File size: 6,915 Bytes
a82517c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 |
# 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.
#!/bin/bash
echo "Important Notes:"
echo "1: The max batch of engine size is set to 8 in the reference case. "
echo "2: The MIN_LEN/OPT_LEN/MAX_LEN for LLM, DiT, VLLN-VLSelfAttention models is set to 80/296/300."
echo "If your inference batch size exceeds 8 or the MIN_LEN/OPT_LEN/MAX_LEN for LLM, DiT, VLLN-VLSelfAttention not fit your use case, please set it to your actual batch size and length variables."
export PATH=/usr/src/tensorrt/bin:$PATH
VIDEO_VIEWS=${VIDEO_VIEWS:-1} # Options: 1 (default), 2
# Define length variables
# If only one video view, use these lengths; else (for two video views) use different lengths
if [ "$VIDEO_VIEWS" = "2" ]; then
# two video views
MIN_LEN=80
OPT_LEN=568
MAX_LEN=600
else # one video view (default)
MIN_LEN=80
OPT_LEN=296
MAX_LEN=300
fi
# Define precision settings (can be overridden via environment variables)
VIT_DTYPE=${VIT_DTYPE:-fp8} # Options: fp16, fp8
LLM_DTYPE=${LLM_DTYPE:-nvfp4} # Options: fp16, nvfp4, nvfp4_full, fp8
DIT_DTYPE=${DIT_DTYPE:-fp8} # Options: fp16, fp8
# Define max batch size (default 8, will be overridden for nvfp4 LLM variants)
MAX_BATCH=${MAX_BATCH:-8}
echo "Building TensorRT engines with the following precisions:"
echo " ViT: ${VIT_DTYPE}"
echo " LLM: ${LLM_DTYPE}"
echo " DiT: ${DIT_DTYPE}"
echo " Video Views: ${VIDEO_VIEWS}"
echo " MAX_BATCH: ${MAX_BATCH}"
echo " MIN_LEN: ${MIN_LEN}"
echo " OPT_LEN: ${OPT_LEN}"
echo " MAX_LEN: ${MAX_LEN}"
if [ ! -e /usr/src/tensorrt/bin/trtexec ]; then
echo "The file /usr/src/tensorrt/bin/trtexec does not exist. Please install tensorrt"
exit 1
fi
mkdir -p gr00t_engine
# VLLN-VLSelfAttention
echo "------------Building vlln_vl_self_attention Model--------------------"
trtexec --useCudaGraph --verbose --stronglyTyped --separateProfileRun --noDataTransfers --onnx=gr00t_onnx/action_head/vlln_vl_self_attention.onnx --saveEngine=gr00t_engine/vlln_vl_self_attention.engine --minShapes=backbone_features:1x${MIN_LEN}x2048 --optShapes=backbone_features:1x${OPT_LEN}x2048 --maxShapes=backbone_features:${MAX_BATCH}x${MAX_LEN}x2048 > gr00t_engine/vlln_vl_self_attention.log 2>&1
# DiT Model
echo "------------Building DiT Model (${DIT_DTYPE})--------------------"
trtexec --useCudaGraph --verbose --stronglyTyped --separateProfileRun --noDataTransfers --onnx=gr00t_onnx/action_head/DiT_${DIT_DTYPE}.onnx --saveEngine=gr00t_engine/DiT_${DIT_DTYPE}.engine --minShapes=sa_embs:1x49x1536,vl_embs:1x${MIN_LEN}x2048,timesteps_tensor:1 --optShapes=sa_embs:1x49x1536,vl_embs:1x${OPT_LEN}x2048,timesteps_tensor:1 --maxShapes=sa_embs:${MAX_BATCH}x49x1536,vl_embs:${MAX_BATCH}x${MAX_LEN}x2048,timesteps_tensor:${MAX_BATCH} > gr00t_engine/DiT_${DIT_DTYPE}.log 2>&1
# State Encoder
echo "------------Building State Encoder--------------------"
trtexec --useCudaGraph --verbose --stronglyTyped --separateProfileRun --noDataTransfers --onnx=gr00t_onnx/action_head/state_encoder.onnx --saveEngine=gr00t_engine/state_encoder.engine --minShapes=state:1x1x64,embodiment_id:1 --optShapes=state:1x1x64,embodiment_id:1 --maxShapes=state:${MAX_BATCH}x1x64,embodiment_id:${MAX_BATCH} > gr00t_engine/state_encoder.log 2>&1
# Action Encoder
echo "------------Building Action Encoder--------------------"
trtexec --useCudaGraph --verbose --stronglyTyped --separateProfileRun --noDataTransfers --onnx=gr00t_onnx/action_head/action_encoder.onnx --saveEngine=gr00t_engine/action_encoder.engine --minShapes=actions:1x16x32,timesteps_tensor:1,embodiment_id:1 --optShapes=actions:1x16x32,timesteps_tensor:1,embodiment_id:1 --maxShapes=actions:${MAX_BATCH}x16x32,timesteps_tensor:${MAX_BATCH},embodiment_id:${MAX_BATCH} > gr00t_engine/action_encoder.log 2>&1
# Action Decoder
echo "------------Building Action Decoder--------------------"
trtexec --useCudaGraph --verbose --stronglyTyped --separateProfileRun --noDataTransfers --onnx=gr00t_onnx/action_head/action_decoder.onnx --saveEngine=gr00t_engine/action_decoder.engine --minShapes=model_output:1x49x1024,embodiment_id:1 --optShapes=model_output:1x49x1024,embodiment_id:1 --maxShapes=model_output:${MAX_BATCH}x49x1024,embodiment_id:${MAX_BATCH} > gr00t_engine/action_decoder.log 2>&1
# VLM-ViT
echo "------------Building VLM-ViT (${VIT_DTYPE})--------------------"
trtexec --useCudaGraph --verbose --stronglyTyped --separateProfileRun --noDataTransfers --onnx=gr00t_onnx/eagle2/vit_${VIT_DTYPE}.onnx --saveEngine=gr00t_engine/vit_${VIT_DTYPE}.engine --minShapes=pixel_values:1x3x224x224,position_ids:1x256 --optShapes=pixel_values:${VIDEO_VIEWS}x3x224x224,position_ids:${VIDEO_VIEWS}x256 --maxShapes=pixel_values:${MAX_BATCH}x3x224x224,position_ids:${MAX_BATCH}x256 > gr00t_engine/vit_${VIT_DTYPE}.log 2>&1
# VLM-LLM
echo "------------Building VLM-LLM (${LLM_DTYPE})--------------------"
# Validate LLM_DTYPE
if [[ ! "$LLM_DTYPE" =~ ^(fp16|nvfp4|nvfp4_full|fp8)$ ]]; then
echo "Error: LLM_DTYPE must be 'fp16', 'nvfp4', 'nvfp4_full', or 'fp8', got '${LLM_DTYPE}'"
exit 1
fi
# Override max batch size for nvfp4 variants (require fixed shapes for Myelin)
if [[ "$LLM_DTYPE" =~ ^nvfp4 ]]; then
LLM_MAX_BATCH=1 # Fixed shapes (required for Myelin)
else
LLM_MAX_BATCH=${MAX_BATCH} # Use the default MAX_BATCH
fi
trtexec --useCudaGraph --verbose --stronglyTyped --separateProfileRun --noDataTransfers \
--onnx=gr00t_onnx/eagle2/llm_${LLM_DTYPE}.onnx \
--saveEngine=gr00t_engine/llm_${LLM_DTYPE}.engine \
--minShapes=inputs_embeds:1x${MIN_LEN}x2048,attention_mask:1x${MIN_LEN} \
--optShapes=inputs_embeds:1x${OPT_LEN}x2048,attention_mask:1x${OPT_LEN} \
--maxShapes=inputs_embeds:${LLM_MAX_BATCH}x${MAX_LEN}x2048,attention_mask:${LLM_MAX_BATCH}x${MAX_LEN} \
> gr00t_engine/llm_${LLM_DTYPE}.log 2>&1
echo ""
echo "============================================================"
echo "TensorRT Engine Build Complete!"
echo "============================================================"
echo "Built engines with the following configurations:"
echo " ViT: ${VIT_DTYPE}"
echo " LLM: ${LLM_DTYPE}"
echo " DiT: ${DIT_DTYPE}"
echo ""
echo "Engines saved in: gr00t_engine/"
echo "Build logs saved in: gr00t_engine/*.log"
echo "============================================================"
|