bench stringclasses 1
value | task_dir stringclasses 3
values | infra_name stringclasses 3
values | reward_class stringclasses 1
value | gpus int64 1 1 | max_submissions int64 16 16 | ref_speedup float64 1.13 1.24 | summary stringclasses 3
values |
|---|---|---|---|---|---|---|---|
lh | mi210-jais3-routed-moe-w8a16 | fused_moe_w8a16_routed | performance | 1 | 16 | 1.130592 | This is the routed-expert computation of one Mixture-of-Experts layer on one tensor-parallel rank of a vLLM server on AMD Instinct MI210, for a block-quantized FP8 model with BF16 activations. The implementation in moe_w8a16.py and its two tile tables is what the server runs today: already tuned, correct, and hard to i... |
lh | mi210-jais3-moe-m1024 | fused_moe_w8a16_m1024 | performance | 1 | 16 | 1.241187 | This is the routed-expert computation of one Mixture-of-Experts layer on one tensor-parallel rank of a vLLM server on AMD Instinct MI210, for a block-quantized FP8 model with BF16 activations. The implementation in moe_w8a16.py and its two tile tables is what the server runs today, and it is correct. You must make the ... |
lh | mi210-jais3-moe-m256 | fused_moe_w8a16_m256 | performance | 1 | 16 | 1.215727 | This is the routed-expert computation of one Mixture-of-Experts layer on one tensor-parallel rank of a vLLM server on AMD Instinct MI210, for a block-quantized FP8 model with BF16 activations. The implementation in moe_w8a16.py and its two tile tables is what the server runs today, and it is correct. You must make the ... |
Φ-Bench MI210 rows (draft)
Draft task rows in the format of Φ-Bench, with the same 8 columns, built from GEAK agent runs on AMD Instinct MI210. MI210 is stated in each summary and task package.
| Task | Scored on | ref_speedup |
Reference patch |
|---|---|---|---|
mi210-jais3-routed-moe-w8a16 |
Whole layer, 1 to 8192 tokens | 1.130592 | Large, several kernel and table changes |
mi210-jais3-moe-m1024 |
1024 tokens only | 1.241187 | One table entry, 14 lines changed |
mi210-jais3-moe-m256 |
256 tokens only | 1.215727 | One new table entry, 21 lines added |
Status: not ready for scoring.
ref_speedupis provisional. It comes from a separate kernel speed check, not from this package's own verifier, and must be replaced by the median of at least 5 verifier runs. A solution that only ties it scores 0.- The image has not been built, and the full verifier, including its no-change and wrong-patch self-checks, has not been run.
- The two one-size tasks were checked once on one MI210 with the shipped task runner and row-only metric: 1.253x and
1.188x, correctness passing. That is not calibration; the 256-token result is 2.3% under its
ref_speedup. - Correctness in the one-size tasks covers every token count, but only the named size is timed, so a solution can make other sizes slower without losing points.
- The one-size tasks may be easy: with 16 submissions, searching table settings may be enough. No solver has been run.
scripts/run_task.pyfrom Φ-Bench only passes NVIDIA GPUs. Use thedocker runlines in each task's Dockerfile.
The vendored code under environment/repo/ keeps its own license (vLLM, Apache-2.0). The license label on this
dataset has not been confirmed.
tasks/*/oracle.patch is the reference solution.
- Downloads last month
- 104