From 9c10b0ab88cfae620df282ebf2dbac8ff308067a Mon Sep 17 00:00:00 2001 From: Mikei386 <44135113+Mikei386@users.noreply.github.com> Date: Sun, 30 Aug 2026 18:39:17 +0200 Subject: [PATCH] Tune Qwen medium prefill batching --- .env.example | 2 ++ compose.yaml | 4 ++-- config/install.env.example | 2 ++ 3 files changed, 6 insertions(+), 2 deletions(-) diff --git a/.env.example b/.env.example index 6664b35..7d9831b 100644 --- a/.env.example +++ b/.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 diff --git a/compose.yaml b/compose.yaml index c4308df..b8a9385 100644 --- a/compose.yaml +++ b/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 diff --git a/config/install.env.example b/config/install.env.example index 5422a90..07540cf 100644 --- a/config/install.env.example +++ b/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