "Explain KV caching briefly."},
    ]
    inputs = tok.apply_chat_template(
        messages,
        add_generation_prompt=True,
        return_tensors="pt",
        return_dict=True,
        reasoning_effort="low",
    )

    local_rank = int(os.environ["LOCAL_RANK"])
    device = torch.device(f"cuda:{local_rank}")
    inputs = {k: v.to(device, non_blocking=True) for k, v in inputs.items()}

    with torch.inference_mode():
        out = model.generate(**inputs, max_new_tokens=128)
        torch.cuda.synchronize(device)

    # keep output from rank 0 only
    dist.barrier()
    if dist.get_rank() == 0:
        print(tok.decode(out[0][inputs["input_ids"].shape[-1]:]))

def main():
    initialize_process()
    try:
        run_inference()
    finally:
        dist.destroy_process_group()

if __name__ == "__main__":
    main()