Tune Qwen medium prefill batching

2026-08-30 18:39:17 +02:00
parent 6af25cf30c
commit edb8b2868f
3 changed files with 6 additions and 2 deletions
@@ -181,9 +181,9 @@ services:
      - --threads-batch
      - "${LLAMA_THREADS_BATCH:-6}"
      - --batch-size
-      - "64"
+      - "${MEDIUM_BATCH_SIZE:-2048}"
      - --ubatch-size
-      - "32"
+      - "${MEDIUM_UBATCH_SIZE:-128}"
      - --parallel
      - "1"
      - --jinja