Describe the bug
"batch_size_attention // self.slice_size" will forget to compute the last batch of attention。
should be this:
"(batch_size_attention - 1) // self.slice_size + 1"
I have submit a pull request for this issue:
#8836
Reproduction
importtorchfromdiffusers.models.attention_processorimportAttention, SlicedAttnProcessorattn=Attention(320, None, heads=5, dim_head=64)
slice_proc=SlicedAttnProcessor(slice_size=2)
hidden_states=torch.randn(1, 4096, 320)
result1=slice_proc(attn=attn, hidden_states=hidden_states)
hidden_states=torch.cat([hidden_states]*2, dim=0)
result2=slice_proc(attn=attn, hidden_states=hidden_states)
print(torch.sum(torch.abs(result1-result2[:1]))) ### Ideally, the numerical difference should be 0
Logs
No response
System Info
- __ Diffusers version: 0.29.0
- Platform: Linux-4.14.105-1-tlinux3-0013-x86_64-with-glibc2.28
- Running on a notebook?: No
- Running on Google Colab?: No
- Python version: 3.9.19
- PyTorch version (GPU?): 2.1.2+cu118 (True)
- Flax version (CPU?/GPU?/TPU?): not installed (NA)
- Jax version: not installed
- JaxLib version: not installed
- Huggingface_hub version: 0.23.3
- Transformers version: 4.41.2
- Accelerate version: 0.31.0
- PEFT version: not installed
- Bitsandbytes version: not installed
- Safetensors version: 0.4.3
- xFormers version: 0.0.23.post1+cu118
- Accelerator: Tesla V100-SXM2-32GB, 32510 MiB VRAM
- Using GPU in script?:
- Using distributed or parallel set-up in script?:
Who can help?
@DN6@yiyixuxu@sayakpaul
Describe the bug
"batch_size_attention // self.slice_size" will forget to compute the last batch of attention。
should be this:
"(batch_size_attention - 1) // self.slice_size + 1"
I have submit a pull request for this issue:
#8836
Reproduction
Logs
No response
System Info
Who can help?
@DN6@yiyixuxu@sayakpaul