Question about wan_va_server.py

[liyizhep]

I noticed that the infer() function in wan_va_server.py contains three logic branches: reset, compute KV cache, and infer one chunk.

I understand that the reset branch is triggered at the very beginning of a task when the client sends a dictionary where the reset key is set to True and the prompt key contains the text description.

Regarding the other two branches, I have the following questions:

Dictionary Structure: When the inference process enters the compute KV cache or infer one chunk branches, what specific key-value pairs must be included in the dictionary sent by the client, and what is the exact format for each?

Timing: At what specific stages or under what conditions should the client send the dictionaries required for compute KV cache and infer one chunk respectively?

[liyizhep]

I have another question: The paper mentions the introduction of a Forward Dynamics Model (FDM) to imagine future visuals as a way to calibrate stale predictions.

Is this part implemented in the server-side code? Specifically, where in the codebase is this logic reflected or integrated?

[lilin-hitcrt]

I noticed that the infer() function in wan_va_server.py contains three logic branches: reset, compute KV cache, and infer one chunk.

I understand that the reset branch is triggered at the very beginning of a task when the client sends a dictionary where the reset key is set to True and the prompt key contains the text description.

Regarding the other two branches, I have the following questions:

Dictionary Structure: When the inference process enters the compute KV cache or infer one chunk branches, what specific key-value pairs must be included in the dictionary sent by the client, and what is the exact format for each?

Timing: At what specific stages or under what conditions should the client send the dictionaries required for compute KV cache and infer one chunk respectively?

Hi

Thank you for your interest in our work! We suggest you refer to the code of robotwin client or libero client.

[liyizhep]

I noticed that the infer() function in wan_va_server.py contains three logic branches: reset, compute KV cache, and infer one chunk.
I understand that the reset branch is triggered at the very beginning of a task when the client sends a dictionary where the reset key is set to True and the prompt key contains the text description.
Regarding the other two branches, I have the following questions:
Dictionary Structure: When the inference process enters the compute KV cache or infer one chunk branches, what specific key-value pairs must be included in the dictionary sent by the client, and what is the exact format for each?
Timing: At what specific stages or under what conditions should the client send the dictionaries required for compute KV cache and infer one chunk respectively?

Hi

Thank you for your interest in our work! We suggest you refer to the code of robotwin client or libero client.

Thank you for your response. Based on the client-side code you provided, I now understand the specific dictionary format that the client should send in each scenario.

However, both sets of code you provided seem to use a synchronous mechanism. I was wondering if there is any FDM-grounded Async code as mentioned in the papers? Specifically, do the current training and server-side implementations incorporate FDM?

[ludekcizinsky]

same question regarding the async and the FDM model