DeepSpeed

7c1b0f2c - Validate positive inference and HybridEngine max output tokens (#8343)

Validate positive inference and HybridEngine max output tokens (#8343) ## Description Reject zero and negative `max_out_tokens` during initial configuration validation. These values previously passed parsing and failed later during generation or model conversion. Both `DeepSpeedInferenceConfig` (including its `max_tokens` alias) and `HybridEngineConfig` now require a positive value. HybridEngine therefore rejects invalid values when parsing the top-level `DeepSpeedConfig`, before constructing an inference configuration during model conversion. The defaults remain 1024 for inference and 512 for HybridEngine. Regression coverage is included in the existing inference and runtime configuration test files. Inference tests reject -1 and 0 through both field names. HybridEngine tests construct the top-level `DeepSpeedConfig` with HybridEngine enabled, reject -1 and 0, accept 1/512/1024, and preserve the default of 512. This covers the parsed configuration consumed by `DeepSpeedHybridEngine.new_inference_container`. Fixes #8339. ## Testing - `python -m pytest -q -m 'inference or not inference' tests/unit/runtime/test_ds_config_model.py tests/unit/inference/test_inference_config.py --tb=short`: **25 passed**, including the regression coverage and existing CUDA Graph test. Executed on an NVIDIA RTX A1000 Laptop GPU (4 GiB, driver 595.84), Python 3.12.13, PyTorch 2.13.0+cu130. - Confirmed that the HybridEngine -1/0 rejection tests fail without the new constraint and pass with it. - `pre-commit run --files deepspeed/inference/config.py deepspeed/runtime/config.py tests/unit/inference/test_inference_config.py tests/unit/runtime/test_ds_config_model.py`: passed. - `python ci/tests_fetcher.py --base origin/master`: selects the full `tests/unit/v1` scope because `deepspeed/runtime/config.py` is shared runtime configuration. No full HybridEngine model conversion or generation run was performed. --------- Signed-off-by: taking-lying-flat <1615405@qq.com> Co-authored-by: Masahiro Tanaka <81312776+tohtana@users.noreply.github.com>