Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-18 08:44:33 +02:00
Code Issues Packages Projects Releases Wiki Activity

Rename ROCm Flash atten hijack to CK Flash atten and enable AOTriton memory and flash atten by default

Browse Source
This commit is contained in:
Disty0
2025-02-13 22:01:06 +03:00
parent 5ee6768eba
commit f94196bcd1
5 changed files with 36 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/shared.py
+2 -6
View File
@@ -452,14 +452,10 @@ def get_default_modes():
else: # cuda, rocm, zluda, ipex, openvino
default_cross_attention = "Scaled-Dot-Product"
if devices.backend == "rocm":
default_sdp_options = ['Memory attention', 'Math attention']
elif devices.backend == "zluda":
if devices.backend == "zluda":
default_sdp_options = ['Math attention', 'Dynamic attention']
else:
default_sdp_options = ['Flash attention', 'Memory attention', 'Math attention']
if (cmd_opts.lowvram or cmd_opts.medvram) and ('Flash attention' not in default_sdp_options and 'Dynamic attention' not in default_sdp_options):
default_sdp_options.append('Dynamic attention')
return default_offload_mode, default_diffusers_offload_min_gpu_memory, default_cross_attention, default_sdp_options
@@ -537,7 +533,7 @@ options_templates.update(options_section(('cuda', "Compute Settings"), {
"cross_attention_sep": OptionInfo("<h2>Cross Attention</h2>", "", gr.HTML),
"cross_attention_optimization": OptionInfo(startup_cross_attention, "Attention optimization method", gr.Radio, lambda: {"choices": shared_items.list_crossattention(native)}),
"sdp_options": OptionInfo(startup_sdp_options, "SDP options", gr.CheckboxGroup, {"choices": ['Flash attention', 'Memory attention', 'Math attention', 'Dynamic attention', 'Sage attention'], "visible": native}),
"sdp_options": OptionInfo(startup_sdp_options, "SDP options", gr.CheckboxGroup, {"choices": ['Flash attention', 'Memory attention', 'Math attention', 'Dynamic attention', 'CK Flash attention', 'Sage attention'], "visible": native}),
"xformers_options": OptionInfo(['Flash attention'], "xFormers options", gr.CheckboxGroup, {"choices": ['Flash attention'] }),
"dynamic_attention_slice_rate": OptionInfo(0.5, "Dynamic Attention slicing rate in GB", gr.Slider, {"minimum": 0.01, "maximum": max(gpu_memory,4), "step": 0.01, "visible": native}),
"dynamic_attention_trigger_rate": OptionInfo(1, "Dynamic Attention trigger rate in GB", gr.Slider, {"minimum": 0.01, "maximum": max(gpu_memory,4)*2, "step": 0.01, "visible": native}),