Use conservative four-token MTP drafts

2026-08-25 04:20:15 -07:00
parent 3cf6353050
commit f86ed70c71
3 changed files with 7 additions and 8 deletions
@@ -55,7 +55,7 @@ CACHE_RAM=8192

 # Embedded-MTP (qwen35 NextN) qualified profile. The head is in the main GGUF,
 # so the draft runs with all layers on ROCm0 (no --model-draft).
-MTP_N_MAX=16
+MTP_N_MAX=4
 MTP_N_MIN=0
 MTP_P_MIN=0.60
 SPEC_MTP_STRICT_QWEN=on