File size: 10,217 Bytes
e479c46 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 | #!/usr/bin/env python3
# SPDX-FileCopyrightText: Copyright (c) 2026 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.
"""Quick verification: compare PyTorch vs TRT action head outputs for N1.7."""
from dataclasses import dataclass
import os
import sys
import torch
from torch.nn.functional import cosine_similarity
import tyro
# Make sibling imports work
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from _trt_contract import resolve_batch_size
from export_onnx_n1d7 import prepare_observation
from gr00t.data.dataset.lerobot_episode_loader import LeRobotEpisodeLoader
from gr00t.data.embodiment_tags import EmbodimentTag
from gr00t.deployment.modes import VerifyMode
from gr00t.policy.gr00t_policy import Gr00tPolicy
@dataclass
class VerifyConfig:
"""Configuration for TRT verification."""
model_path: str
"""Path to model checkpoint (required)."""
dataset_path: str = "demo_data/libero_demo"
"""Path to dataset."""
engine_dir: str = "./gr00t_trt_deployment/engines"
"""Directory with TRT engines."""
mode: VerifyMode = VerifyMode.action_head
"""TRT setup mode. 'dit_only' loads only the DiT engine; 'vit_llm_only' keeps the action head in PyTorch."""
embodiment_tag: EmbodimentTag = EmbodimentTag.LIBERO_PANDA
"""Embodiment tag to use."""
batch_size: int = 1
"""Batch size for TRT inference. If > 1, tiles the observation and takes slice [0] for comparison."""
def _tile_observation(obs, n):
"""Tile a single observation dict to batch size n."""
tiled = {}
for modality, entries in obs.items():
tiled[modality] = {}
for key, val in entries.items():
if isinstance(val, list):
# language: [["text"]] -> [["text"]] * n
tiled[modality][key] = val * n
else:
# numpy/tensor: repeat along batch dim 0
import numpy as np
if isinstance(val, np.ndarray):
tiled[modality][key] = np.repeat(val, n, axis=0)
else:
tiled[modality][key] = val.repeat(n, *([1] * (val.ndim - 1)))
return tiled
def main(args: VerifyConfig | None = None):
if args is None:
args = tyro.cli(VerifyConfig)
print("=" * 60)
print("N1.7 TRT Verification")
print("=" * 60)
resolve_batch_size(args.engine_dir, args.batch_size, source="verify_n1d7_trt")
# Step 1: Load policy and get PyTorch reference output
print("\n[1] Loading policy...")
policy = Gr00tPolicy(
embodiment_tag=args.embodiment_tag,
model_path=args.model_path,
device="cuda",
)
print("[2] Loading dataset...")
dataset = LeRobotEpisodeLoader(
dataset_path=args.dataset_path,
modality_configs=policy.get_modality_config(),
)
# --- Capture ViT input/output and backbone output from PyTorch ---
pt_backbone_features = None
pt_vit_output = None
pt_vit_input = None
def _capture_backbone_hook(module, args, output):
nonlocal pt_backbone_features
pt_backbone_features = output["backbone_features"].detach().clone()
return output
def _capture_vit_hook(module, args, kwargs, output):
nonlocal pt_vit_output, pt_vit_input
# Capture input pixel_values
if args:
pt_vit_input = args[0].detach().clone()
elif "pixel_values" in kwargs:
pt_vit_input = kwargs["pixel_values"].detach().clone()
# Capture output (image_embeds after merger)
if isinstance(output, tuple):
pt_vit_output = output[0].detach().clone()
else:
pt_vit_output = output.detach().clone()
return output
backbone_hook = policy.model.backbone.register_forward_hook(_capture_backbone_hook)
vit_hook = policy.model.backbone.model.model.visual.register_forward_hook(
_capture_vit_hook, with_kwargs=True
)
print("[3] Running PyTorch inference...")
obs = prepare_observation(policy, dataset, traj_idx=0)
torch.manual_seed(42)
with torch.inference_mode():
result = policy.get_action(obs)
backbone_hook.remove()
vit_hook.remove()
# get_action returns (action_dict, info_dict)
action_dict = result[0] if isinstance(result, tuple) else result
print(f" Action keys: {list(action_dict.keys())}")
# Concatenate all action arrays into a single tensor for comparison
pt_arrays = []
for k in sorted(action_dict.keys()):
v = action_dict[k]
t = torch.tensor(v) if not isinstance(v, torch.Tensor) else v
pt_arrays.append(t.float().flatten())
print(f" {k}: shape={v.shape if hasattr(v, 'shape') else len(v)}")
pt_action = torch.cat(pt_arrays)
# Step 2: Setup TRT engines and run
print("\n[4] Loading TRT engines...")
from trt_model_forward import setup_tensorrt_engines
setup_tensorrt_engines(policy, args.engine_dir, mode=args.mode)
# --- Capture backbone output from TRT ---
trt_backbone_features = None
def _capture_trt_backbone_hook(module, args, output):
nonlocal trt_backbone_features
trt_backbone_features = output["backbone_features"].detach().clone()
return output
backbone_hook2 = policy.model.backbone.register_forward_hook(_capture_trt_backbone_hook)
# Run ViT TRT with the same pixel_values captured during PyTorch pass
trt_vit_output = None
if pt_vit_input is not None and getattr(policy.model.backbone, "vit_engine", None) is not None:
vit_dtype = policy.model.backbone.vit_engine.dtype_of("pixel_values")
pv = pt_vit_input.to(vit_dtype).cuda().contiguous()
# For batch_size > 1, tile pixel_values to match engine's expected num_patches
if args.batch_size > 1:
pv = pv.repeat(args.batch_size, 1)
policy.model.backbone.vit_engine.set_runtime_tensor_shape("pixel_values", pv.shape)
vit_result = policy.model.backbone.vit_engine(pv)
# Take first batch's merged patches for comparison
num_merged = (
pt_vit_output.shape[0]
if pt_vit_output is not None
else vit_result["image_embeds"].shape[0] // args.batch_size
)
trt_vit_output = vit_result["image_embeds"][:num_merged].detach().clone()
print("[5] Running TRT inference...")
obs2 = prepare_observation(policy, dataset, traj_idx=0)
if args.batch_size > 1:
print(f" Tiling observation to batch_size={args.batch_size}")
obs2 = _tile_observation(obs2, args.batch_size)
torch.manual_seed(42)
with torch.inference_mode():
result2 = policy.get_action(obs2)
backbone_hook2.remove()
action_dict2 = result2[0] if isinstance(result2, tuple) else result2
trt_arrays = []
for k in sorted(action_dict2.keys()):
v = action_dict2[k]
# For batch_size > 1, take slice [0] to compare against single PyTorch output
if args.batch_size > 1 and hasattr(v, "shape") and v.shape[0] == args.batch_size:
v = v[0:1]
t = torch.tensor(v) if not isinstance(v, torch.Tensor) else v
trt_arrays.append(t.float().flatten())
trt_act = torch.cat(trt_arrays)
# Step 3a: Compare ViT outputs
if pt_vit_output is not None and trt_vit_output is not None:
vit_pt = pt_vit_output.float().flatten()
vit_trt = trt_vit_output.float().flatten()
vit_cosine = cosine_similarity(vit_pt.unsqueeze(0), vit_trt.unsqueeze(0)).item()
vit_l1 = (vit_pt - vit_trt).abs().mean().item()
vit_linf = (vit_pt - vit_trt).abs().max().item()
print("\n[6a] ViT output comparison (image_embeds):")
print(f" Cosine Similarity: {vit_cosine:.6f}")
print(f" L1 Mean Error: {vit_l1:.6f}")
print(f" L∞ Max Error: {vit_linf:.6f}")
else:
print("\n[6a] ViT comparison skipped (PyTorch ViT was deleted before capture)")
# Step 3b: Compare backbone outputs (before vl_self_attention)
if pt_backbone_features is not None and trt_backbone_features is not None:
bb_pt = pt_backbone_features.float().flatten()
# For batch_size > 1, take slice [0] to match PyTorch single-batch output
trt_bb = trt_backbone_features[:1] if args.batch_size > 1 else trt_backbone_features
bb_trt = trt_bb.float().flatten()
bb_cosine = cosine_similarity(bb_pt.unsqueeze(0), bb_trt.unsqueeze(0)).item()
bb_l1 = (bb_pt - bb_trt).abs().mean().item()
bb_linf = (bb_pt - bb_trt).abs().max().item()
print("\n[6b] Backbone output comparison (LLM output, before vl_self_attention):")
print(f" Cosine Similarity: {bb_cosine:.6f}")
print(f" L1 Mean Error: {bb_l1:.6f}")
print(f" L∞ Max Error: {bb_linf:.6f}")
# Step 4: Compare final action outputs
print("\n[6b] Final action output comparison:")
pt_flat = pt_action.float().flatten()
trt_flat = trt_act.float().flatten()
cosine = cosine_similarity(pt_flat.unsqueeze(0), trt_flat.unsqueeze(0)).item()
l1 = (pt_flat - trt_flat).abs().mean().item()
linf = (pt_flat - trt_flat).abs().max().item()
print(f"\n Cosine Similarity: {cosine:.6f}")
print(f" L1 Mean Error: {l1:.6f}")
print(f" L∞ Max Error: {linf:.6f}")
if cosine > 0.999:
print("\n PASS — TRT matches PyTorch")
elif cosine > 0.99:
print("\n WARN — Minor drift detected")
else:
print("\n FAIL — Significant divergence")
return cosine
if __name__ == "__main__":
config = tyro.cli(VerifyConfig)
main(config)
|