nvfp4 w4a4 doesn't work on Qwen/Qwen3-235b-A22B

[ehartford]

Describe the bug
NVFP4 quantization fails on Qwen3-235B (MoE) because FX tracing hits torch.vmap proxies in transformers masking_utils

Expected behavior
It should work

Environment

package	version
llm-compressor	0.6.0 (latest PyPI)
transformers	4.43.0
torch	2.3.1 + CUDA 12.1
python	3.11
GPU	A100 / H100 (same on both)

To Reproduce

from transformers import AutoModelForCausalLM, AutoTokenizer
from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier
import torch, datasets

model_id  = "Qwen/Qwen3-235B-A22B"
model     = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto")
tok       = AutoTokenizer.from_pretrained(model_id)
ds        = datasets.load_dataset("HuggingFaceH4/ultrachat_200k",
                                  split="train_sft[:8]")  # tiny calib

recipe = QuantizationModifier(targets="Linear", scheme="NVFP4", ignore=["lm_head"])
oneshot(model=model, dataset=ds, recipe=recipe, max_seq_length=2048,
        num_calibration_samples=8)          # --> TraceError / vmap-proxy error

Errors

ValueError: vmap(wrapped, in_dims=(0, None, None, None), ...)(<inputs>):
Got in_dim=0 for an input but the input is of type
<class 'transformers.utils.fx.HFProxy'>. We cannot vmap over non-Tensor arguments

Additional context

[dsikka]

@ehartford As listed in the ReadMe, MoE support has not landed yet.
This is not yet a supported case or expected to work.

Please consider reading the docs/readme for up-to-date information on supported features.

You are free to try a WIP branch: #1596

[ehartford]

How about adding a sensible error message like "nvfp4 is not yet supported for MoE models"