sd-webui: Pony from its single file; unload llama-server's models first

sd.cpp 929 detects an SDXL diffusers directory as SD 1.x and fails its
metadata validation (leejet/stable-diffusion.cpp#1938): all three of
imggen's SDXL directories do. Pony now loads from the single-file
checkpoint, which detects as SDXL; SDXL base and RealVis stay broken
until the upstream fix, noted beside them.

Before sd-server starts, every model llama-server's router holds is
unloaded: a resident Gemma left 1.1 GB of 12 GB free and the first unet
segment failed. Unloading is safe, the router loads it again on the next
chat request; a chat sent while sd-server runs does bring it back.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

1 files changed, 15 insertions, 2 deletions

diff --git a/bin/sd-webui b/bin/sd-webui
index 0d72547..ffc4c6f 100755
--- a/bin/sd-webui
+++ b/bin/sd-webui
@@ -57,16 +57,29 @@ case $(printf '%s\n' \
                     --t5xxl $M/umt5-xxl-encoder-Q8_0.gguf --vae $M/vae/wan2.2_vae.safetensors
                     --cfg-scale 5.0 --steps 30 --flow-shift 5.0 --diffusion-fa
                     -W 1280 -H 704 --video-frames 49 --temporal-tiling) ;;
-  # imggen's diffusers weights, symlinked into the layout sd.cpp loads as a directory
+  # imggen's diffusers weights, symlinked into the layout sd.cpp loads as a directory.
+  # ponytail: broken in sd.cpp 929, which detects an SDXL diffusers directory as
+  # SD 1.x (leejet/stable-diffusion.cpp#1938); single-file checkpoints load fine.
   "SDXL base")  MODEL=(-m $M/sdxl-links/sdxl "${SDXL[@]}") ;;
   RealVisXL*)   MODEL=(-m $M/sdxl-links/realvis "${SDXL[@]}") ;;
-  Pony*)        MODEL=(-m $M/sdxl-links/pony "${SDXL[@]}" --clip-skip 2) ;;
+  # The single-file checkpoint, not imggen's directory, for the bug above.
+  Pony*)        MODEL=(-m $M/ponyDiffusionV6XL_v6StartWithThisOne.safetensors "${SDXL[@]}" --clip-skip 2) ;;
   *) exit 0 ;;
 esac
 
 PORT=${PORT:-7860}
 URL=http://127.0.0.1:$PORT/
 
+# Free the GPU: a model llama-server's router holds (Gemma for the assistant)
+# left Pony 1.1 GB of 12 GB and failed the first step. Unloading is safe, the
+# router loads it again on the next chat request.
+# ponytail: one sweep at start; a chat sent while sd-server runs reloads it.
+LLAMA=${LLAMA_URL:-http://127.0.0.1:8181}
+for id in $(curl -sf -m 5 "$LLAMA/models" | jq -r '.data[] | select(.status.value != "unloaded") | .id'); do
+  curl -sf -m 30 -X POST "$LLAMA/models/unload" -H 'Content-Type: application/json' \
+    -d "$(jq -nc --arg m "$id" '{model: $m}')" >/dev/null && echo "sd-webui: unloaded $id from llama-server"
+done
+
 # Open the browser once the server answers (model loading takes a while).
 (
   until curl -sf -o /dev/null "$URL"; do sleep 1; done