flag for model log rotate

* Support old vLLM payloads

start_server.sh

@@ -46,6 +46,21 @@ JSON
     exit 1
 }
 
+function install_vastai_sdk() {
+    if [ -n "${SDK_VERSION:-}" ]; then
+        echo "Installing vastai-sdk version ${SDK_VERSION}"
+        if ! uv pip install "vastai-sdk==${SDK_VERSION}"; then
+            report_error_and_exit "Failed to install vastai-sdk==${SDK_VERSION}"
+        fi
+    else
+        echo "Installing default vastai-sdk"
+        if ! uv pip install vastai-sdk; then
+            report_error_and_exit "Failed to install vastai-sdk"
+        fi
+    fi
+}
+
+
 [ -n "$BACKEND" ] && [ -z "$HF_TOKEN" ] && report_error_and_exit "HF_TOKEN must be set when BACKEND is set!"
 [ -z "$CONTAINER_ID" ] && report_error_and_exit "CONTAINER_ID must be set!"
 [ "$BACKEND" = "comfyui" ] && [ -z "$COMFY_MODEL" ] && report_error_and_exit "For comfyui backends, COMFY_MODEL must be set!"
@@ -63,7 +78,8 @@ echo_var DEBUG_LOG
 echo_var PYWORKER_LOG
 echo_var MODEL_LOG
 
-if [ -e "$MODEL_LOG" ]; then
+ROTATE_MODEL_LOG="${ROTATE_MODEL_LOG:-false}"
+if [ "$ROTATE_MODEL_LOG" = "true" ] && [ -e "$MODEL_LOG" ]; then
     echo "Rotating model log at $MODEL_LOG to $MODEL_LOG.old"
     if ! cat "$MODEL_LOG" >> "$MODEL_LOG.old"; then
         report_error_and_exit "Failed to rotate model log"
@@ -123,6 +139,8 @@ then
         report_error_and_exit "Failed to install Python requirements"
     fi
 
+    install_vastai_sdk
+
     if ! touch ~/.no_auto_tmux; then
         report_error_and_exit "Failed to create ~/.no_auto_tmux"
     fi

workers/openai/worker.py

@@ -28,6 +28,12 @@ MODEL_INFO_LOG_MSGS = [
 nltk.download("words")
 WORD_LIST = nltk.corpus.words.words()
 
+def request_parser(request):
+    data = request
+    if request.get("input") is not None:
+        data = request.get("input")
+    return data
+
 
 def completions_benchmark_generator() -> dict:
     prompt = " ".join(random.choices(WORD_LIST, k=int(250)))
@@ -54,18 +60,20 @@ worker_config = WorkerConfig(
             route="/v1/completions",
             workload_calculator= lambda data: data.get("max_tokens", 0),
             allow_parallel_requests=True,
-            max_queue_time=60.0,
+            request_parser=request_parser,
+            max_queue_time=600.0,
             benchmark_config=BenchmarkConfig(
                 generator=completions_benchmark_generator,
-                concurrency=100,
-                runs=2
+                concurrency=10,
+                runs=3
             )
         ),
         HandlerConfig(
             route="/v1/chat/completions",
             workload_calculator= lambda data: data.get("max_tokens", 0),
             allow_parallel_requests=True,
-            max_queue_time=60.0,
+            request_parser=request_parser,
+            max_queue_time=600.0,
         )
     ],
     log_action_config=LogActionConfig(

workers/tgi/worker.py

@@ -52,17 +52,18 @@ worker_config = WorkerConfig(
         HandlerConfig(
             route="/generate",
             allow_parallel_requests=True,
-            max_queue_time=60.0,
+            max_queue_time=600.0,
             benchmark_config=BenchmarkConfig(
                 generator=benchmark_generator,
-                concurrency=50
+                concurrency=10,
+                runs=3
             ),
             workload_calculator= lambda x: x["parameters"]["max_new_tokens"]
         ),
         HandlerConfig(
             route="/generate_stream",
             allow_parallel_requests=True,
-            max_queue_time=60.0,
+            max_queue_time=600.0,
             workload_calculator= lambda x: x["parameters"]["max_new_tokens"]
         )
     ],