Add new start_server.sh

Updated requirements to only require vastai-sdk

requirements.txt

@@ -1,13 +1 @@
-aiohttp==3.10.1
-aiodns~=3.6.0
-pycares~=4.11.0
-anyio~=4.4
-lib~=4.0
-nltk~=3.9
-psutil~=6.0
-pycryptodome~=3.20
-Requests~=2.32
-transformers~=4.52
-utils==1.0.*
-hf_transfer>=0.1.9
 vastai-sdk>=0.3.0

start_server.sh

@@ -2,6 +2,13 @@
 
 set -e -o pipefail
 
+# Check for force update flag
+FORCE_UPDATE=false
+if [ -f "/.force_update" ]; then
+    echo "Force update flag detected at /.force_update"
+    FORCE_UPDATE=true
+fi
+
 WORKSPACE_DIR="${WORKSPACE_DIR:-/workspace}"
 
 SERVER_DIR="$WORKSPACE_DIR/vast-pyworker"
@@ -46,6 +53,39 @@ JSON
     exit 1
 }
 
+function install_vastai_sdk() {
+    local force_flag=""
+    if [ "$FORCE_UPDATE" = true ]; then
+        force_flag="--force-reinstall"
+        echo "Force reinstalling vastai-sdk"
+    fi
+
+    # If SDK_BRANCH is set, install vastai-sdk from the vast-sdk repo at that branch/tag/commit.
+    if [ -n "${SDK_BRANCH:-}" ]; then
+        if [ -n "${SDK_VERSION:-}" ]; then
+            echo "WARNING: Both SDK_BRANCH and SDK_VERSION are set; using SDK_BRANCH=${SDK_BRANCH}"
+        fi
+        echo "Installing vastai-sdk from https://github.com/vast-ai/vast-sdk/ @ ${SDK_BRANCH}"
+        if ! uv pip install $force_flag "vastai-sdk @ git+https://github.com/vast-ai/vast-sdk.git@${SDK_BRANCH}"; then
+            report_error_and_exit "Failed to install vastai-sdk from vast-ai/vast-sdk@${SDK_BRANCH}"
+        fi
+        return 0
+    fi
+
+    if [ -n "${SDK_VERSION:-}" ]; then
+        echo "Installing vastai-sdk version ${SDK_VERSION}"
+        if ! uv pip install $force_flag "vastai-sdk==${SDK_VERSION}"; then
+            report_error_and_exit "Failed to install vastai-sdk==${SDK_VERSION}"
+        fi
+        return 0
+    fi
+
+    echo "Installing default vastai-sdk"
+    if ! uv pip install $force_flag vastai-sdk; then
+        report_error_and_exit "Failed to install vastai-sdk"
+    fi
+}
+
 [ -n "$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 +103,8 @@ 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" >> "$MODEL_LOG.old"; then
         report_error_and_exit "Failed to rotate model log"
@@ -104,10 +145,27 @@ then
         if ! git clone "${PYWORKER_REPO:-https://github.com/vast-ai/pyworker}" "$SERVER_DIR"; then
             report_error_and_exit "Failed to clone pyworker repository"
         fi
+    elif [ "$FORCE_UPDATE" = true ]; then
+        echo "Force updating pyworker repository"
+        if ! (cd "$SERVER_DIR" && git fetch --all); then
+            report_error_and_exit "Failed to fetch pyworker repository updates"
+        fi
     fi
     if [[ -n ${PYWORKER_REF:-} ]]; then
-        if ! (cd "$SERVER_DIR" && git checkout "$PYWORKER_REF"); then
+        if [ "$FORCE_UPDATE" = true ]; then
-            report_error_and_exit "Failed to checkout pyworker reference: $PYWORKER_REF"
+            echo "Force updating to pyworker reference: $PYWORKER_REF"
+            if ! (cd "$SERVER_DIR" && git fetch --all && git checkout "$PYWORKER_REF" && git pull); then
+                report_error_and_exit "Failed to force update pyworker reference: $PYWORKER_REF"
+            fi
+        else
+            if ! (cd "$SERVER_DIR" && git checkout "$PYWORKER_REF"); then
+                report_error_and_exit "Failed to checkout pyworker reference: $PYWORKER_REF"
+            fi
+        fi
+    elif [ "$FORCE_UPDATE" = true ]; then
+        echo "Force updating pyworker to latest"
+        if ! (cd "$SERVER_DIR" && git pull); then
+            report_error_and_exit "Failed to pull latest pyworker changes"
         fi
     fi
 
@@ -123,6 +181,8 @@ 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
@@ -137,6 +197,44 @@ else
     fi
     echo "environment activated"
     echo "venv: $VIRTUAL_ENV"
+
+    # Handle force update for existing environment
+    if [ "$FORCE_UPDATE" = true ]; then
+        echo "Performing force update on existing environment"
+
+        # Update pyworker repository
+        if [[ -d $SERVER_DIR ]]; then
+            echo "Force updating pyworker repository"
+            if ! (cd "$SERVER_DIR" && git fetch --all); then
+                report_error_and_exit "Failed to fetch pyworker repository updates"
+            fi
+
+            if [[ -n ${PYWORKER_REF:-} ]]; then
+                echo "Force updating to pyworker reference: $PYWORKER_REF"
+                if ! (cd "$SERVER_DIR" && git checkout "$PYWORKER_REF" && git pull); then
+                    report_error_and_exit "Failed to force update pyworker reference: $PYWORKER_REF"
+                fi
+            else
+                echo "Force updating pyworker to latest"
+                if ! (cd "$SERVER_DIR" && git pull); then
+                    report_error_and_exit "Failed to pull latest pyworker changes"
+                fi
+            fi
+        fi
+
+        # Force reinstall SDK
+        install_vastai_sdk
+    fi
+fi
+
+# Remove force update flag after successful update
+if [ "$FORCE_UPDATE" = true ]; then
+    echo "Removing force update flag"
+    if ! rm -f "/.force_update"; then
+        echo "WARNING: Failed to remove /.force_update, continuing anyway"
+    else
+        echo "Force update completed successfully"
+    fi
 fi
 
 if [ "$USE_SSL" = true ]; then

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,
+                concurrency=10,
-                runs=2
+                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"]
         )
     ],