Updated AMD ROCm (markdown)

AMD-ROCm.md

@@ -114,6 +114,26 @@ sudo docker run -it \
 > It will install the necessary libraries at the first run so it will take a while depending on your internet.  
 > Resulting docker image will use 3.2 GB disk space (uncompressed) for the docker image and 20 GB for the venv.
 
+## ROCm performance tuning on Linux
+
+### MIOpen database tuning 
+
+On first use, when using a resolution for the first time, or when upgrading Pytorch versions, ROCm runs a series of benchmarks to select the most efficient approach. This can lead to slow (up to 5-8 minutes) startup times, in particular if you use a refine pass at high resolution, but it only happens once per resolution (subsequent runs are much faster). If for any case this is undesirable, you can set the environment variable `MIOPEN_FIND_MODE` to `FAST`. This will reduce a lot the startup time on first use at the price of worse performance when generating.
+On the other hand, for best performance during generation (but slower startup on first use), you can set the variable `MIOPEN_FIND_ENFORCE` to `SEARCH`. 
+
+### Reduce VRAM consumption
+
+If you use the `bf16` data type (*Settings > Compute Settings > Execution Precision > Device precision type*), which is autodetected on RDNA3 and newer cards, there is the chance that VRAM usage will be very high (16+ GB) when decoding the final image and when upscaling with non-latent upscalers. To workaround the problem, ensure to set `Device precision type` as `fp16`, and disable VAE upcasting in *Variable Auto Encoder > VAE upcasting*. 
+Setting `fp16` has also a noticeable impact on performance. 
+
+### Composable Kernel (CK) Flash attention
+
+On RDNA3 and newer hardware, you can set the option to use CK Flash Attention for improved performance in *Compute Settings > Cross Attention > SDP Options* by toggling `CK Flash attention` and restarting SD.Next. Notice that enabling this option requires `rocm-hip-sdk` installed as it will download and compile an additional Python package from source on startup. 
+ 
+In case you want to install it manually, activate the virtual environment then run `pip`:
+    
+`pip install git+https://github.com/Disty0/flash-attention@navi_rotary_fix`
+
 ## ROCm Windows Support
 
 This is a guide to build rocBLAS based on the ROCm Official Documentations.
@@ -402,4 +422,4 @@ and tensile [tuning .tex](https://github.com/ROCm/Tensile/blob/develop/tuning_do
 A pdf version available in [here](https://github.com/likelovewant/ROCmLibs-for-gfx1103-AMD780M-APU/blob/main/tensile_tuning.pdf)
 
 Please feel welcome to edit this post and contribute optimized logic links. Remember to carefully consider the
-impact of any edits or additions.
+impact of any edits or additions.