[RFC]: Command "dreamverse-server --host 0.0.0.0 --port 8009": GPU 0 warmup failed

[nanwanj]

Motivation.

We followed https://github.com/hao-ai-lab/FastVideo/blob/main/apps/dreamverse/README.md#quick-start-local-gpu to start dreamverse server on a B200 vm.

when running command "dreamverse-server --host 0.0.0.0 --port 8009", there will be error like:

1;36m(Worker pid=418572)�[0;0m INFO 08-06 00:16:34.816 [lora_pipeline.py:252] Converting transformer to LoRA Transformer
�[1;36m(Worker pid=418572)�[0;0m ERROR 08-06 00:16:34.820 [multiproc_executor.py:584] WorkerMultiprocProc failed.
�[1;36m(Worker pid=418572)�[0;0m ERROR 08-06 00:16:34.820 [multiproc_executor.py:584] Traceback (most recent call last):
�[1;36m(Worker pid=418572)�[0;0m ERROR 08-06 00:16:34.820 [multiproc_executor.py:584]   File "/home/ubuntu/FastVideo/fastvideo/worker/multiproc_executor.py", line 566, in worker_main
�[1;36m(Worker pid=418572)�[0;0m ERROR 08-06 00:16:34.820 [multiproc_executor.py:584]     worker.worker_busy_loop()
�[1;36m(Worker pid=418572)�[0;0m ERROR 08-06 00:16:34.820 [multiproc_executor.py:584]   File "/home/ubuntu/FastVideo/fastvideo/worker/multiproc_executor.py", line 692, in worker_busy_loop
�[1;36m(Worker pid=418572)�[0;0m ERROR 08-06 00:16:34.820 [multiproc_executor.py:584]     output_batch = self.worker.execute_forward(forward_batch, fastvideo_args)
�[1;36m(Worker pid=418572)�[0;0m ERROR 08-06 00:16:34.820 [multiproc_executor.py:584]                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
�[1;36m(Worker pid=418572)�[0;0m ERROR 08-06 00:16:34.820 [multiproc_executor.py:584]   File "/home/ubuntu/FastVideo/fastvideo/worker/gpu_worker.py", line 75, in execute_forward
�[1;36m(Worker pid=418572)�[0;0m ERROR 08-06 00:16:34.820 [multiproc_executor.py:584]     output_batch = self.pipeline.forward(forward_batch, self.fastvideo_args)
�[1;36m(Worker pid=418572)�[0;0m ERROR 08-06 00:16:34.820 [multiproc_executor.py:584]                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
�[1;36m(Worker pid=418572)�[0;0m ERROR 08-06 00:16:34.820 [multiproc_executor.py:584]   File "/home/ubuntu/FastVideo/.venv/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 124, in decorate_context
�[1;36m(Worker pid=418572)�[0;0m ERROR 08-06 00:16:34.820 [multiproc_executor.py:584]     return func(*args, **kwargs)
�[1;36m(Worker pid=418572)�[0;0m ERROR 08-06 00:16:34.820 [multiproc_executor.py:584]            ^^^^^^^^^^^^^^^^^^^^^

Full log will be attached.

log.txt

Proposed Change.

Can you please help and let's know what's going on here?

Thanks.

Feedback Period.

No response

CC List.

No response

Any Other Things.

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues.

[aryan5v]

Thanks for the detailed report. We traced the warmup failure to the interaction between NVFP4 weight purging and the LTX2 refinement LoRA:

• NVFP4 conversion purges the original dense transformer weights by default.
• The refinement LoRA is applied afterward, and the LoRA wrapper tries to capture the missing base weight (weight=None).
• That produces the reported AttributeError during DreamVerse startup.

I opened PR #1705, which carries an opt-in transformer_retain_original_weights setting through the typed API and enables it for DreamVerse's bundled LTX2 configuration. It also adds regression coverage for typed config propagation and LoRA wrapping of retained NVFP4 weights.

I validated the fix on a single Modal B200: NVFP4 retained the required dense weights, the refinement LoRA applied to all 1,344 layers, all three startup warmup segments completed, and the run reported ISSUE_1685_VALIDATION=PASS.

The B200 check used the current source mounted into the existing CUDA image; the release image should still be rebuilt and checked after the PR merges.