import os,sys,time,torch,json
os.environ.setdefault("HF_HOME","/workspace/.hf_home")
import torch.multiprocessing as mp
def worker(rank,q):
    torch.cuda.set_device(rank)
    from transformers import AutoModelForCausalLM, BitsAndBytesConfig
    from peft import LoraConfig, get_peft_model
    from liger_kernel.transformers import _apply_liger_kernel_to_instance
    M="/workspace/models/gemma4-qat-unquantized"
    bnb=BitsAndBytesConfig(load_in_4bit=True,bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=torch.bfloat16,bnb_4bit_use_double_quant=True)
    m=AutoModelForCausalLM.from_pretrained(M,quantization_config=bnb,dtype=torch.bfloat16,
        device_map={"":rank},attn_implementation="sdpa")
    _apply_liger_kernel_to_instance(model=m); m.config.use_cache=False
    m.enable_input_require_grads()
    mm=get_peft_model(m,LoraConfig(r=128,lora_alpha=256,lora_dropout=0.05,bias="none",
        task_type="CAUSAL_LM",target_modules=["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"]))
    mm.train()
    opt=torch.optim.AdamW([p for p in mm.parameters() if p.requires_grad],lr=1e-5)
    ids=torch.randint(0,250000,(1,2048),device=rank)
    for i in range(6):
        if i==2: torch.cuda.synchronize(); t=time.time()
        out=mm(input_ids=ids,labels=ids); out.loss.backward(); opt.step(); opt.zero_grad(set_to_none=True)
    torch.cuda.synchronize(); dt=(time.time()-t)/4
    q.put((rank, round(2048/dt), round(torch.cuda.max_memory_allocated(rank)/2**30,1)))
if __name__=="__main__":
    mp.set_start_method("spawn"); q=mp.Queue()
    ps=[mp.Process(target=worker,args=(r,q)) for r in range(2)]
    [p.start() for p in ps]; res=[q.get() for _ in range(2)]; [p.join() for p in ps]
    tot=sum(r[1] for r in res)
    for r in sorted(res): print(f"  GPU{r[0]}: {r[1]} tok/s  peak {r[2]} GiB")
    print(f"AGGREGATE {tot} tok/s across 2 GPUs")