version: "1.0.0" name: flaggems-fusion-metrics description: Run, resume, merge, or export the FlagGems fusion compute-TFLOPS and memory-bandwidth metric suite. Use for fusion metric coverage, per-operator or full pytest benchmark runs, CSV/run.json reports, and formula Markdown exports. Includes a maintenance reference for infrequent formula or architecture changes.

FlagGems Fusion Metrics

Use scripts/run_fusion_metrics.py as the only operational entrypoint. Keep formula truth in $TRITON_REPO_DIR/FlagGems/benchmark/fusion_metrics_formulas.yaml.

Run

Activate the configured Python environment before every Python command:

source ~/myenv/bin/activate
python3 \
  "$AGENT_DIR/agentfortritoncpu/skills/flaggems-fusion-metrics/scripts/run_fusion_metrics.py" \
  --ops silu_and_mul

Use --all for the full configured set and --dry-run to inspect selection, NUMA bindings, and commands. Results default to $AGENT_DIR/logs/fusion_metrics/; use --output-dir to override. Keep --omp-threads at or below 32; the runner sets child OMP_NUM_THREADS.

Compute report targets follow the selected pipeline and dtype after applying the scale 0.2 / 304 * 32: SME/float32 is 2.52631579 TFLOPS, SME/bfloat16 is 5.05263158 TFLOPS, and SVE/float32 is 0.71578947 TFLOPS. Use --target-tflops or FUSION_TARGET_TFLOPS only to apply an explicit uniform override; otherwise the runner selects the scaled value per result row.

Memory report default target bandwidth is 30.72 GB/s. Override with --target-gbps or FUSION_TARGET_GBPS when needed.

When --dtypes or FUSION_DTYPES explicitly selects a dtype and an operator run fails, retry that operator with BF16, then retry without --dtypes if BF16 also fails. The terminal prints each fallback, and each attempt keeps its own raw stdout/record log. Runs without an explicit dtype keep the normal default dtype selection. Other pipeline/dtype combinations fail preflight until an explicit override is provided.

Optional defaults come from scripts/.env, with precedence CLI > scripts/.env > code defaults. Allowed fields are FUSION_WARMUP, FUSION_ITER, FUSION_MODE, FUSION_LEVEL, FUSION_DTYPES, FUSION_JOBS, FUSION_OMP_THREADS, FUSION_NUMA_NODES, FUSION_TARGET_TFLOPS, FUSION_TARGET_GBPS, FUSION_TIMEOUT, FUSION_SHAPE_FILE, and FUSION_OUTPUT_DIR.

The runner loads the fixed YAML and sets the formula config and selected formula key explicitly for each pytest child. The formula evaluator binds the actual benchmark args/kwargs, computes one logical_flops or logical_bytes result, and leaves result collection to the existing benchmark path. The runner preserves raw stdout/record logs and publishes the existing compute, memory, coverage, and run manifest structures.

Export formulas

Export documentation without environment, NUMA, or pytest preflight:

source ~/myenv/bin/activate
python3 \
  "$AGENT_DIR/agentfortritoncpu/skills/flaggems-fusion-metrics/scripts/run_fusion_metrics.py" \
  --export-formulas "$AGENT_DIR/logs/fusion_metrics/formulas.md"

Always provide an explicit output file.

For infrequent formula, schema, evaluator, or runner architecture maintenance, read references/design.md completely. Do not load it for routine runs.