I'm trying to load the quantized model RedHatAI/Qwen2.5-VL-7B-Instruct-quantized.w8a8 but encountering a dtype compatibility issue during model initialization. The model appears to be quantized using llmcompressor with W8A8 quantization scheme.

Note: I need to load this model without vLLM because I may need to add custom hooks for my research, so I'm looking for a direct loading method using transformers/llmcompressor.

Error Message

RuntimeError: expected a floating-point or complex dtype, but got dtype=torch.int8

Full Stack Trace:

File "/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py", line 366, in _init_weights
    module.weight.data.normal_(mean=0.0, std=std)
File "/torch/_refs/__init__.py", line 6214, in normal_
    return normal(mean, std, self.shape, out=self, generator=generator)
...
RuntimeError: expected a floating-point or complex dtype, but got dtype=torch.int8

Traceback

The error occurs during model weight initialization where transformers tries to call normal_() on int8 tensors. The normal_() function in PyTorch only works with floating-point tensors, but the quantized model contains int8 weights.

Specific failure point:

• File: modeling_qwen2_5_vl.py, line 366
• Function: _init_weights()
• Operation: module.weight.data.normal_(mean=0.0, std=std)
• Issue: Trying to apply normal distribution to int8 tensors

Model Information

Based on the model's config.json:

• Quantization method: compressed-tensors
• Format: int-quantized
• Scheme: W8A8 (8-bit weights and activations)
• Base model: Qwen/Qwen2.5-VL-7B-Instruct
• Compression ratio: ~1.2x
• Ignored layers: All visual layers (visual.blocks.*, visual.merger.*, lm_head)

What I've Tried

1. llmcompressor methods:

# Method 1: TraceableQwen2_5_VLForConditionalGeneration
from llmcompressor.transformers.tracing import TraceableQwen2_5_VLForConditionalGeneration
model = TraceableQwen2_5_VLForConditionalGeneration.from_pretrained(
    model_path, device_map="auto", torch_dtype="auto", trust_remote_code=True
)

# Method 2: SparseAutoModelForCausalLM  
from llmcompressor.transformers import SparseAutoModelForCausalLM
model = SparseAutoModelForCausalLM.from_pretrained(
    model_path, device_map="auto", torch_dtype="auto", trust_remote_code=True
)

2. Standard transformers methods:

# Method 3: Various dtype configurations
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,  # Also tried: torch.float16, "auto", None
    trust_remote_code=True,
    device_map="auto"
)

# Method 4: AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(
    model_path, trust_remote_code=True, torch_dtype="auto"
)

All methods fail at the same weight initialization step, so I wonder should the model be loaded with _fast_init=False or other special parameters?

Additional Observations

1. Warning about ignored layers: The loader warns about missing visual layers, but this seems expected since they were ignored during quantization
2. Model files exist: The quantized model directory contains the expected .safetensors files and configuration
3. Original model works: The base Qwen/Qwen2.5-VL-7B-Instruct loads and works perfectly

Environment

• Python: 3.10
• PyTorch: 2.7.0+cu126
• Transformers: 4.52.4
• LLMCompressor: 0.6.0
• Compressed-tensors: 0.10.2

This model was likely created using llmcompressor's oneshot quantization:

from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.transformers import oneshot

recipe = [
    GPTQModifier(
        targets="Linear",
        scheme="W8A8", 
        sequential_targets=["Qwen2_5_VLDecoderLayer"],
        ignore=["lm_head", "re:visual.*"],
    ),
]

If this is more of an llmcompressor-specific model loading issue rather than a transformers compatibility issue, please let me know and I'll file this issue in the llmcompressor repository instead.