AttributeError in _get_actual_bias when using FP8 quantized T5 with phase-level CPU offload

[vbhakta8]

Environment

• LightX2V version: latest main branch
• Python: 3.11
• PyTorch: 2.x
• GPU: NVIDIA CUDA

Bug Description

When running WAN 2.2 I2V with FP8 quantized T5 encoder and CPU offload at phase granularity, the _get_actual_bias() method in mm_weight.py raises an AttributeError because it accesses self.bias directly without checking if the attribute exists.

Configuration

{
  "t5_cpu_offload": true,
  "t5_offload_granularity": "phase",
  "t5_quantized": true,
  "t5_quant_scheme": "fp8-q8f"
}

Error Traceback

File "lightx2v/models/input_encoders/hf/wan/t5/model.py", line 524, in forword_attn_with_offload
    q = attn_phase.attn_q.apply(x.squeeze(0)).view(b, -1, n, c)
File "lightx2v/common/ops/mm/mm_weight.py", line 1319, in apply
    self._get_actual_bias(),
File "lightx2v/common/ops/mm/mm_weight.py", line 152, in _get_actual_bias
    if self.bias is None:
       ^^^^^^^^^
AttributeError: 'MMWeightWfp8channelAfp8channeldynamicQ8F' object has no attribute 'bias'

Root Cause

In MMWeightTemplate._get_actual_bias() (line 152), the code directly accesses self.bias without first checking if the attribute exists:

def _get_actual_bias(self, bias=None):
    if bias is not None:
        ...
    else:
        if self.bias is None:  # <-- AttributeError if self.bias doesn't exist!
            return None

When using create_cpu_buffer=True with phase-level offload, the load_quantized() method only initializes self.bias = None if bias is in base_attrs. However, _update_base_attrs() only adds bias to base_attrs if bias_name is not None. For attention layers without a bias term (like T5 attention Q/K/V projections), the bias attribute is never created.

Proposed Fix

Change line 152 in lightx2v/common/ops/mm/mm_weight.py:

# Before
if self.bias is None:

# After  
if not hasattr(self, "bias") or self.bias is None:

Workarounds

• Use t5_offload_granularity: "block" instead of "phase"
• Use a different T5 quant scheme (e.g., fp8-vllm or int8-vllm)
• Disable T5 quantization: t5_quantized: false

[vbhakta8]

#782