diff --git a/Dockerfile b/Dockerfile index 44e0a21..250727a 100644 --- a/Dockerfile +++ b/Dockerfile @@ -45,5 +45,8 @@ RUN apt-get update && apt-get install -y git \ # Make sure we have the latest up to date chat template COPY glm_5.1_chat_template.jinja /opt/chat_template.jinja -# Working LMCache config -COPY lmcache-config.yaml /opt/lmcache-config.yaml \ No newline at end of file +# GLM 5.1 LMCache config +COPY lmcache-config-glm-51.yaml /opt/lmcache-config-glm-51.yaml + +# DEEPSEEK v4 LMCache config +COPY lmcache-config-dsv4.yaml /opt/lmcache-config-dsv4.yaml \ No newline at end of file diff --git a/lmcache-config-dsv4.yaml b/lmcache-config-dsv4.yaml new file mode 100644 index 0000000..6940107 --- /dev/null +++ b/lmcache-config-dsv4.yaml @@ -0,0 +1,10 @@ +chunk_size: 256 +local_cpu: true +max_local_cpu_size: 20.0 +remote_url: "redis://10.66.0.100:6379" +remote_serde: naive +save_decode_cache: true +use_layerwise: true +save_unfull_chunk: true +blocking_timeout_secs: 60 +cache_policy: LRU \ No newline at end of file diff --git a/lmcache-config.yaml b/lmcache-config-glm-51.yaml similarity index 75% rename from lmcache-config.yaml rename to lmcache-config-glm-51.yaml index 1229f3c..41958cf 100644 --- a/lmcache-config.yaml +++ b/lmcache-config-glm-51.yaml @@ -5,3 +5,5 @@ save_decode_cache: false enable_lazy_memory_allocator: true lazy_memory_initial_ratio: 1.0 use_gpu_connector_v3: true +remote_url: "redis://10.66.0.100:6379" +remote_serde: naive \ No newline at end of file