docs: add aggregated example turning on MTP with DeepSeek R1

examples/tensorrt_llm/README.md

@@ -118,6 +118,16 @@ cd /workspace/examples/tensorrt_llm
 dynamo serve graphs.disagg_router:Frontend -f ./configs/disagg_router.yaml
 ```
 
+#### Aggregated serving with Multi-Token Prediction(MTP) and DeepSeek R1
+```bash
+cd /workspace/examples/tensorrt_llm
+dynamo serve graphs.disagg_router:Frontend -f configs/deepseek_r1/mtp/mtp_agg.yaml
+```
+Notes:
+- There is a noticeable latency for the first two inference requests. Please send warm-up requests before starting the benchmark.
+- Please keep the `cuda_graph_padding_enabled` setting as `false` in the model engine's configuration. There is a known bug, and the fix will be included in the next release of TensorRT-LLM.
+- Disaggregated support for MTP in Dynamo + TensorRT-LLM is coming soon.
+
 #### Multi-Node Disaggregated Serving
 
 In the following example, we will demonstrate how to run a Disaggregated Serving

examples/tensorrt_llm/configs/deepseek_r1/mtp/mtp_agg.yaml

@@ -0,0 +1,29 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+Frontend:
+  served_model_name: "nvidia/DeepSeek-R1-FP4"
+  endpoint: dynamo.TensorRTLLMWorker.generate
+  port: 8000
+  router: round-robin
+
+TensorRTLLMWorker:
+  served_model_name: "nvidia/DeepSeek-R1-FP4"
+  engine_args: "configs/deepseek_r1/mtp/mtp_agg_llm_api_config.yaml"
+  router: round-robin
+  ServiceArgs:
+    workers: 1
+    resources:
+      gpu: 4

examples/tensorrt_llm/configs/deepseek_r1/mtp/mtp_agg_llm_api_config.yaml

@@ -0,0 +1,54 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# NOTE: FP4 only supported starting with Blackwell GPUs.
+# https://huggingface.co/nvidia/DeepSeek-R1-FP4
+# You can also specify the full path to locally downloaded weights
+# instead of a HuggingFace ID here.
+
+model_name: "nvidia/DeepSeek-R1-FP4"
+backend: pytorch
+tensor_parallel_size: 4
+moe_expert_parallel_size: 4
+enable_attention_dp: true
+max_batch_size: 256
+# 8448 = 8192 ISL + 256 OSL
+max_num_tokens: 8448
+max_seq_len: 8448
+kv_cache_config:
+  free_gpu_memory_fraction: 0.30
+
+# Enable the MTP(Multi-Token Prediction) in the model engine
+speculative_config:
+  decoding_type: MTP
+  num_nextn_predict_layers: 1
+
+use_cuda_graph: true
+# Please keep cuda_graph_padding_enabled setting as 'false' when MTP is turned on.
+# There is known bug with MTP and cuda_graph_padding_enabled.
+# Tensorrt LLM team is working on a fix in the next release.
+cuda_graph_padding_enabled: false
+cuda_graph_batch_sizes:
+- 1
+- 2
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+print_iter_log: true
+kv_cache_dtype: fp8