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 1.25 | summary stringclasses 3
values |
|---|---|---|---|---|---|---|---|
lh | mi210-jais3-moe-m256 | fused_moe_w8a16_m256 | performance | 1 | 16 | 1.201437 | 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-m1024 | fused_moe_w8a16_m1024 | performance | 1 | 16 | 1.249117 | 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-routed-moe-w8a16 | fused_moe_w8a16_routed | performance | 1 | 16 | 1.132514 | 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... |
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 |
Anchor | Reference patch | Status |
|---|---|---|---|---|---|
mi210-jais3-routed-moe-w8a16 |
Whole layer, 1 to 8192 tokens | 1.132514 | calibrated | Large, several kernel and table changes | Usable draft |
mi210-jais3-moe-m256 |
256 tokens only | 1.201437 | calibrated | One new table entry, 21 lines added | Usable draft |
mi210-jais3-moe-m1024 |
1024 tokens only | 1.249117 | calibrated | One table entry, 14 lines changed | Too easy (see below) |
Calibration
All three tasks were calibrated with their own verifier (tests/test.sh) on one MI210, alone on its node, in images
built from each task's environment/Dockerfile (with Enroot, on the pinned vllm/vllm-openai-rocm:v0.32.0 image; every
check in the Dockerfile passed). ref_speedup is the median of 5 oracle runs:
mi210-jais3-routed-moe-w8a16: 1.1334, 1.1329, 1.1320, 1.1325, 1.1322.mi210-jais3-moe-m256: 1.1745, 1.1956, 1.2014, 1.2045, 1.2051.mi210-jais3-moe-m1024: 1.2504, 1.2491, 1.2464, 1.2521, 1.2441.
Self-checks on all three: the unchanged repository scores 0 (speedup 0.9988, 1.0075 and 1.0002), and negative.patch
fails correctness and scores 0. Landing the reference with solution/solve.sh and scoring it as a submission gives
1.1322, 1.2136 and 1.2403, within about 1% of each anchor: runs and days vary by that much, so the reference alone
can score slightly above or below its own anchor. Details are in each task's tests/ref_speedup.caveat.md.
Known limits
mi210-jais3-moe-m1024is too easy. Its reference entry is an exact copy of the table's existing 8192-token entry. A search with no kernel insight (copies of existing entries, then random values from the table) found it among its first tries and scored 1.2516 with the verifier, above the anchor. It is kept for now; do not read a solver's score on it as evidence of skill.mi210-jais3-moe-m256survived that search (96 tries, nothing above 1.1x on its row; the finalist scored 0.998). No real solver has been run on it yet.- In the one-size tasks, correctness covers every token count but only the named size is timed, so a solution can make other sizes slower without losing points.
scripts/run_task.pyfrom Φ-Bench only passes NVIDIA GPUs. Use thedocker runlines in each task's Dockerfile.- Each task's Docker image has not been built with Docker itself, only with Enroot from the same 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
- 218