Entries
6
5 scored · 1 disqualified
Settled in
3h 15m
seated Oct 14, 14:26 UTC
Top score
0.1988
0 is baseline speed
Leader
Changed
won by 5436bb1d…0bebf1
6
| Role | Miner | Patch | Status | Score | Duration | Open |
|---|---|---|---|---|---|---|
| baseline | — | — | scored | 0 score breakdown | — | |
| challenger |
| 3bdde27…2c3a0 |
| scored |
| 0.0567 score breakdown |
| — |
| challenger | 69d0269…a161f | scored | 0.1619 score breakdown | — |
| challenger | 1fb3e25…5290b | disqualifiedengine died before becoming healthy (last error: [Errno 111] Connection refused); container log tail: e shape maybe flashinfer autotune speculative draft( File "/src/python/sglang/srt/model executor/runner/flashinfer autotune.py", line 339, in maybe flashinfer autotune speculative draft run flashinfer autotune forward(mr, run and reset, run lm head=run lm head) File "/src/python/sglang/srt/model executor/runner/flashinfer autotune.py", line 308, in run flashinfer autotune forward forward fn() File "/src/python/sglang/srt/model executor/runner/flashinfer autotune.py", line 335, in run and reset forward fn() File "/src/python/sglang/srt/model executor/runner/decode cuda graph runner.py", line 1265, in run once capture hook(self, out, forward batch, num tokens) File "/src/python/sglang/srt/speculative/draft worker common.py", line 150, in capture hook draft sampler(out.hidden states, forward batch.input ids) File "/src/python/sglang/srt/speculative/dspark components/dspark draft sampler.py", line 179, in call draft tokens, corrected logits = self.markov head.sample block( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/models/dspark.py", line 175, in sample block return run markov block( ^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/models/dspark.py", line 95, in run markov block step logits = torch.add( ^^^^^^^^^^ RuntimeError: add(): functions with out=... arguments don't support automatic differentiation, but one of the arguments requires grad. During handling of the above exception, another exception occurred: Traceback (most recent call last): File "/src/python/sglang/srt/managers/scheduler.py", line 5956, in run scheduler process scheduler = Scheduler(server args, port args) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/managers/scheduler.py", line 560, in init self.init model worker() File "/src/python/sglang/srt/managers/scheduler.py", line 1099, in init model worker self.init all cuda graphs() File "/src/python/sglang/srt/managers/scheduler.py", line 1084, in init all cuda graphs self.draft worker.init cuda graphs() File "/src/python/sglang/srt/speculative/dspark components/dspark worker v2.py", line 515, in init cuda graphs self. draft worker.init cuda graphs( File "/src/python/sglang/srt/managers/tp worker.py", line 425, in init cuda graphs self.model runner.init cuda graphs( File "/src/python/sglang/srt/model executor/model runner.py", line 1130, in init cuda graphs capture = capture cuda graphs( ^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/model executor/model runner components/cuda graph setup.py", line 412, in capture cuda graphs decode = capture decode graph(model runner=model runner) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/model executor/model runner components/cuda graph setup.py", line 737, in capture decode graph runner = graph runners[model runner.device](model runner) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/model executor/runner/decode cuda graph runner.py", line 513, in init raise Exception( Exception: Capture cuda graph failed: add(): functions with out=... arguments don't support automatic differentiation, but one of the arguments requires grad. Possible solutions: 1. set --mem-fraction-static to a smaller value (e.g., 0.8 or 0.7) 2. set --cuda-graph-max-bs-decode to a smaller value (e.g., 16) 3. disable decode CUDA graph by --cuda-graph-backend-decode=disabled. (Not recommended. Huge performance loss) Open an issue on GitHub https://github.com/sgl-project/sglang/issues/new/choose [2026-10-14 16:31:34] Received sigquit from a child process. It usually means the child failed. [2026-10-14 16:31:34] kill process tree called: parent pid=1, include parent=True, pid=1 [2026-10-14 16:31:34] kill process tree called: parent pid=1, include parent=False, pid=1 | — score breakdown | — |
| challenger | 5436bb1…bebf1 | scored | 0.1988 score breakdown | — |
| challenger | 5f6deea…2258a | scored | 0.1136 score breakdown | — |
21/21 steps
Setup
4/4Waiting on a provider to hand over a machine.
Installing the harness and shipping the repo.
Fetching the baseline and candidate engine images.
Staging the campaign model onto the pod.
Baseline
2/2Loading weights and waiting for the server to answer.
Second baseline
2/2Loading weights and waiting for the server to answer.
Candidate 1 of 5
2/2Loading weights and waiting for the server to answer.
Candidate 2 of 5
2/2Loading weights and waiting for the server to answer.
Candidate 3 of 5
2/2Loading weights and waiting for the server to answer.
Candidate 4 of 5
2/2Loading weights and waiting for the server to answer.
Candidate 5 of 5
2/2Loading weights and waiting for the server to answer.
Scorer
2/2Loading weights and waiting for the server to answer.
Teardown
1/1Destroying the pod and its volume.
Full workload replay against the latency gates.
Full workload replay against the latency gates.
Full workload replay against the latency gates.
Full workload replay against the latency gates.
Full workload replay against the latency gates.
Full workload replay against the latency gates.
Full workload replay against the latency gates.
The pinned baseline image teacher-forces captured outputs against the campaign thresholds.