migrate GDN workloads, moe and gdn baselines from flashinfer-trace
#1
by yongwww - opened
No description provided.
yongwww changed pull request status to closed
yongwww changed pull request status to open
yongwww changed pull request title from ddd to migrate GDN workloads, moe and gdn baselines from flashinfer-trace
I am still testing with the change, will update shortly
gdn works well. The moe flashinfer_wrapper_9sdjf3 timed out for all workloads
flashinfer-bench run \ --local ./mlsys26-contest \ # flashinfer-trace also timed out --definitions moe_fp8_block_scale_ds_routing_topk8_ng8_kg4_e32_h7168_i2048 \ --solutions flashinfer_wrapper_9sdjf3 \ --use-isolated-runner \ --log-level DEBUG \ --timeout 300 \ --atol 1 --rtol 0.3 --required-matched-ratio 0.9
[03:48:12] INFO flashinfer_bench.bench.runner.isolated_runner: Initialized benchmark multi-process on 4 CUDA devices and 4 workers
[03:48:12] INFO flashinfer_bench.cli.main: Running benchmark on FlashInfer Trace Dataset: /workspace/FIB/mlsys26-contest
[03:48:12] INFO flashinfer_bench.bench.benchmark: Processing definition: moe_fp8_block_scale_ds_routing_topk8_ng8_kg4_e32_h7168_i2048 with 1 solutions
[03:53:18] WARNING flashinfer_bench.bench.benchmark: Solution 'flashinfer_wrapper_9sdjf3' for workload b8f4f012-a32e-4356-b4e1-7665b3d598af: TIMEOUT
[03:58:24] WARNING flashinfer_bench.bench.benchmark: Solution 'flashinfer_wrapper_9sdjf3' for workload e05c6c03-5603-4a1c-b34c-dcce0ecaeea4: TIMEOUT
How can we get more detailed output here? @ubospica @zanderjiang
yongwww changed pull request status to merged
yongwww deleted the
refs/pr/1 ref