ValueError: Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?)

🐛 Describe the bug

Getting ValueError: Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?) when doing inference using HF from_pretrained() with device_map="auto".

Error

File "/home/coder/Liger-Kernel/src/liger_kernel/ops/swiglu.py", line 111, in forward
    a, b, c = swiglu_forward(a, b)
              ^^^^^^^^^^^^^^^^^^^^
  File "/home/coder/Liger-Kernel/src/liger_kernel/ops/swiglu.py", line 74, in swiglu_forward
    _swiglu_forward_kernel[(n_rows,)](
  File "/usr/local/lib/python3.11/site-packages/triton/runtime/jit.py", line 345, in <lambda>
    return lambda *args, **kwargs: self.run(grid=grid, warmup=False, *args, **kwargs)
                                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.11/site-packages/triton/runtime/jit.py", line 691, in run
    kernel.run(grid_0, grid_1, grid_2, stream, kernel.function, kernel.packed_metadata, launch_metadata,
  File "/usr/local/lib/python3.11/site-packages/triton/backends/nvidia/driver.py", line 365, in __call__
    self.launch(*args, **kwargs)
ValueError: Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?)

Reproduce

from liger_kernel.transformers import AutoLigerKernelForCausalLM
from transformers import AutoTokenizer, AutoModelForCausalLM
from liger_kernel.transformers import apply_liger_kernel_to_qwen2
import torch

model_name = "Qwen/Qwen2.5-Math-1.5B-Instruct"
device = "cuda"  # the device to load the model onto

apply_liger_kernel_to_qwen2(cross_entropy=False,
                            fused_linear_cross_entropy=False,
                            rms_norm=True,
                            rope=True,
                            swiglu=True)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    trust_remote_code=True,
    use_cache=False,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Find the value of $x$ that satisfies the equation $4x+5 = 6x+7$."

# CoT
messages = [
    {"role": "system", "content": "Please reason step by step, and put your final answer within \\boxed{}."},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

print(f"{response=}")

Versions

Operating System: Linux-4.4.0-x86_64-with-glibc2.36
Python version: 3.11.5
PyTorch version: 2.5.1+cu124
CUDA version: 12.4
Triton version: 3.1.0
Transformers version: 4.46.3
GPU: L4 x 4
Liger-Kernel version: 0.4.2
Accelerate version: 1.1.1

Start by reproducing the provided Qwen inference example with device_map="auto", then inspect src/liger_kernel/ops/swiglu.py at swiglu_forward and the forward method around lines 74 and 111. Check how device placement reaches the Triton kernel; done means the example runs inference without passing a CPU tensor to Triton.