[Bug] Qwen Image FP8 Offload Crash: infer_block() Unexpected Keyword Argument temb on RTX 5090 (main branch)

[metacryptom]

Title

Qwen Image FP8 Offload Crash: infer_block() Unexpected Keyword Argument temb on RTX 5090 (main branch)

---

Description

When enabling offload in examples/qwen_image/qwen_2511_fp8.py and running the Qwen Image FP8 example, the program crashes during inference. The error indicates an API/signature mismatch between the offload inference path and the base transformer inference implementation, where infer_block() is called with an unsupported keyword argument temb.

---

Steps to Reproduce

1. Checkout the main branch of LightX2V.
2. Modify examples/qwen_image/qwen_2511_fp8.py to enable offload (so the pipeline uses the offload inference path).
3. Run:

   python3 examples/qwen_image/qwen_2511_fp8.py
   

---

Expected Result

With offload enabled, the pipeline should run normally through the diffusion steps (e.g., progress beyond step_index: 1 / 8) and complete image generation without crashing.

---

Actual Result

The program crashes during inference at the first step with:

• TypeError: QwenImageTransformerInfer.infer_block() got an unexpected keyword argument 'temb'

(Previously, an additional mismatch was observed: infer_with_blocks_offload() takes 7 positional arguments but 8 were given. After adjusting the call site, the run proceeds to the current temb keyword error.)

---

Environment Information

• Operating System: Ubuntu (version not specified)
• GPU: NVIDIA RTX 5090
• Python: 3.10
• Branch: main
• Commit ID: main branch (exact commit not pinned)

---

Log Information

2025-12-24 04:26:27.828 | INFO     | lightx2v.models.runners.qwen_image.qwen_image_runner:run:175 - ==> step_index: 1 / 8
Traceback (most recent call last):
  File "/data/aimodels/Qwen/LightX2V/examples/qwen_image/qwen_2511_fp8.py", line 48, in <module>
    pipe.generate(
  File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 120, in decorate_context
    return func(*args, **kwargs)
  File "/data/aimodels/Qwen/LightX2V/lightx2v/pipeline.py", line 377, in generate
    self.runner.run_pipeline(input_info)
  File "/data/aimodels/Qwen/LightX2V/lightx2v/models/runners/qwen_image/qwen_image_runner.py", line 299, in run_pipeline
    latents, generator = self.run_dit()
  File "/data/aimodels/Qwen/LightX2V/lightx2v/models/runners/qwen_image/qwen_image_runner.py", line 94, in _run_dit_local
    latents, generator = self.run(total_steps)
  File "/data/aimodels/Qwen/LightX2V/lightx2v/models/runners/qwen_image/qwen_image_runner.py", line 181, in run
    self.model.infer(self.inputs)
  File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 120, in decorate_context
    return func(*args, **kwargs)
  File "/data/aimodels/Qwen/LightX2V/lightx2v/models/networks/qwen_image/model.py", line 356, in infer
    noise_pred = self._infer_cond_uncond(latents_input, inputs["text_encoder_output"]["prompt_embeds"], infer_condition=True)
  File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 120, in decorate_context
    return func(*args, **kwargs)
  File "/data/aimodels/Qwen/LightX2V/lightx2v/models/networks/qwen_image/model.py", line 374, in _infer_cond_uncond
    hidden_states = self.transformer_infer.infer(
  File "/data/aimodels/Qwen/LightX2V/lightx2v/models/networks/qwen_image/infer/transformer_infer.py", line 255, in infer
    hidden_states = self.infer_func(block_weights, hidden_states, encoder_hidden_states, temb_img_silu, temb_txt_silu, image_rotary_emb)#, self.scheduler.modulate_index)
  File "/data/aimodels/Qwen/LightX2V/lightx2v/models/networks/qwen_image/infer/offload/transformer_infer.py", line 40, in infer_with_blocks_offload
    encoder_hidden_states, hidden_states = self.infer_block(
TypeError: QwenImageTransformerInfer.infer_block() got an unexpected keyword argument 'temb'

---

Additional Information

• The only intentional change was enabling offload in examples/qwen_image/qwen_2511_fp8.py.
• The stack trace suggests the offload implementation calls infer_block() with a temb keyword, but the base infer_block() signature does not accept it, indicating the offload path and base inference code may be out of sync on the main branch.