docs: tweaks for outdated recommendations in template/main README.md (#…

…390)

* docs: explain how top n and models_to_skip interact

* fix: better model loading suggestions

* docs: more up to date settings recommendations

* style: fix

README.md

@@ -106,29 +106,35 @@ Tailor settings to your GPU, following these pointers:
 - **24GB+ VRAM** (e.g. 3090, 4090):
 
   ```yaml
+  - queue_size: 1 # <32GB RAM: 0, 32GB: 1, >32GB: 2
   - safety_on_gpu: true
+  - high_memory_mode: true
   - high_performance_mode: true
-  - post_process_job_overlap: true
   - unload_models_from_vram_often: false
-  - max_threads: 1 # If not using Flux/Cascade, otherwise 2 max
-  - queue_size: 2 # Or 3 if 64GB+ RAM
-  - max_batch: 8 # Or higher
+  - max_threads: 1 # 2 is often viable for xx90 cards
+  - post_process_job_overlap: true
+  - queue_size: 2 # Set to 1 if max_threads: 2
+  - max_power: 64 # Reduce if max_threads: 2
+  - max_batch: 8 # Increase if max_threads: 1, decrease if max_threads: 2
+  - allow_sdxl_controlnet: true
   ```
 
 - **12-16GB VRAM** (e.g. 3080 Ti, 4070 Ti, 4080):
 
   ```yaml
+  - queue_size: 1 # <32GB RAM: 0, 32GB: 1, >32GB: 2
   - safety_on_gpu: true # Consider false if using Cascade/Flux
   - moderate_performance_mode: true
   - unload_models_from_vram_often: false
   - max_threads: 1
+  - max_power: 50
   - max_batch: 4 # Or higher
   ```
 
 - **8-10GB VRAM** (e.g. 2080, 3060, 4060, 4060 Ti):
 
   ```yaml
-  - queue_size: 1 # Max, or only offer Flux
+  - queue_size: 1 # <32GB RAM: 0, 32GB: 1, >32GB: 2
   - safety_on_gpu: false
   - max_threads: 1
   - max_power: 32 # No higher
@@ -144,6 +150,11 @@ Tailor settings to your GPU, following these pointers:
   - `limit_max_steps: true` caps total steps per job based on model.
   - `preload_timeout: 120` allows longer model load times. Avoid misusing to prevent kudos loss or maintenance mode.
 
+- **Systems with less than 32GB of System RAM**:
+  - Be sure to only run SD15 models and queue_size: 0.
+    - Set `load_large_models: false`
+    - To your `models_to_skip` add `ALL SDXL`, `ALL SD21`, and the 'unpruned' models (see config) to prevent running out of memory
+
 ### Important Notes
 
 - Use an SSD, especially for multiple models. HDDs should offer one model only, loading within 60s.

bridgeData_template.yaml

@@ -250,18 +250,38 @@ disable_disk_cache: false # Currently unused in reGen
 #   "ALL INPAINTING MODELS" - All models marked for inpainting.
 #   "ALL SFW MODELS" - All models marked as SFW.
 #   "ALL NSFW MODELS" - All models marked as NSFW.
+#
 # The official model reference (in JSON format) is at https://github.com/Haidra-Org/AI-Horde-image-model-reference/blob/main/stable_diffusion.json.
 # The model name must match the name in the model reference or be a magic constant.
 # If you use `TOP` or `ALL` load commands, you should set a `civitai_api_token`. See that configuration entry for more information.
+#
+# SD15 models are generally 1-2GB each, SD21 models are generally 5-37B each, and SDXL models are generally 7B each.
+# Flux.1-Schnell fp8 (Compact) is 17GB, and Cascade is 14gb total.
+#
+#
+# Important note: models_to_skip are only removed and other models are *not* added back in.
+#                 If you set `TOP 10`, and add `Flux.1-Schnell fp8 (Compact)` to `models_to_skip`,
+#                 you will only have 9 models loaded.
+#
 models_to_load:
-  - "top 2"
-  #- "ALL MODELS"
-  #- "TOP 3"
-  #- "ALL SFW"
-  #- "Flux.1-Schnell fp8 (Compact)"
-  #- "stable_diffusion"
-  #- "Anything Diffusion"
-  #- "stable_diffusion_inpainting"
+    - "top 2"
+    #- "ALL MODELS"
+    #- "TOP 3"
+    #- "ALL SFW"
+    #- "Flux.1-Schnell fp8 (Compact)"
+    #- "stable_diffusion"
+    #- "Anything Diffusion"
+    #- "stable_diffusion_inpainting" # requires `allow_painting` to be true
+# <32gb RAM:
+  # Only SD15 models (set `load_large_models: false` and add SDXL, SD21, and all of the "Un-pruned" models to your `models_to_skip` below if using TOP/ALL commands)
+  # Be sure to set your queue_size to 0
+# 32gb+ RAM, 8GB VRAM:
+  # SD15, SD21
+  # If running SDXL
+  # If running Flux, be sure to free up as much VRAM as possible before starting the worker.
+# 32gb+ RAM, 10GB+ VRAM:
+  # All model types can be run.
+
 
 # If you use a meta command, such as ALL or TOP n, you can allow very large models, such as cascade or flux to be included.
 # By default, these models are excluded due to their large size.
@@ -271,15 +291,15 @@ load_large_models: false
 
 # Models to skip when `dynamic_models` is true or TOP n models are selected in models_to_load.
 # Avoid loading models due to VRAM constraints, NSFW content, or other reasons.
+# Models skipped are not replaced by other models, meaning TOP n will load fewer models. See the note above models_to_load.
 models_to_skip:
   - "pix2pix" # Not currently supported
   - "SDXL_beta::stability.ai#6901" # Do not remove this, as this model would never work
   - "A to Zovya RPG" # This model is known to cause problems with reGen
   # - "Stable Cascade 1.0" # This a *very* VRAM intensive model
   # - ALL NSFW MODELS
   #- "stable_diffusion_inpainting"  # Inpainting is generally quite heavy along with other models for smaller GPUs.
-  #- "stable_diffusion_2.1"  # Stable diffusion 2.1 has bigger memory requirements than 1.5, so if your card cannot lift, it, disable it
-  #- "stable_diffusion_2.0"  # Same as Stable diffusion 2.1
+  #- "ALL SD21"  # Stable diffusion 2.1 has bigger memory requirements than 1.5, so if your card cannot lift, it, disable it
   # - HASDX # Un-pruned SD1.5 model (5gb+)
   # - Anygen # Un-pruned SD1.5 model (5gb+)
   # - PFG # Un-pruned SD1.5 model (5gb+)