How to inference longer than 5s?

[NeverMoreLCH]

Great work!

The inference.py likes only supporting generate videos with 81 frames (5s). If I want to generate videos such as 10s, how to modify the inference.py? I've tried to achieve this by setting num_frames in default_config.yaml to 161, but it didn't works.

Thanks!

[Sugar929]

hello, have you solved this problem?

[wikiwen]

Same problem @guandeh17 @xunhuang1995

[huchenjiang]

@NeverMoreLCH

I tried the following steps and it works.

1. change the value of local_attn_size in CausalWanModel other than -1. (When local_attn_size is set to -1, the new kv cache is always append to current_end and will exceed the kv_cache_size)

Self-Forcing/wan/modules/causal_model.py

Lines 223 to 228 in 33593df

	else:
	# Assign new keys/values directly up to current_end
	local_end_index = kv_cache["local_end_index"].item() + current_end - kv_cache["global_end_index"].item()
	local_start_index = local_end_index - num_new_tokens
	kv_cache["k"][:, local_start_index:local_end_index] = roped_key
	kv_cache["v"][:, local_start_index:local_end_index] = v

2. add --num_output_frames xxx in the inference script