feat: add unified backend architecture with DynamoBackend (#8003)

Signed-off-by: Tanmay Verma <tanmayv@nvidia.com>

examples/backends/trtllm/launch/agg.sh

@@ -20,6 +20,7 @@ export MODALITY=${MODALITY:-"text"}
 
 
 ENABLE_OTEL=false
+USE_UNIFIED=false
 EXTRA_ARGS=()
 while [[ $# -gt 0 ]]; do
     case $1 in
@@ -27,10 +28,15 @@ while [[ $# -gt 0 ]]; do
             ENABLE_OTEL=true
             shift
             ;;
+        --unified)
+            USE_UNIFIED=true
+            shift
+            ;;
         -h|--help)
             echo "Usage: $0 [OPTIONS]"
             echo "Options:"
             echo "  --enable-otel        Enable OpenTelemetry tracing"
+            echo "  --unified            Use unified_main entry point (Worker)"
             echo "  -h, --help           Show this help message"
             echo ""
             echo "Any additional options are passed through to dynamo.trtllm."
@@ -71,8 +77,12 @@ python3 -m dynamo.frontend &
 
 # run worker
 # Additional command line args can be passed
+WORKER_MODULE="dynamo.trtllm"
+if [ "$USE_UNIFIED" = true ]; then
+    WORKER_MODULE="dynamo.trtllm.unified_main"
+fi
 OTEL_SERVICE_NAME=dynamo-worker \
-python3 -m dynamo.trtllm \
+python3 -m "$WORKER_MODULE" \
   --model-path "$MODEL_PATH" \
   --served-model-name "$SERVED_MODEL_NAME" \
   --modality "$MODALITY" \

examples/backends/vllm/launch/agg.sh

@@ -13,6 +13,7 @@ source "$SCRIPT_DIR/../../../common/launch_utils.sh" # print_launch_banner, wait
 
 # Default model
 MODEL="Qwen/Qwen3-0.6B"
+USE_UNIFIED=false
 
 # Parse command line arguments
 EXTRA_ARGS=()
@@ -22,6 +23,20 @@ while [[ $# -gt 0 ]]; do
             MODEL="$2"
             shift 2
             ;;
+        --unified)
+            USE_UNIFIED=true
+            shift
+            ;;
+        -h|--help)
+            echo "Usage: $0 [OPTIONS]"
+            echo "Options:"
+            echo "  --model <name>       Specify model (default: $MODEL)"
+            echo "  --unified            Use unified_main entry point (Worker)"
+            echo "  -h, --help           Show this help message"
+            echo ""
+            echo "Any additional options are passed through to dynamo.vllm."
+            exit 0
+            ;;
         *)
             EXTRA_ARGS+=("$1")
             shift
@@ -47,8 +62,12 @@ python -m dynamo.frontend &
 
 # run worker
 # --enforce-eager is added for quick deployment. for production use, need to remove this flag
+WORKER_MODULE="dynamo.vllm"
+if [ "$USE_UNIFIED" = true ]; then
+    WORKER_MODULE="dynamo.vllm.unified_main"
+fi
 DYN_SYSTEM_PORT=${DYN_SYSTEM_PORT:-8081} \
-    python -m dynamo.vllm --model "$MODEL" --enforce-eager \
+    python -m "$WORKER_MODULE" --model "$MODEL" --enforce-eager \
     --max-model-len "$MAX_MODEL_LEN" \
     --max-num-seqs "$MAX_CONCURRENT_SEQS" \
     $GPU_MEM_ARGS \

lib/llm/src/protocols/common.rs

@@ -55,7 +55,7 @@ pub enum FinishReason {
     #[serde(rename = "error")]
     Error(String),
 
-    #[serde(rename = "cancelled")]
+    #[serde(rename = "cancelled", alias = "abort")]
     Cancelled,
 
     #[serde(rename = "content_filter")]
@@ -83,7 +83,7 @@ impl std::str::FromStr for FinishReason {
             "eos" => Ok(FinishReason::EoS),
             "length" => Ok(FinishReason::Length),
             "stop" => Ok(FinishReason::Stop),
-            "cancelled" => Ok(FinishReason::Cancelled),
+            "cancelled" | "abort" => Ok(FinishReason::Cancelled),
             s if s.starts_with("error: ") => Ok(FinishReason::Error(s[7..].to_string())),
             _ => Err(anyhow::anyhow!("Invalid FinishReason variant: '{}'", s)),
         }

tests/serve/test_sample.py

+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+import dataclasses
+import logging
+import os
+
+import pytest
+
+from tests.serve.common import (
+    WORKSPACE_DIR,
+    params_with_model_mark,
+    run_serve_deployment,
+)
+from tests.utils.constants import DefaultPort
+from tests.utils.engine_process import EngineConfig
+from tests.utils.payload_builder import chat_payload_default, completion_payload_default
+
+logger = logging.getLogger(__name__)
+
+sample_dir = os.path.join(WORKSPACE_DIR, "examples/backends/sample")
+
+sample_configs = {
+    "aggregated": EngineConfig(
+        name="aggregated",
+        directory=sample_dir,
+        script_name="agg.sh",
+        script_args=["--model-name", "Qwen/Qwen3-0.6B"],
+        marks=[
+            pytest.mark.gpu_0,
+            pytest.mark.timeout(300),
+            pytest.mark.pre_merge,
+        ],
+        model="Qwen/Qwen3-0.6B",
+        frontend_port=DefaultPort.FRONTEND.value,
+        request_payloads=[
+            chat_payload_default(),
+            completion_payload_default(),
+        ],
+    ),
+}
+
+
+@pytest.fixture(params=params_with_model_mark(sample_configs))
+def sample_config_test(request):
+    """Fixture that provides different sample test configurations"""
+    return sample_configs[request.param]
+
+
+@pytest.mark.e2e
+def test_sample_deployment(
+    sample_config_test,
+    request,
+    runtime_services_dynamic_ports,
+    dynamo_dynamic_ports,
+    predownload_models,
+):
+    """Test sample backend deployment using the unified Worker."""
+    config = dataclasses.replace(
+        sample_config_test, frontend_port=dynamo_dynamic_ports.frontend_port
+    )
+    run_serve_deployment(config, request, ports=dynamo_dynamic_ports)

tests/serve/test_sglang.py

@@ -89,6 +89,26 @@ sglang_configs = {
             metric_payload_default(min_num_requests=6, backend="sglang"),
         ],
     ),
+    "aggregated_unified": SGLangConfig(
+        name="aggregated_unified",
+        directory=sglang_dir,
+        script_name="agg.sh",
+        script_args=["--unified"],
+        marks=[
+            pytest.mark.gpu_1,
+            pytest.mark.profiled_vram_gib(3.7),
+            pytest.mark.requested_sglang_kv_tokens(96),
+            pytest.mark.timeout(195),
+            pytest.mark.pre_merge,
+        ],
+        model="Qwen/Qwen3-0.6B",
+        env={},
+        frontend_port=DefaultPort.FRONTEND.value,
+        request_payloads=[
+            chat_payload_default(),
+            completion_payload_default(),
+        ],
+    ),
     "disaggregated": SGLangConfig(
         name="disaggregated",
         directory=sglang_dir,

tests/serve/test_trtllm.py

@@ -100,6 +100,27 @@ trtllm_configs = {
             metric_payload_default(min_num_requests=6, backend="trtllm"),
         ],
     ),
+    "aggregated_unified": TRTLLMConfig(
+        name="aggregated_unified",
+        directory=trtllm_dir,
+        script_name="agg.sh",
+        script_args=["--unified"],
+        marks=[
+            pytest.mark.gpu_1,
+            pytest.mark.trtllm,
+            pytest.mark.profiled_vram_gib(3.9),
+            pytest.mark.requested_trtllm_kv_tokens(2592),
+            pytest.mark.timeout(300),
+            pytest.mark.pre_merge,
+        ],
+        model="Qwen/Qwen3-0.6B",
+        frontend_port=DefaultPort.FRONTEND.value,
+        delayed_start=5,
+        request_payloads=[
+            chat_payload_default(),
+            completion_payload_default(),
+        ],
+    ),
     "disaggregated": TRTLLMConfig(
         name="disaggregated",
         directory=trtllm_dir,

tests/serve/test_vllm.py

@@ -104,6 +104,24 @@ vllm_configs = {
             metric_payload_default(min_num_requests=6, backend="vllm"),
         ],
     ),
+    "aggregated_unified": VLLMConfig(
+        name="aggregated_unified",
+        directory=vllm_dir,
+        script_name="agg.sh",
+        script_args=["--unified"],
+        marks=[
+            pytest.mark.gpu_1,
+            pytest.mark.profiled_vram_gib(3.8),
+            pytest.mark.requested_vllm_kv_cache_bytes(1_119_388_000),
+            pytest.mark.timeout(360),
+            pytest.mark.pre_merge,
+        ],
+        model="Qwen/Qwen3-0.6B",
+        request_payloads=[
+            chat_payload_default(),
+            completion_payload_default(),
+        ],
+    ),
     "aggregated_logprobs": VLLMConfig(
         name="aggregated_logprobs",
         directory=vllm_dir,