import os,sys,time,torch,json
os.environ.setdefault("HF_HOME","/workspace/.hf_home")
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
M="/workspace/models/gemma4-qat-unquantized"; rank=int(sys.argv[1])
torch.cuda.init(); _=torch.zeros(1,device="cuda:0")
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={"":0},attn_implementation="sdpa")
print(f"BASE_4BIT_WEIGHTS {torch.cuda.memory_allocated(0)/2**30:.2f} GiB",flush=True)
m.config.use_cache=False
m.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant":False})
m=prepare_model_for_kbit_training(m,use_gradient_checkpointing=True)
mm=get_peft_model(m,LoraConfig(r=rank,lora_alpha=rank*2,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"]))
tr=sum(p.numel() for p in mm.parameters() if p.requires_grad)
print(f"rank={rank} trainable={tr/1e6:.1f}M  after-LoRA {torch.cuda.memory_allocated(0)/2**30:.2f} GiB",flush=True)
opt=torch.optim.AdamW([p for p in mm.parameters() if p.requires_grad],lr=1e-5)
res={}
for seq,mb in ((2048,1),(2048,2),(2048,4),(2048,8),(4096,4),(8192,2)):
    try:
        torch.cuda.empty_cache(); torch.cuda.reset_peak_memory_stats(0)
        ids=torch.randint(0,250000,(mb,seq),device=0)
        for i in range(3):
            if i==1: 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)/2; pk=torch.cuda.max_memory_allocated(0)/2**30
        res[f"s{seq}_mb{mb}"]={"tok_s":round(mb*seq/dt),"peak_gib":round(pk,1)}
        print(f"  seq={seq} mb={mb}: {mb*seq/dt:8.0f} tok/s/GPU  peak {pk:5.1f} GiB  {dt:.2f}s/step",flush=True)
        del ids,out
    except torch.cuda.OutOfMemoryError:
        print(f"  seq={seq} mb={mb}: OOM",flush=True); torch.cuda.empty_cache()
res["trainable_M"]=round(tr/1e6,1)
json.dump(res,open(f"/workspace/aditi/logs/bench_r{rank}.json","w"),indent=2)
print("BENCH_DONE",flush=True)