From d81341b9b3bc728ac0eb71a629bebda9852c2cd8 Mon Sep 17 00:00:00 2001 From: kesmeey <107767849+kesmeey@users.noreply.github.com> Date: Thu, 18 Dec 2025 10:57:53 +0800 Subject: [PATCH] =?UTF-8?q?[CI]=E3=80=90Hackathon=209th=20Sprint=20No.14?= =?UTF-8?q?=E3=80=91=E5=8A=9F=E8=83=BD=E6=A8=A1=E5=9D=97=20=20fastdeploy/r?= =?UTF-8?q?l/rollout=5Fmodel.py=20=E5=8D=95=E6=B5=8B=E8=A1=A5=E5=85=85=20(?= =?UTF-8?q?#5552)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit * Add rollout model unit tests * test: update rl rollout_model tests * test: fix cache_type_branches unsupported platform case * test: fix rl rollout_model test indent * Delete tests/spec_decode/test_mtp_proposer.py * chore: format test_rollout_model * chore: translate rollout test comments to English * test: guard rollout_model import by disabling auto registry * chore: reorder imports in rl rollout test * test: isolate env for RL rollout tests * style: format rollout RL tests with black * update * test: remove RL rollout unit tests causing collection issues * test: add lightweight rollout_model RL unit tests * fix(coverage): filter test file paths and handle collection failures - Only extract real test file paths (tests/.../test_*.py) from pytest collect output - Filter out ERROR/collecting prefixes to prevent garbage in failed_tests.log - Add proper error handling for pytest collection failures - Exit early if no test files can be extracted - Preserve collection error output for debugging * update * style: fix code style issues in test_rollout_model.py - Remove unused 'os' import - Remove trailing blank lines --------- Co-authored-by: YuBaoku <49938469+EmmonsCurse@users.noreply.github.com> --- scripts/coverage_run.sh | 13 +- tests/rl/test_rollout_model.py | 215 +++++++++++++++++++++++++++++++++ 2 files changed, 225 insertions(+), 3 deletions(-) create mode 100644 tests/rl/test_rollout_model.py diff --git a/scripts/coverage_run.sh b/scripts/coverage_run.sh index 6f3700f86..d0e94c363 100644 --- a/scripts/coverage_run.sh +++ b/scripts/coverage_run.sh @@ -14,9 +14,16 @@ failed_tests_file="failed_tests.log" ################################## # 执行 pytest，每个文件单独跑 -################################## -# 收集 pytest 文件 -TEST_FILES=$(python -m pytest --collect-only -q -c ${PYTEST_INI} ${tests_path} --rootdir=${run_path} --disable-warnings | grep -Eo '^.*test_.*\.py' | sort | uniq) +# 使用 pytest 的 --collect-only 输出，并从每行中提取真正的测试文件路径（形如 tests/.../test_*.py）。 +# 注意：pytest 在收集失败时会输出形如 "ERROR tests/xxx/test_xxx.py::test_xxx ..." 的行， +# 为了避免把前缀 "ERROR"/"FAILED"/"collecting" 等误当成文件名，这里只保留行中出现的 +# "tests/.../test_*.py" 这一段，其他前后内容直接丢弃。 +TEST_FILES=$( + python -m pytest --collect-only -q -c "${PYTEST_INI}" "${tests_path}" --rootdir="${run_path}" --disable-warnings 2>&1 \ + | grep -E 'tests/.+\/test_.*\.py' \ + | sed -E 's@.*(tests/[^: ]*test_[^: ]*\.py).*@\1@' \ + | sort -u +) failed_pytest=0 diff --git a/tests/rl/test_rollout_model.py b/tests/rl/test_rollout_model.py new file mode 100644 index 000000000..4d720bcc2 --- /dev/null +++ b/tests/rl/test_rollout_model.py @@ -0,0 +1,215 @@ +""" +Unit tests for `fastdeploy.rl.rollout_model`. + +These tests focus purely on Python-side mapping / quantization logic and +intentionally avoid any heavy engine or model initialization. They should not +modify global environment in a way that affects other test modules. +""" + +import types + +import pytest # type: ignore + +# Conservative guard: skip locally if paddle is missing; CI has paddle installed. +try: # pragma: no cover - env probe + import paddle # noqa: F401 +except Exception as e: # pragma: no cover - env probe + pytest.skip(f"Skip RL rollout tests, paddle import failed: {e}", allow_module_level=True) + + +from fastdeploy.rl.rollout_model import ( + BaseRLModel, + Ernie4_5_MoeForCausalLMRL, + Ernie4_5_VLMoeForConditionalGenerationRL, + Glm4MoeForCausalLMRL, + Qwen2_5_VLForConditionalGenerationRL, + Qwen2ForCausalLMRL, + Qwen3ForCausalLMRL, + Qwen3MoeForCausalLMRL, + RolloutModel, +) + + +def _dummy_instance( + cls, + model_config_kwargs, + state_keys, + parallel_config=None, + quant_name="wint8", +): + """Create a lightweight instance without running heavy model init.""" + inst = cls.__new__(cls) + BaseRLModel.__init__(inst) + inst.fd_config = types.SimpleNamespace( + model_config=types.SimpleNamespace(**model_config_kwargs), + parallel_config=parallel_config or types.SimpleNamespace(tensor_parallel_size=1), + quant_config=types.SimpleNamespace(name=lambda: quant_name), + ) + inst.state_dict = lambda: {k: 0 for k in state_keys} + return inst + + +def test_rollout_model_quantization_and_state_dict_fallback(): + """RolloutModel wrapper should safely delegate to underlying rollout_model.""" + # Cover default branch when rollout_model lacks quantization/state_dict + fallback = RolloutModel.__new__(RolloutModel) + fallback.rollout_model = types.SimpleNamespace() + assert fallback.get_quantization_infer_keys() == {} + + # Cover delegate branch when rollout_model implements quantization/state_dict + forwarded = RolloutModel.__new__(RolloutModel) + forwarded.rollout_model = types.SimpleNamespace( + get_quantization_infer_keys=lambda: {"k": "v"}, + state_dict=lambda: {"p": 1}, + ) + assert forwarded.get_quantization_infer_keys() == {"k": "v"} + assert forwarded.state_dict() == {"p": 1} + + +def test_base_rl_name_and_quantization_keys_and_error(): + model = BaseRLModel.__new__(BaseRLModel) + BaseRLModel.__init__(model) + + # Cover BaseRLModel.name and wint8 branch + assert BaseRLModel.name() == "BaseRLModel" + model.fd_config = types.SimpleNamespace(quant_config=types.SimpleNamespace(name=lambda: "wint8")) + model.state_dict = lambda: { + "a.weight_scale": 1, + "b.weight_scale": 2, + "c.weight": 3, + } + assert model.get_quantization_infer_keys() == ["a.weight", "b.weight"] + + # Cover non-wint8 branch raising error + model.fd_config = types.SimpleNamespace(quant_config=types.SimpleNamespace(name=lambda: "fp16")) + with pytest.raises(ValueError): + model.get_quantization_infer_keys() + + +def test_complete_missing_mappings_skips_scale(): + model = BaseRLModel.__new__(BaseRLModel) + BaseRLModel.__init__(model) + model.state_dict = lambda: { + "kept.weight": 1, + "ignored.weight_scale": 2, + } + model._complete_missing_mappings() + assert model.infer_to_train_mapping["kept.weight"] == "kept.weight" + assert "ignored.weight_scale" not in model.infer_to_train_mapping + + +def test_ernie45_moe_mapping_and_cache(): + dummy = _dummy_instance( + Ernie4_5_MoeForCausalLMRL, + { + "moe_use_aux_free": True, + "moe_num_experts": 2, + "moe_layer_start_index": 1, + "num_hidden_layers": 3, + }, + [ + "ernie.layers.1.mlp.experts.0.up_gate_proj.weight", + "ernie.layers.1.mlp.experts.0.down_proj.weight", + "some.weight", + "scale.weight_scale", + ], + ) + first = dummy.get_name_mappings_to_training() + # Cover gate/gate_correction_bias mapping and MoE experts aggregation + assert "ernie.layers.1.mlp.experts.gate_correction_bias" in first + assert first["some.weight"] == "some.weight" + assert "scale.weight_scale" not in first + # Cover cached path + assert dummy.get_name_mappings_to_training() is first + + +def test_ernie45_vl_moe_text_and_image_mappings(): + dummy = _dummy_instance( + Ernie4_5_VLMoeForConditionalGenerationRL, + { + "moe_use_aux_free": False, + "moe_num_experts": [8, 8], + "moe_layer_start_index": (0, 1), + "moe_layer_end_index": (1, 2), + "num_hidden_layers": 2, + }, + [ + "ernie.layers.0.mlp.experts.0.up_gate_proj.weight", + "ernie.layers.1.mlp.experts.2.down_proj.weight", + ], + parallel_config=types.SimpleNamespace(tensor_parallel_size=4), + ) + mappings = dummy.get_name_mappings_to_training() + # Cover fused MoE text/image expert mappings + assert "ernie.layers.0.mlp.text_fused_moe.experts.up_gate_proj_weight" in mappings + assert "ernie.layers.1.mlp.image_fused_moe.experts.down_proj_weight" in mappings + + +def test_qwen2_mapping_builds_and_completes(): + dummy = _dummy_instance( + Qwen2ForCausalLMRL, + {"num_hidden_layers": 2}, + ["qwen2.layers.0.mlp.gate_up_fused_proj.weight"], + ) + mappings = dummy.get_name_mappings_to_training() + # Cover up_gate_proj -> gate_up_fused_proj mapping + assert "qwen2.layers.0.mlp.up_gate_proj.weight" in mappings + assert mappings["qwen2.layers.0.mlp.up_gate_proj.weight"] == "qwen2.layers.0.mlp.gate_up_fused_proj.weight" + + +def test_qwen3moe_mapping_aux_free(): + dummy = _dummy_instance( + Qwen3MoeForCausalLMRL, + {"moe_use_aux_free": True, "num_experts": 1, "num_hidden_layers": 1}, + [ + "model.layers.0.mlp.experts.0.up_gate_proj.weight", + "model.layers.0.mlp.experts.0.down_proj.weight", + ], + ) + mappings = dummy.get_name_mappings_to_training() + # Cover gate/gate_correction_bias handling and expert merge + assert "model.layers.0.mlp.gate.weight" in mappings + + +def test_qwen3_mapping_basic(): + dummy = _dummy_instance( + Qwen3ForCausalLMRL, + {"num_hidden_layers": 1}, + ["model.layers.0.mlp.gate_up_fused_proj.weight"], + ) + mappings = dummy.get_name_mappings_to_training() + assert "model.layers.0.mlp.up_gate_proj.weight" in mappings + + +def test_qwen25_vl_mapping_basic(): + dummy = _dummy_instance( + Qwen2_5_VLForConditionalGenerationRL, + {"num_hidden_layers": 1}, + ["model.layers.0.mlp.gate_up_fused_proj.weight"], + ) + mappings = dummy.get_name_mappings_to_training() + assert "model.layers.0.mlp.up_gate_proj.weight" in mappings + + +def test_glm4moe_mapping_removes_gate_correction(): + dummy = _dummy_instance( + Glm4MoeForCausalLMRL, + { + "n_routed_experts": 1, + "first_k_dense_replace": 0, + "num_hidden_layers": 1, + }, + [ + "model.layers.0.mlp.experts.0.up_gate_proj.weight", + "model.layers.0.mlp.experts.0.down_proj.weight", + "model.layers.0.mlp.experts.gate_correction_bias", + ], + ) + mappings = dummy.get_name_mappings_to_training() + # Cover gate/experts aggregation and dropping gate_correction_bias + assert "model.layers.0.mlp.experts.up_gate_proj_weight" in mappings + assert "model.layers.0.mlp.experts.gate_correction_bias" not in mappings + + +if __name__ == "__main__": + pytest.main()