Fix autotuner latency ranking and the FLOPS metric name

[Aravind-11]

Fixes #8766

The bug

"metric": "latency" keeps the largest latency, so ds_config_optimal.json is the slowest trial. get_best_space_record, get_best_space_records, ResourceManager.parse_results, and BaseTuner.tune all use >. Throughput and FLOPS are larger-is-better. Latency is the step time, so the smaller number is the better config. find_estimated_top_configs already reverses the sort for latency. The selectors that publish the winner do not.

The engine writes FLOPS as FLOPS_per_gpu (DeepSpeedEngine._autotuning_exit). The README tells users to set "FLOPS", and AUTOTUNING_METRIC_FLOPS is "flops". results[metric] raises KeyError for either name. On the fast path that read is in run_tuning_micro_batch_sizes, outside the tuner's try, so the run aborts.

The fix

metric_is_better in deepspeed/autotuning/utils.py is the one comparison. Latency replaces the incumbent only when it is smaller. Every other metric replaces it only when it is larger. None never wins.

metric_value maps "flops" and "FLOPS" to FLOPS_per_gpu. An exact key already present in the file still wins.

The cross-stage check treats a previous metric of 0 as "no previous stage", not as a measured latency of zero. A real latency is not compared against that sentinel. Throughput ranking is unchanged.

Testing

CPU tests in tests/unit/autotuning/test_autotuning.py:

• "FLOPS", "flops", and "FLOPS_per_gpu" read the engine's field
• latency keeps the smaller value, including the global winner across spaces
• throughput still keeps the larger value
• parse_results does both
• BaseTuner.tune keeps the smaller latency and does not let a missing metric replace a real one
• the 0 sentinel does not count as a regression for latency or throughput

tests/unit/autotuning/test_autotuning.py passes (18 tests). yapf and flake8 are clean on the changed files. No GPU run.

cc @loadams