Use Audio8 native compile batch policy

Dockerfile

@@ -13,7 +13,7 @@ RUN apt-get update && \
 WORKDIR /home/user/app
 
 ARG SGLANG_OMNI_REVISION=68a572348837f7b004857b4b07993c20ade4c017
-ARG AUDIO8_TTS_REVISION=5bdcbaddee56599483d7ecf422a9595dc5e5ab32
+ARG AUDIO8_TTS_REVISION=59ca6d27d441302eaa568bcb140b63c50a879859
 
 RUN git init /opt/sglang-omni && \
     git -C /opt/sglang-omni remote add origin https://github.com/sgl-project/sglang-omni.git && \
@@ -71,7 +71,6 @@ ENV HOME=/home/user \
     AUDIO8_TTS_MAX_RUNNING_REQUESTS=8 \
     AUDIO8_TTS_DISABLE_CUDA_GRAPH=0 \
     AUDIO8_TTS_ENABLE_TORCH_COMPILE=1 \
-    AUDIO8_TTS_TORCH_COMPILE_MAX_BS=1 \
     AUDIO8_TTS_VERIFY_FAST_KV=1 \
     FLASHINFER_WORKSPACE_BASE=/tmp/audio8-flashinfer \
     TORCHINDUCTOR_CACHE_DIR=/tmp/audio8-torchinductor \

README.md

@@ -17,8 +17,8 @@ The service uses the official Audio8 adapter with SGLang Omni 0.1.0 and
 SGLang 0.5.8.
 
 Reference audio uploaded through the web interface is limited to 30 seconds.
-The A10G runtime enables `torch.compile` for batch size 1 and uses CUDA Graph
-for larger batches.
+The A10G runtime enables `torch.compile` with SGLang's native batch policy and
+uses CUDA Graph without an Audio8-specific compile batch-size limit.
 
 ## API