frank-default-small: ubatch=1024 (tight 24GB VRAM, 786k ctx pool)

scripts/frank-default-small.env

@@ -43,9 +43,9 @@ VERBOSITY=4
 # Store SWA/context checkpoints in RAM to hold peak GPU memory down.
 CTX_CHECKPOINTS=4
 
-# Big prefill batches (ubatch 2048, batch 4096).
+# Big prefill batches (ubatch 1024, batch 4096).
 BATCH_SIZE=4096
-UBATCH_SIZE=2048
+UBATCH_SIZE=1024
 
 # Eager RAM load (MoE experts on CPU; ~31 GB RAM).
 NO_MMAP=1