docs: TRTLLM Example of Llama4+Eagle3 (Speculative Decoding)

examples/tensorrt_llm/configs/llama4/eagle/eagle_agg.yaml

@@ -0,0 +1,31 @@
+# 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: "meta-llama/Llama-4-Maverick-17B-128E"
+  endpoint: dynamo.TensorRTLLMWorker.generate
+  port: 8000
+  router: round-robin
+
+TensorRTLLMWorker:
+  served_model_name: "meta-llama/Llama-4-Maverick-17B-128E"
+  model-path: "meta-llama/Llama-4-Maverick-17B-128E"
+  extra-engine-args: "configs/llama4/eagle/engine_configs/agg_config.yaml"
+  router: round-robin
+  ServiceArgs:
+    workers: 1
+    resources:
+      gpu: 8

examples/tensorrt_llm/configs/llama4/eagle/eagle_disagg.yaml

@@ -0,0 +1,48 @@
+# 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: "meta-llama/Llama-4-Maverick-17B-128E"
+  endpoint: dynamo.TensorRTLLMWorker.generate
+  port: 8000
+  router: round-robin
+
+TensorRTLLMWorker:
+  served_model_name: "meta-llama/Llama-4-Maverick-17B-128E"
+  model-path: "meta-llama/Llama-4-Maverick-17B-128E"
+  # Path to a YAML file containing additional keyword arguments to pass to the TRTLLM engine.
+  # The fields in `extra-engine-args` holds higher priority than the above TRTLLM engine fields.
+  extra-engine-args: "configs/llama4/eagle/engine_configs/decode_config.yaml"
+  router: round-robin
+  enable-disagg: true
+  ServiceArgs:
+    workers: 1
+    resources:
+      gpu: 8
+
+TensorRTLLMPrefillWorker:
+  # 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-path: "meta-llama/Llama-4-Maverick-17B-128E"
+  # Path to a YAML file containing additional keyword arguments to pass to the TRTLLM engine.
+  # The fields in `extra-engine-args` holds higher priority than the above TRTLLM engine fields.
+  extra-engine-args: "configs/llama4/eagle/engine_configs/prefill_config.yaml"
+  router: round-robin
+  ServiceArgs:
+    workers: 1
+    resources:
+      gpu: 8

examples/tensorrt_llm/configs/llama4/eagle/engine_configs/agg_config.yaml

@@ -0,0 +1,55 @@
+# 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.
+
+backend: pytorch
+tensor_parallel_size: 4
+moe_expert_parallel_size: 1
+max_batch_size: 256
+# 8448 = 8192 ISL + 256 OSL
+max_num_tokens: 8448
+max_seq_len: 8448
+
+# Enable Speculative Decoding in the model engine
+speculative_config:
+  decoding_type: Eagle
+  max_draft_len: 1
+  pytorch_weights_path: nvidia/Llama-4-Maverick-17B-128E-Eagle3
+  eagle3_one_model: False
+
+kv_cache_config:
+  free_gpu_memory_fraction: 0.5
+  enable_block_reuse: false  
+
+disable_overlap_scheduler: true
+
+use_cuda_graph: true
+cuda_graph_padding_enabled: true
+cuda_graph_batch_sizes:
+- 1
+- 2
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+print_iter_log: true
+kv_cache_dtype: fp8

examples/tensorrt_llm/configs/llama4/eagle/engine_configs/decode_config.yaml

@@ -0,0 +1,57 @@
+# 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.
+
+backend: pytorch
+tensor_parallel_size: 4
+moe_expert_parallel_size: 4
+max_batch_size: 256
+# Note: When MPT is enabled and `cuda_graph_batch_sizes` is specified, `max_num_tokens` must satisfy the following formula:
+# max_num_tokens >= max(cuda_graph_batch_sizes) * (num_nextn_predict_layers + 1)
+# This is a known issue in TensorRT-LLM and will be resolved in the next release.
+max_num_tokens: 512
+# 8704 = 8192 ISL + 512 OSL
+max_seq_len: 8704
+disable_overlap_scheduler: true
+
+# Enable Speculative Decoding in the model engine
+speculative_config:
+  decoding_type: Eagle
+  max_draft_len: 1
+  pytorch_weights_path: nvidia/Llama-4-Maverick-17B-128E-Eagle3
+  eagle3_one_model: False
+
+kv_cache_config:
+  free_gpu_memory_fraction: 0.5
+  enable_block_reuse: false  
+
+use_cuda_graph: true
+cuda_graph_padding_enabled: true
+cuda_graph_batch_sizes:
+- 1
+- 2
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+print_iter_log: true
+kv_cache_dtype: fp8

examples/tensorrt_llm/configs/llama4/eagle/engine_configs/prefill_config.yaml

@@ -0,0 +1,40 @@
+# 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.
+
+backend: pytorch
+tensor_parallel_size: 4
+moe_expert_parallel_size: 4
+max_batch_size: 1
+max_num_tokens: 8192
+max_seq_len: 8192
+print_iter_log: true
+kv_cache_dtype: fp8
+disable_overlap_scheduler: true
+
+# Enable Speculative Decoding in the model engine
+speculative_config:
+  decoding_type: Eagle
+  max_draft_len: 1
+  pytorch_weights_path: nvidia/Llama-4-Maverick-17B-128E-Eagle3
+  eagle3_one_model: False
+
+kv_cache_config:
+  free_gpu_memory_fraction: 0.5
+  enable_block_reuse: false