theluyuan/DiffSynth-Studio
1
0
Fork 0
mirror of https://github.com/modelscope/DiffSynth-Studio.git synced 2026-03-19 06:39:43 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix

Browse Source
This commit is contained in:
co63oc
2025-02-26 14:18:36 +08:00
parent bed770248b
commit 4268f5466b
9 changed files with 14 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
diffsynth/models/kolors_text_encoder.py
View File

@@ -980,7 +980,7 @@ class Embedding(torch.nn.Module):
# Embeddings.
words_embeddings = self.word_embeddings(input_ids)
embeddings = words_embeddings
# Data format change to avoid explicit tranposes : [b s h] --> [s b h].
# Data format change to avoid explicit transposes : [b s h] --> [s b h].
embeddings = embeddings.transpose(0, 1).contiguous()
# If the input flag for fp32 residual connection is set, convert for float.
if self.fp32_residual_connection:

4
diffsynth/models/stepvideo_dit.py
View File

@@ -398,7 +398,7 @@ class RoPE1D:
* tokens: batch_size x ntokens x nheads x dim
* positions: batch_size x ntokens (t position of each token)
output:
* tokens after appplying RoPE2D (batch_size x ntokens x nheads x dim)
* tokens after applying RoPE2D (batch_size x ntokens x nheads x dim)
"""
D = tokens.size(3)
assert positions.ndim == 2 # Batch, Seq
@@ -428,7 +428,7 @@ class RoPE3D(RoPE1D):
* tokens: batch_size x ntokens x nheads x dim
* rope_positions: list of (f, h, w)
output:
* tokens after appplying RoPE2D (batch_size x ntokens x nheads x dim)
* tokens after applying RoPE2D (batch_size x ntokens x nheads x dim)
"""
assert sum(ch_split) == tokens.size(-1);

6
diffsynth/models/stepvideo_text_encoder.py
View File

@@ -88,7 +88,7 @@ class LLaMaEmbedding(nn.Module):
embeddings = embeddings.to(self.params_dtype)
self.word_embeddings = self.word_embeddings.to(self.params_dtype)
# Data format change to avoid explicit tranposes : [b s h] --> [s b h].
# Data format change to avoid explicit transposes : [b s h] --> [s b h].
embeddings = embeddings.transpose(0, 1).contiguous()
# If the input flag for fp32 residual connection is set, convert for float.
@@ -326,7 +326,7 @@ class MultiQueryAttention(nn.Module):
dim=-1,
)
# gather on 1st dimention
# gather on 1st dimension
xq = xq.view(seqlen, bsz, self.n_local_heads, self.head_dim)
xkv = xkv.view(seqlen, bsz, self.n_local_groups, 2 * self.head_dim)
xk, xv = xkv.chunk(2, -1)
@@ -357,7 +357,7 @@ class MultiQueryAttention(nn.Module):
output = self.core_attention(xq, xk, xv,
cu_seqlens=cu_seqlens,
max_seq_len=max_seq_len)
# reduce-scatter only support first dimention now
# reduce-scatter only support first dimension now
output = rearrange(output, "b s h d -> s b (h d)").contiguous()
else:
xq, xk, xv = [

2
diffsynth/models/tiler.py
View File

@@ -55,7 +55,7 @@ class TileWorker:
def io_scale(self, model_output, tile_size):
# Determine the size modification happend in forward_fn
# Determine the size modification happened in forward_fn
# We only consider the same scale on height and width.
io_scale = model_output.shape[2] / tile_size
return io_scale

2
diffsynth/pipelines/omnigen_image.py
View File

@@ -16,7 +16,7 @@ class OmniGenCache(DynamicCache):
def __init__(self,
num_tokens_for_img: int, offload_kv_cache: bool=False) -> None:
if not torch.cuda.is_available():
print("No avaliable GPU, offload_kv_cache wiil be set to False, which will result in large memory usage and time cost when input multiple images!!!")
print("No available GPU, offload_kv_cache will be set to False, which will result in large memory usage and time cost when input multiple images!!!")
offload_kv_cache = False
raise RuntimeError("OffloadedCache can only be used with a GPU")
super().__init__()