Set default atten for ROCm to SDP

CHANGELOG.md

@@ -23,9 +23,10 @@
   - **Cross-attention** refactored cross-attention methods, thanks @Disty0  
     - for backend:original, its unchanged: SDP, xFormers, Doggettxs, InvokeAI, Sub-quadratic, Split attention  
     - for backend:diffuers, list is now: SDP, xFormers, Batch matrix-matrix, Split attention, Dynamic Attention BMM, Dynamic Attention SDP  
-      note: you may need to update your settings! if you were previously using split-attention, closest match is batch-matrix-matrix  
+      note: you may need to update your settings! Attention Slicing is renamed to Split attention  
+    - for ROCm, updated default cross-attention to Scaled Dot Product  
   - **Dynamic Attention Slicing**, thanks @Disty0  
-    - dynamically slices attention queries in order to save vram based on query size and slice rate in GB  
+    - dynamically slices attention queries in order to keep them under the slice rate  
       slicing gets only triggered if the query size is larger than the slice rate to gain performance  
       *Dynamic Attention Slicing BMM* uses *Batch matrix-matrix*  
       *Dynamic Attention Slicing SDP* uses *Scaled Dot Product*  

modules/sd_hijack_dynamic_atten.py

@@ -99,8 +99,10 @@ def sliced_scaled_dot_product_attention(query, key, value, attn_mask=None, dropo
 
 class DynamicAttnProcessorSDP:
     r"""
-    dynamically slices attention queries based on query size and slice rate in GB
+    dynamically slices attention queries in order to keep them under the slice rate
     slicing will not get triggered if the query size is smaller than the slice rate to gain performance
+
+    slice rate is in GB
     based on AttnProcessor V2
     """
 
@@ -181,8 +183,10 @@ class DynamicAttnProcessorSDP:
 
 class DynamicAttnProcessorBMM:
     r"""
-    dynamically slices attention queries based on query size and slice rate in GB
+    dynamically slices attention queries in order to keep them under the slice rate
     slicing will not get triggered if the query size is smaller than the slice rate to gain performance
+
+    slice rate is in GB
     based on AttnProcessor V1
     """
 

modules/shared.py

@@ -318,9 +318,7 @@ elif devices.backend == "mps":
     cross_attention_optimization_default = "Scaled-Dot-Product" if backend == Backend.DIFFUSERS else "Doggettx's"
 elif devices.backend == "directml":
     cross_attention_optimization_default = "Dynamic Attention BMM" if backend == Backend.DIFFUSERS else "Sub-quadratic"
-elif devices.backend == "rocm":
-    cross_attention_optimization_default = "Dynamic Attention BMM" if backend == Backend.DIFFUSERS else "Sub-quadratic"
-else: # cuda and ipex
+else: # cuda, rocm, ipex
     cross_attention_optimization_default ="Scaled-Dot-Product"