Tune Qwen medium prefill batching

.env.example

@@ -23,6 +23,8 @@ VISION_PROJECTOR_FILE=qwen/mmproj-BF16.gguf
 
 FAST_CONTEXT=76800
 MEDIUM_CONTEXT=160000
+MEDIUM_BATCH_SIZE=2048
+MEDIUM_UBATCH_SIZE=128
 LARGE_CONTEXT=192000
 ULTRA_CONTEXT=262144
 UNCENSORED_CONTEXT=80000

compose.yaml

@@ -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

config/install.env.example

@@ -77,6 +77,8 @@ VISION_PROJECTOR_SHA256=83ee4f4f205fa514161778c41df1ea14144faa0f713510893b63c239
 
 FAST_CONTEXT=76800
 MEDIUM_CONTEXT=160000
+MEDIUM_BATCH_SIZE=2048
+MEDIUM_UBATCH_SIZE=128
 MEDIUM_TENSOR_SPLIT=90,10
 LARGE_CONTEXT=192000
 LARGE_TENSOR_SPLIT=86,14