theluyuan/DiffSynth-Studio
1
0
Fork 0
mirror of https://github.com/modelscope/DiffSynth-Studio.git synced 2026-03-25 18:58:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

fea : enable wan video usp for arbitrary seq len

Browse Source
This commit is contained in:
handoku
2025-07-08 16:43:43 +08:00
parent 89397c755a
commit 00279a8375
2 changed files with 26 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
diffsynth/distributed/xdit_context_parallel.py
View File

@@ -26,15 +26,12 @@ def pad_freqs(original_tensor, target_len):
def rope_apply(x, freqs, num_heads): def rope_apply(x, freqs, num_heads):
x = rearrange(x, "b s (n d) -> b s n d", n=num_heads) x = rearrange(x, "b s (n d) -> b s n d", n=num_heads)
s_per_rank = x.shape[1]
x_out = torch.view_as_complex(x.to(torch.float64).reshape( x_out = torch.view_as_complex(x.to(torch.float64).reshape(
x.shape[0], x.shape[1], x.shape[2], -1, 2)) x.shape[0], x.shape[1], x.shape[2], -1, 2))
sp_size = get_sequence_parallel_world_size()
sp_rank = get_sequence_parallel_rank() sp_rank = get_sequence_parallel_rank()
freqs = pad_freqs(freqs, s_per_rank * sp_size) freqs_rank = torch.chunk(freqs, dim=0)[sp_rank] # chunk freqs like x
freqs_rank = freqs[(sp_rank * s_per_rank):((sp_rank + 1) * s_per_rank), :, :]
x_out = torch.view_as_real(x_out * freqs_rank).flatten(2) x_out = torch.view_as_real(x_out * freqs_rank).flatten(2)
return x_out.to(x.dtype) return x_out.to(x.dtype)
@@ -73,6 +70,9 @@ def usp_dit_forward(self,
return custom_forward return custom_forward
# Context Parallel # Context Parallel
chunks = torch.chunk(x, get_sequence_parallel_world_size(), dim=1)
seq_lens = [chunk.shape[1] for chunk in chunks]
x = torch.chunk( x = torch.chunk(
x, get_sequence_parallel_world_size(), x, get_sequence_parallel_world_size(),
dim=1)[get_sequence_parallel_rank()] dim=1)[get_sequence_parallel_rank()]
@@ -98,7 +98,15 @@ def usp_dit_forward(self,
x = self.head(x, t) x = self.head(x, t)
# Context Parallel # Context Parallel
max_len = seq_lens[0]
b, s, c = x.shape
if s != max_len:
padding_tensor = torch.ones(b, max_len - s, c, dtype=x.dtype, device=x.device)
x = torch.cat([x, padding_tensor], dim=1)
x = get_sp_group().all_gather(x, dim=1) x = get_sp_group().all_gather(x, dim=1)
# remove pad
x = torch.cat([x[:,max_len*id:seq_lens[id],:] for id in range(seq_lens)])
# unpatchify # unpatchify
x = self.unpatchify(x, (f, h, w)) x = self.unpatchify(x, (f, h, w))

15
diffsynth/pipelines/wan_video.py
View File

@@ -594,7 +594,12 @@ def model_fn_wan_video(
# blocks # blocks
if use_unified_sequence_parallel: if use_unified_sequence_parallel:
if dist.is_initialized() and dist.get_world_size() > 1: if dist.is_initialized() and dist.get_world_size() > 1:
x = torch.chunk(x, get_sequence_parallel_world_size(), dim=1)[get_sequence_parallel_rank()] chunks = torch.chunk(x, get_sequence_parallel_world_size(), dim=1)
seq_lens = [chunk.shape[1] for chunk in chunks]
x = torch.chunk(
x, get_sequence_parallel_world_size(),
dim=1)[get_sequence_parallel_rank()]
if tea_cache_update: if tea_cache_update:
x = tea_cache.update(x) x = tea_cache.update(x)
else: else:
@@ -612,6 +617,14 @@ def model_fn_wan_video(
x = dit.head(x, t) x = dit.head(x, t)
if use_unified_sequence_parallel: if use_unified_sequence_parallel:
if dist.is_initialized() and dist.get_world_size() > 1: if dist.is_initialized() and dist.get_world_size() > 1:
max_len = seq_lens[0]
b, s, c = x.shape
if s != max_len:
padding_tensor = torch.ones(b, max_len - s, c, dtype=x.dtype, device=x.device)
x = torch.cat([x, padding_tensor], dim=1)
x = get_sp_group().all_gather(x, dim=1) x = get_sp_group().all_gather(x, dim=1)
# remove pad
x = torch.cat([x[:,max_len*id:seq_lens[id],:] for id in range(seq_lens)])
x = dit.unpatchify(x, (f, h, w)) x = dit.unpatchify(x, (f, h, w))
return x return x