Test ordered one-slot DFlash2 placements and preserve results

2026-09-20 21:10:08 +02:00
parent 3d5931146b
commit 4007242709
25 changed files with 360 additions and 5 deletions
@@ -0,0 +1,19 @@
+[
+  {
+    "label": "medium-slot1-priority3-draft-cuda0",
+    "model": "pure",
+    "ctx": 160000,
+    "single": false,
+    "split": "70,30",
+    "ubatch": 128,
+    "parallel": 1,
+    "draft_max": 7,
+    "draft_kv": "q4_0",
+    "draft_device": "CUDA0",
+    "quality": false,
+    "prompts": [
+      4096
+    ],
+    "minimum_headroom_mib": 512
+  }
+]