chore: Add example TRTLLM configs for Deepseek R1 (GB200) (#1099)

examples/tensorrt_llm/configs/deepseek_r1/agg.yaml

+# 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:
+  # This is the client-facing model name, you can set this to anything you'd like.
+  served_model_name: "nvidia/DeepSeek-R1-FP4"
+  endpoint_chat: dynamo.Processor.chat/completions
+  endpoint_completions: dynamo.Processor.completions
+  port: 8000
+
+Processor:
+  engine_args: "configs/deepseek_r1/agg_llm_api_config.yaml"
+  router: round-robin
+  # Parallelize preprocessing/tokenization to avoid bottlenecks
+  ServiceArgs:
+    workers: 5
+
+TensorRTLLMWorker:
+  engine_args: "configs/deepseek_r1/agg_llm_api_config.yaml"
+  router: round-robin
+  ServiceArgs:
+    workers: 1
+    resources:
+      gpu: 4

examples/tensorrt_llm/configs/deepseek_r1/agg_llm_api_config.yaml

+# 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
+
+# TP/EP/PP/DP
+tensor_parallel_size: 4
+moe_expert_parallel_size: 4
+pipeline_parallel_size: 1
+enable_attention_dp: false
+
+max_batch_size: 256
+# 8448 = 8192 ISL + 256 OSL
+max_num_tokens: 8448
+max_seq_len: 8448
+
+kv_cache_config:
+  # With dp attention disabled: high free_gpu_memory_fraction is fine.
+  free_gpu_memory_fraction: 0.85
+  # With dp attention enabled: large ISL at high concurrency may need
+  # free_gpu_memory_fraction low to have enough available memory.
+  # free_gpu_memory_fraction: 0.30
+
+pytorch_backend_config:
+  use_cuda_graph: true
+  cuda_graph_padding_enabled: true
+  # NOTE: For larger max batch size, you may want to add larger cuda graph
+  # batch sizes below to match.
+  cuda_graph_batch_sizes:
+  - 1
+  - 2
+  - 4
+  - 8
+  - 16
+  - 32
+  - 64
+  - 128
+  - 256
+  print_iter_log: true
+  enable_overlap_scheduler: true
+  kv_cache_dtype: fp8

examples/tensorrt_llm/configs/deepseek_r1/disagg.yaml

+# 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:
+  # This is the client-facing model name, you can set this to anything you'd like.
+  served_model_name: "nvidia/DeepSeek-R1-FP4"
+  endpoint_chat: dynamo.Processor.chat/completions
+  endpoint_completions: dynamo.Processor.completions
+  port: 8000
+
+Processor:
+  engine_args: "configs/deepseek_r1/agg_llm_api_config.yaml"
+  router: round-robin
+  remote-prefill: true
+  # Parallelize preprocessing/tokenization to avoid bottlenecks
+  ServiceArgs:
+    workers: 5
+
+TensorRTLLMWorker:
+  engine_args: "configs/deepseek_r1/agg_llm_api_config.yaml"
+  llmapi-disaggregated-config: "configs/deepseek_r1/disagg_llm_api_config.yaml"
+  remote-prefill: true
+  # NOTE: When testing/benchmarking multiple prefill workers, you can set
+  # this number to the exact amount of prefill workers if you want Dynamo to
+  # wait until all the prefill workers are ready before marking the decode
+  # worker ready.
+  min-prefill-workers: 1
+  router: round-robin
+  ServiceArgs:
+    workers: 1
+    resources:
+      gpu: 4
+
+TensorRTLLMPrefillWorker:
+  engine_args: "configs/deepseek_r1/agg_llm_api_config.yaml"
+  llmapi-disaggregated-config: "configs/deepseek_r1/disagg_llm_api_config.yaml"
+  router: round-robin
+  ServiceArgs:
+    workers: 1
+    resources:
+      gpu: 4

examples/tensorrt_llm/configs/deepseek_r1/disagg_llm_api_config.yaml

+# 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.
+
+# Example Configs for Context & Generation on GB200 nodes
+# - Context on 1xGB200 (4xB00)
+# - Generation on 1xGB200 (4xB200)
+
+# NOTE: Fields like hostname, ports, urls, num_instances, etc. only used by trtllm-serve, not by dynamo
+
+backend: pytorch
+
+context_servers:
+  # Context/prefill processes many tokens at once, so for a large ISL, a large
+  # batch size may not be needed to saturate GPU utilization.
+  max_batch_size: 1
+  max_num_tokens: 8192
+  max_seq_len: 8192
+
+  # TP/EP/PP/DP
+  tensor_parallel_size: 4
+  moe_expert_parallel_size: 4
+  pipeline_parallel_size: 1
+  enable_attention_dp: true
+
+  free_gpu_memory_fraction: 0.75
+  pytorch_backend_config:
+    print_iter_log: true
+    # NOTE: This dtype must match in both context/generation configs
+    kv_cache_dtype: fp8
+
+generation_servers:
+  # Generation/decode processes one token per request at a time, so a larger
+  # batch size helps to saturate GPU utilization.
+  max_batch_size: 256
+  max_num_tokens: 256
+  # 8448 = 8192 ISL + 256 OSL
+  max_seq_len: 8448
+
+  # TP/EP/PP/DP
+  tensor_parallel_size: 4
+  moe_expert_parallel_size: 4
+  pipeline_parallel_size: 1
+  enable_attention_dp: false
+
+  # With dp attention disabled: high free_gpu_memory_fraction is fine.
+  free_gpu_memory_fraction: 0.85
+  # With dp attention enabled: large ISL at high concurrency may need
+  # free_gpu_memory_fraction low to have enough available memory.
+  # free_gpu_memory_fraction: 0.30
+
+  pytorch_backend_config:
+    use_cuda_graph: true
+    cuda_graph_padding_enabled: true
+    # NOTE: For larger max batch size, you may want to add larger cuda graph
+    # batch sizes below to match.
+    cuda_graph_batch_sizes:
+    - 1
+    - 2
+    - 4
+    - 8
+    - 16
+    - 32
+    - 64
+    - 128
+    - 256
+    print_iter_log: true
+    enable_overlap_scheduler: true
+    # NOTE: This dtype must match in both context/generation configs
+    kv_cache_dtype: fp8