Adapt to latest vllm changes

- Remove --eager-enforce on hpu to improve performance
- Refactor to the upstream docker entrypoint changes

Fixes issue #631.

Signed-off-by: Lianhao Lu <[email protected]>

Author: lianhao

helm-charts/common/agent/values.yaml

@@ -14,7 +14,7 @@ tgi:
 vllm:
   enabled: false
   LLM_MODEL_ID: "mistralai/Mistral-7B-Instruct-v0.3"
-  extraCmdArgs: ["/bin/bash", "-c", "python3 -m vllm.entrypoints.openai.api_server --enforce-eager --model mistralai/Mistral-7B-Instruct-v0.3 --tensor-parallel-size 1 --host 0.0.0.0 --port 2080 --download-dir /data --block-size 128 --max-num-seqs 4096 --max-seq_len-to-capture 8192 --enable-auto-tool-choice --tool-call-parser mistral"]
+  extraCmdArgs: ["--tensor-parallel-size", "1", "--block-size", "128", "--max-num-seqs", "4096", "--max-seq_len-to-capture", "8192", "--enable-auto-tool-choice", "--tool-call-parser", "mistral"]
 
 replicaCount: 1
 llm_endpoint_url: ""

helm-charts/common/llm-uservice/ci-vllm-gaudi-values.yaml

@@ -13,7 +13,7 @@ vllm:
     tag: "latest"
   LLM_MODEL_ID: Intel/neural-chat-7b-v3-3
   OMPI_MCA_btl_vader_single_copy_mechanism: none
-  extraCmdArgs: ["--enforce-eager","--tensor-parallel-size","1","--block-size","128","--max-num-seqs","256","--max-seq_len-to-capture","2048"]
+  extraCmdArgs: ["--tensor-parallel-size","1","--block-size","128","--max-num-seqs","256","--max-seq_len-to-capture","2048"]
   resources:
     limits:
       habana.ai/gaudi: 1

helm-charts/common/vllm/gaudi-values.yaml

@@ -13,9 +13,7 @@ image:
 
 # VLLM_CPU_KVCACHE_SPACE: "40"
 OMPI_MCA_btl_vader_single_copy_mechanism: none
-extraCmdArgs: ["--enforce-eager","--tensor-parallel-size","1","--block-size","128","--max-num-seqs","256","--max-seq_len-to-capture","2048"]
-# Workaround for current HPU image with start command /bin/bash
-# extraCmdArgs: ["/bin/bash","-c","python3 -m vllm.entrypoints.openai.api_server --enforce-eager --model Intel/neural-chat-7b-v3-3 --tensor-parallel-size 1 --host 0.0.0.0 --port 2080 --download-dir /data --block-size 128 --max-num-seqs 256 --max-seq_len-to-capture 2048"]
+extraCmdArgs: ["--tensor-parallel-size","1","--block-size","128","--max-num-seqs","256","--max-seq_len-to-capture","2048"]
 resources:
   limits:
     habana.ai/gaudi: 1

helm-charts/common/vllm/values.yaml

@@ -73,7 +73,7 @@ resources: {}
   #   cpu: 100m
   #   memory: 128Mi
 
-extraCmdArgs: ["--enforce-eager", "--dtype", "auto"]
+extraCmdArgs: []
 
 livenessProbe:
   httpGet: