Merge branch 'main' into add-hacky-deployments-script

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

@@ -46,6 +46,33 @@ JSON
     exit 1
 }
 
+function install_vastai_sdk() {
+    # 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/ @ ${SDK_BRANCH}"
+        if ! uv pip install "vastai-sdk @ git+https://github.com/vast-ai/vast-sdk.git@${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 ${SDK_VERSION}"
+        if ! uv pip install "vastai-sdk==${SDK_VERSION}"; then
+            report_error_and_exit "Failed to install vastai-sdk==${SDK_VERSION}"
+        fi
+        return 0
+    fi
+
+    echo "Installing default vastai-sdk"
+    if ! uv pip install vastai-sdk; then
+        report_error_and_exit "Failed to install vastai-sdk"
+    fi
+}
+
 [ -n "$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 +90,8 @@ 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 +151,8 @@ 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
@@ -184,6 +214,80 @@ fi
 
 export REPORT_ADDR WORKER_PORT USE_SSL UNSECURED
 
+# ─── SDK Deployment Mode ───────────────────────────────────────────────
+if [ "$IS_DEPLOYMENT" = "true" ]; then
+    echo "=== SDK Deployment Mode ==="
+    echo "DEPLOYMENT_ID: $DEPLOYMENT_ID"
+
+    DEPLOY_DIR="/workspace/deployment"
+    mkdir -p "$DEPLOY_DIR"
+
+    VAST_API_BASE="${VAST_API_BASE:-https://console.vast.ai}"
+
+    # Download deployment code, retrying until the blob is available on S3.
+    # The s3_key exists in the DB as soon as the deployment is created, but the
+    # actual upload may still be in flight from the client side.
+    echo "Downloading deployment code..."
+    RETRY=0
+    while true; do
+        DOWNLOAD_RESPONSE=$(curl -sS \
+            -H "Authorization: Bearer $CONTAINER_API_KEY" \
+            "${VAST_API_BASE}/api/v0/deployment/${DEPLOYMENT_ID}/download_url/")
+        DOWNLOAD_URL=$(python3 -c "
+import sys, json
+try:
+    d = json.load(sys.stdin)
+    print(d.get('download_url') or '')
+except: print('')
+" <<< "$DOWNLOAD_RESPONSE")
+
+        if [ -z "$DOWNLOAD_URL" ] || [ "$DOWNLOAD_URL" = "None" ]; then
+            RETRY=$((RETRY + 1))
+            echo "No download URL yet (attempt $RETRY), retrying in 10s... response: $DOWNLOAD_RESPONSE"
+            sleep 10
+            continue
+        fi
+
+        # Got a URL — try the actual S3 download
+        HTTP_CODE=$(curl -sS -L -o "$DEPLOY_DIR/deployment.tar.gz" -w "%{http_code}" "$DOWNLOAD_URL")
+        if [ "$HTTP_CODE" = "200" ]; then
+            break
+        fi
+
+        RETRY=$((RETRY + 1))
+        echo "S3 download returned HTTP $HTTP_CODE (attempt $RETRY), blob not yet uploaded. Retrying in 10s..."
+        rm -f "$DEPLOY_DIR/deployment.tar.gz"
+        sleep 10
+    done
+
+    cd "$DEPLOY_DIR" && tar xzf deployment.tar.gz
+    echo "Deployment code extracted."
+
+    # Source secrets if present
+    if [ -f "$DEPLOY_DIR/.secrets" ]; then
+        echo "Sourcing secrets..."
+        source "$DEPLOY_DIR/.secrets"
+    fi
+
+    # Run on_start.sh to completion if present
+    if [ -f "$DEPLOY_DIR/on_start.sh" ]; then
+        echo "Running on_start.sh..."
+        chmod +x "$DEPLOY_DIR/on_start.sh"
+        bash "$DEPLOY_DIR/on_start.sh"
+        echo "on_start.sh completed."
+    fi
+
+    # Install SDK (uses the install_vastai_sdk function which supports SDK_BRANCH/SDK_VERSION)
+    install_vastai_sdk
+
+    # Run deployment in serve mode
+    export VAST_DEPLOYMENT_MODE=serve
+    echo "Starting deployment: python3 $DEPLOY_DIR/deployment.py"
+    python3 "$DEPLOY_DIR/deployment.py"
+    exit $?
+fi
+# ─── End SDK Deployment Mode ───────────────────────────────────────────
+
 if ! cd "$SERVER_DIR"; then
     report_error_and_exit "Failed to cd into SERVER_DIR: $SERVER_DIR"
 fi

workers/openai/worker.py

@@ -60,20 +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,
-            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"]
         )
     ],