add 0.5.1 supported models (#2151)

Author: Qing Lan

tests/integration/llm/download_llm.py

@@ -12,9 +12,11 @@ def llm_download(model_id, token, allow_patterns):
             "*.json", "*.pt", "*.safetensors", "*.txt", "*.model", "*.tiktoken"
         ]
 
+    ignore_patterns = [".git*"]
     snapshot_download(repo_id=model_id,
                       local_dir=local_model_path,
                       allow_patterns=allow_patterns,
+                      ignore_patterns=ignore_patterns,
                       token=token)
 
 

tests/integration/llm/prepare.py

@@ -766,7 +766,7 @@
         "option.model_id": "s3://djl-llm/llama-2-70b-hf/",
     },
     "codellama-34b-aiccl": {
-        "option.model_id": "codellama/CodeLlama-34b-hf",
+        "option.model_id": "s3://djl-llm/CodeLlama-34b-Instruct-hf/",
     },
     "falcon-40b-aiccl": {
         "option.model_id": "s3://djl-llm/falcon-40b/",

wlm/src/main/java/ai/djl/serving/wlm/LmiConfigRecommender.java

@@ -27,7 +27,7 @@
 public final class LmiConfigRecommender {
 
     private static final Logger logger = LoggerFactory.getLogger(LmiConfigRecommender.class);
-    // TODO: model list is up to date with vLLM 0.4.2
+    // TODO: model list is up to date with vLLM 0.5.1
     private static final Map<String, String> MODEL_TO_ROLLING_BATCH =
             Map.ofEntries(
                     Map.entry("falcon", "lmi-dist"),
@@ -59,7 +59,16 @@ public final class LmiConfigRecommender {
                     Map.entry("qwen2_moe", "lmi-dist"),
                     Map.entry("stablelm", "lmi-dist"),
                     Map.entry("xverse", "lmi-dist"),
-                    Map.entry("starcoder2", "lmi-dist"));
+                    Map.entry("starcoder2", "lmi-dist"),
+                    // vllm 0.5.1
+                    Map.entry("arctic", "lmi-dist"),
+                    Map.entry("gemma2", "lmi-dist"),
+                    Map.entry("jamba", "lmi-dist"),
+                    Map.entry("phi3small", "lmi-dist"),
+                    Map.entry("llava", "lmi-dist"),
+                    Map.entry("llava_next", "lmi-dist"),
+                    Map.entry("paligemma", "lmi-dist"),
+                    Map.entry("phi3_v", "lmi-dist"));
 
     private static final Set<String> OPTIMIZED_TASK_ARCHITECTURES =
             Set.of("ForCausalLM", "LMHeadModel", "ForConditionalGeneration");