Benchmark test — tests/benchmark/test_bagel.py Single-forward latency / throughput over the bagel_model text backbone at the loader's fixed shape (DEFAULT_SEQ_LEN = 128, batch 1, bf16), reusing the loader's get_mesh_config / load_shard_spec so the benchmark shards identically to the component test. Emit the standard perf metrics JSON.
Harness wiring — extend the relevant module under tests/benchmark/benchmarks/ rather than hand-rolling timing, matching how LongCat extended video_gen_benchmark.py.
Matrix registration — one entry in .github/workflows/perf-bench-matrix.json:
{
"name": "bagel_tp",
"pytest": "tests/benchmark/test_bagel.py::test_bagel",
"runs-on": "<see open decision 1>",
"group": "regular"
}
Note test-defaults supplies runs-on: [n150-perf, p150-perf], bs: 12, df: bfloat16, lp: 128. BAGEL must override runs-on (it cannot fit a single chip) and almost certainly bs.
Benchmark test — tests/benchmark/test_bagel.py Single-forward latency / throughput over the bagel_model text backbone at the loader's fixed shape (DEFAULT_SEQ_LEN = 128, batch 1, bf16), reusing the loader's get_mesh_config / load_shard_spec so the benchmark shards identically to the component test. Emit the standard perf metrics JSON.
Harness wiring — extend the relevant module under tests/benchmark/benchmarks/ rather than hand-rolling timing, matching how LongCat extended video_gen_benchmark.py.
Matrix registration — one entry in .github/workflows/perf-bench-matrix.json:
{
"name": "bagel_tp",
"pytest": "tests/benchmark/test_bagel.py::test_bagel",
"runs-on": "<see open decision 1>",
"group": "regular"
}
Note test-defaults supplies runs-on: [n150-perf, p150-perf], bs: 12, df: bfloat16, lp: 128. BAGEL must override runs-on (it cannot fit a single chip) and almost certainly bs.