File size: 10,521 Bytes

e479c46

# 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.

"""CPU-only tests for the export/TRT single-source contract helpers.

``_trt_contract`` has no heavy deps (json / os / logging only), so we import
it directly after putting ``scripts/deployment`` on ``sys.path``.
"""

from __future__ import annotations

import json
import os
import sys
import types

import pytest


DEPLOY_DIR = os.path.abspath(os.path.join(os.path.dirname(__file__), "../../../scripts/deployment"))
if DEPLOY_DIR not in sys.path:
    sys.path.insert(0, DEPLOY_DIR)

import _trt_contract as tc  # noqa: E402


def _write_metadata(d, **kwargs):
    meta = {"action_horizon": 16, "sa_seq_len": 17, "batch_size": 1}
    meta.update(kwargs)
    with open(os.path.join(d, "export_metadata.json"), "w") as f:
        json.dump(meta, f)
    return meta


def _fake_policy(action_horizon):
    cfg = types.SimpleNamespace(action_horizon=action_horizon)
    action_head = types.SimpleNamespace(config=cfg, action_horizon=action_horizon)
    model = types.SimpleNamespace(action_head=action_head)
    return types.SimpleNamespace(model=model)


# --- load_export_metadata -------------------------------------------------


def test_load_metadata_from_engine_dir(tmp_path):
    _write_metadata(tmp_path)
    meta = tc.load_export_metadata(str(tmp_path))
    assert meta["action_horizon"] == 16


def test_load_metadata_from_engine_file(tmp_path):
    _write_metadata(tmp_path)
    meta = tc.load_export_metadata(str(tmp_path / "dit_bf16.engine"))
    assert meta["batch_size"] == 1


def test_load_metadata_from_sibling_onnx_dir(tmp_path):
    onnx = tmp_path / "onnx"
    engines = tmp_path / "engines"
    onnx.mkdir()
    engines.mkdir()
    _write_metadata(onnx)
    meta = tc.load_export_metadata(str(engines))
    assert meta["action_horizon"] == 16


def test_load_metadata_absent_returns_none(tmp_path):
    assert tc.load_export_metadata(str(tmp_path)) is None


# --- validate_export_metadata ---------------------------------------------


def _valid_metadata():
    return {
        "schema_version": tc.EXPORT_METADATA_SCHEMA_VERSION,
        "sa_seq_len": 17,
        "vl_seq_len": 280,
        "llm_seq_len": 280,
        "num_patches": 256,
        "num_merged_patches": 64,
        "num_vis_tokens": 64,
        "action_horizon": 16,
        "batch_size": 1,
        "precision": "bf16",
    }


def test_validate_export_metadata_ok():
    tc.validate_export_metadata(_valid_metadata())


def test_validate_export_metadata_wrong_version_raises():
    meta = _valid_metadata()
    meta["schema_version"] = tc.EXPORT_METADATA_SCHEMA_VERSION + 1
    with pytest.raises(ValueError, match="schema_version"):
        tc.validate_export_metadata(meta)


def test_validate_export_metadata_missing_version_raises():
    meta = _valid_metadata()
    del meta["schema_version"]
    with pytest.raises(ValueError, match="schema_version"):
        tc.validate_export_metadata(meta)


@pytest.mark.parametrize(
    "key", [k for k in tc.REQUIRED_EXPORT_METADATA_KEYS if k != "schema_version"]
)
def test_validate_export_metadata_missing_key_raises(key):
    meta = _valid_metadata()
    del meta[key]
    with pytest.raises(ValueError, match="missing required key"):
        tc.validate_export_metadata(meta)


# --- assert_engine_matches_policy -----------------------------------------


def test_engine_matches_policy_ok(tmp_path):
    _write_metadata(tmp_path, action_horizon=16, sa_seq_len=17)
    out = tc.assert_engine_matches_policy(_fake_policy(16), str(tmp_path))
    assert out["action_horizon"] == 16


def test_engine_action_horizon_mismatch_raises(tmp_path):
    _write_metadata(tmp_path, action_horizon=16, sa_seq_len=17)
    with pytest.raises(ValueError, match="disagree on chunk size"):
        tc.assert_engine_matches_policy(_fake_policy(40), str(tmp_path))


def test_engine_corrupt_sa_seq_len_raises(tmp_path):
    _write_metadata(tmp_path, action_horizon=16, sa_seq_len=99)
    with pytest.raises(ValueError, match="corrupt"):
        tc.assert_engine_matches_policy(_fake_policy(16), str(tmp_path))


def test_engine_missing_metadata_warns_returns_none(tmp_path, caplog):
    with caplog.at_level("WARNING"):
        out = tc.assert_engine_matches_policy(_fake_policy(16), str(tmp_path))
    assert out is None
    assert any("no export_metadata.json" in r.getMessage() for r in caplog.records)


def test_corrupt_metadata_treated_as_absent(tmp_path):
    (tmp_path / "export_metadata.json").write_text("{ not valid json ")
    # Corrupt file must not crash; load returns None, validation degrades.
    assert tc.load_export_metadata(str(tmp_path)) is None
    assert tc.assert_engine_matches_policy(_fake_policy(16), str(tmp_path)) is None


def test_action_horizon_mismatch_message_without_sa_seq_len(tmp_path):
    # Metadata has action_horizon but no sa_seq_len: the error must not print
    # the "sa_seq_len=None" placeholder.
    with open(tmp_path / "export_metadata.json", "w") as f:
        json.dump({"action_horizon": 16, "batch_size": 1}, f)
    with pytest.raises(ValueError) as exc:
        tc.assert_engine_matches_policy(_fake_policy(40), str(tmp_path))
    assert "sa_seq_len=None" not in str(exc.value)


# --- assert_engine_bundle_present -----------------------------------------


_FULL_PIPELINE_REQUIRED = (
    "state_encoder.engine",
    "action_encoder.engine",
    "dit_bf16.engine",
    "action_decoder.engine",
)


def test_bundle_present_ok(tmp_path):
    for name in _FULL_PIPELINE_REQUIRED:
        (tmp_path / name).write_bytes(b"stub")
    # All required engines present -> no raise.
    tc.assert_engine_bundle_present(
        str(tmp_path), _FULL_PIPELINE_REQUIRED, mode="n17_full_pipeline"
    )


def test_bundle_missing_dir_raises_with_build_hint(tmp_path):
    missing_dir = tmp_path / "gr00t_trt_deployment" / "engines"
    with pytest.raises(FileNotFoundError) as exc:
        tc.assert_engine_bundle_present(
            str(missing_dir), _FULL_PIPELINE_REQUIRED, mode="n17_full_pipeline"
        )
    msg = str(exc.value)
    assert "n17_full_pipeline" in msg
    assert "build_trt_pipeline.py" in msg  # actionable build hint, not a bare error


def test_bundle_missing_one_file_names_it(tmp_path):
    for name in _FULL_PIPELINE_REQUIRED:
        if name != "dit_bf16.engine":
            (tmp_path / name).write_bytes(b"stub")
    with pytest.raises(FileNotFoundError) as exc:
        tc.assert_engine_bundle_present(
            str(tmp_path), _FULL_PIPELINE_REQUIRED, mode="n17_full_pipeline"
        )
    msg = str(exc.value)
    assert "dit_bf16.engine" in msg
    assert "state_encoder.engine" not in msg  # only the missing file is listed


# --- resolve_batch_size ----------------------------------------------------


def test_resolve_batch_size_default_from_metadata(tmp_path):
    _write_metadata(tmp_path, batch_size=4)
    assert tc.resolve_batch_size(str(tmp_path)) == 4


def test_resolve_batch_size_matching_request(tmp_path):
    _write_metadata(tmp_path, batch_size=2)
    assert tc.resolve_batch_size(str(tmp_path), 2) == 2


def test_resolve_batch_size_mismatch_raises(tmp_path):
    _write_metadata(tmp_path, batch_size=1)
    with pytest.raises(ValueError, match="built .*for batch_size=1"):
        tc.resolve_batch_size(str(tmp_path), 4)


def test_resolve_batch_size_no_metadata_defaults_to_one(tmp_path):
    assert tc.resolve_batch_size(str(tmp_path)) == 1
    # A request with no metadata is accepted (nothing to validate against).
    assert tc.resolve_batch_size(str(tmp_path), 8) == 8


# --- assert_grid_thw_matches ----------------------------------------------


def test_grid_thw_matches_ok():
    tc.assert_grid_thw_matches([[1, 16, 16]], [[1, 16, 16]])


def test_grid_thw_none_baked_skips():
    # Older bundle without a recorded grid: degrade to a no-op, do not raise.
    tc.assert_grid_thw_matches(None, [[1, 8, 8]])


def test_grid_thw_different_layout_same_count_raises():
    # Same patch count (16*16 == 8*32) but a different layout: the static
    # pixel_values shape would NOT catch this; the grid check must.
    with pytest.raises(ValueError, match="image_grid_thw"):
        tc.assert_grid_thw_matches([[1, 16, 16]], [[1, 8, 32]])


def test_grid_thw_more_views_same_layout_ok():
    # Batch scaling tiles the same per-view grid: more views of an already-baked
    # layout are fine (the static pixel_values shape enforces the count). This is
    # the test_trt_full_pipeline[batch=2] path: baked 2 views, runtime 4 views.
    tc.assert_grid_thw_matches([[1, 16, 16], [1, 16, 16]], [[1, 16, 16]] * 4)


def test_grid_thw_extra_view_unbaked_layout_raises():
    # An extra view whose layout was never baked still gets wrong embeddings.
    with pytest.raises(ValueError, match="ViT TRT"):
        tc.assert_grid_thw_matches([[1, 16, 16]], [[1, 16, 16], [1, 8, 32]])


def test_grid_thw_accepts_tensor_like():
    class _FakeTensor:
        def __init__(self, data):
            self._data = data

        def detach(self):
            return self

        def cpu(self):
            return self

        def tolist(self):
            return self._data

    tc.assert_grid_thw_matches([[1, 16, 16]], _FakeTensor([[1, 16, 16]]))
    with pytest.raises(ValueError):
        tc.assert_grid_thw_matches([[1, 16, 16]], _FakeTensor([[2, 16, 16]]))


# --- assert_exec_horizon_within_model -------------------------------------


@pytest.mark.parametrize(
    "exec_h,model_h,ok",
    [(16, 16, True), (8, 16, True), (1, 40, True), (17, 16, False), (0, 16, False)],
)
def test_assert_exec_horizon_within_model(exec_h, model_h, ok):
    if ok:
        tc.assert_exec_horizon_within_model(exec_horizon=exec_h, model_action_horizon=model_h)
    else:
        with pytest.raises(ValueError, match="execution-horizon"):
            tc.assert_exec_horizon_within_model(exec_horizon=exec_h, model_action_horizon=model_h)