Round 2

complete

Campaign: RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHeadRound id:

Entries

4 scored · 3 disqualified

Settled in

2h 17m

seated Sep 17, 12:46 UTC

Top score

0.5099

0 is baseline speed

Leader

Changed

won by 2411df5a…ce7995

Entries

Role	Miner	Patch	Status	Score	Duration	Open
baseline	—	—	scored	0 score breakdown	—

Round 2

challenger

4c3fe52…98761

disqualifiedengine died before becoming healthy (last error: [Errno 111] Connection refused); container log tail: m tokens) File "/src/python/sglang/srt/model executor/runner/prefill cuda graph runner.py", line 1546, in capture one shape self.backend.capture one( File "/src/python/sglang/srt/model executor/runner backend/breakable cuda graph backend.py", line 123, in capture one warmup out = forward fn() ^^^^^^^^^^^^ File "/src/python/sglang/srt/model executor/runner/prefill cuda graph runner.py", line 1530, in run once return self. run forward(forward batch, num tokens) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/utils/ contextlib.py", line 124, in decorate context return func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/model executor/runner/prefill cuda graph runner.py", line 756, in run forward return self.layer model.forward( ^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/utils/ contextlib.py", line 124, in decorate context return func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/models/qwen3 5.py", line 1723, in forward hidden states, residual = layer( ^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1778, in wrapped call impl return self. call impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1789, in call impl return forward call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/models/qwen3 5.py", line 1457, in forward hidden states = self.mlp(hidden states) ^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1778, in wrapped call impl return self. call impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1789, in call impl return forward call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/models/qwen2 moe.py", line 254, in forward gate up, = self.gate up proj(x) ^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1778, in wrapped call impl return self. call impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1789, in call impl return forward call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/layers/linear.py", line 505, in forward output parallel = self.quant method.apply(self, input , bias) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/layers/quantization/unquant.py", line 433, in apply return F.linear(x, layer.weight, bias) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 136.00 MiB. GPU 1 has a total capacity of 31.36 GiB of which 99.06 MiB is free. Including non-PyTorch memory, this process has 31.10 GiB memory in use. Of the allocated memory 28.19 GiB is allocated by PyTorch, with 959.50 MiB allocated in private pools (e.g., CUDA Graphs), and 354.27 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH CUDA ALLOC CONF=expandable segments:True to avoid fragmentation. See documentation for Memory Management (https://docs.pytorch.org/docs/stable/notes/cuda.html#optimizing-memory-usage-with-pytorch-cuda-alloc-conf) [2026-09-17 14:05:37] Received sigquit from a child process. It usually means the child failed. [2026-09-17 14:05:37] kill process tree called: parent pid=1, include parent=True, pid=1 [2026-09-17 14:05:37] kill process tree called: parent pid=1, include parent=False, pid=1

—

challenger

76cd598…2eaa3

disqualifiedengine died before becoming healthy (last error: [Errno 111] Connection refused); container log tail: m tokens) File "/src/python/sglang/srt/model executor/runner/prefill cuda graph runner.py", line 1546, in capture one shape self.backend.capture one( File "/src/python/sglang/srt/model executor/runner backend/breakable cuda graph backend.py", line 123, in capture one warmup out = forward fn() ^^^^^^^^^^^^ File "/src/python/sglang/srt/model executor/runner/prefill cuda graph runner.py", line 1530, in run once return self. run forward(forward batch, num tokens) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/utils/ contextlib.py", line 124, in decorate context return func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/model executor/runner/prefill cuda graph runner.py", line 756, in run forward return self.layer model.forward( ^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/utils/ contextlib.py", line 124, in decorate context return func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/models/qwen3 5.py", line 1723, in forward hidden states, residual = layer( ^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1778, in wrapped call impl return self. call impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1789, in call impl return forward call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/models/qwen3 5.py", line 1457, in forward hidden states = self.mlp(hidden states) ^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1778, in wrapped call impl return self. call impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1789, in call impl return forward call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/models/qwen2 moe.py", line 254, in forward gate up, = self.gate up proj(x) ^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1778, in wrapped call impl return self. call impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1789, in call impl return forward call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/layers/linear.py", line 505, in forward output parallel = self.quant method.apply(self, input , bias) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/layers/quantization/unquant.py", line 433, in apply return F.linear(x, layer.weight, bias) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 136.00 MiB. GPU 3 has a total capacity of 31.36 GiB of which 97.06 MiB is free. Including non-PyTorch memory, this process has 31.11 GiB memory in use. Of the allocated memory 28.19 GiB is allocated by PyTorch, with 959.50 MiB allocated in private pools (e.g., CUDA Graphs), and 354.27 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH CUDA ALLOC CONF=expandable segments:True to avoid fragmentation. See documentation for Memory Management (https://docs.pytorch.org/docs/stable/notes/cuda.html#optimizing-memory-usage-with-pytorch-cuda-alloc-conf) [2026-09-17 14:11:47] Received sigquit from a child process. It usually means the child failed. [2026-09-17 14:11:47] kill process tree called: parent pid=1, include parent=True, pid=1 [2026-09-17 14:11:47] kill process tree called: parent pid=1, include parent=False, pid=1

—

Run

23/23 steps

Setup

4/4

Renting a GPU pod
Waiting on a provider to hand over a machine.
Recorded
Not recorded
Since previous step
—
Phase
provisioning
Preparing the pod
Installing the harness and shipping the repo.
Recorded
Not recorded
Since previous step
—
Phase
bootstrapping
Pulling engine images
Fetching the baseline and candidate engine images.
Recorded
Not recorded
Since previous step
—
Phase
pulling_image
Downloading model weights
Staging the campaign model onto the pod.
Recorded
Not recorded
Since previous step
—
Phase
downloading_model

Baseline

2/2

Baseline, starting the engine
Loading weights and waiting for the server to answer.
Recorded
Not recorded
Since previous step
—
Phase
starting_engine
Baseline, running SLA15:03:46

Incumbent

2/2

Incumbent, starting the engine
Loading weights and waiting for the server to answer.
Recorded
Not recorded
Since previous step
—
Phase
starting_engine
Incumbent, running SLA15:03:46

Candidate 1 of 5

2/2

Candidate 1 of 5, starting the engine
Loading weights and waiting for the server to answer.
Recorded
Not recorded
Since previous step
—
Phase
starting_engine
Candidate 1 of 5, running SLA15:03:46

Candidate 2 of 5

2/2

Candidate 2 of 5, starting the engine
Loading weights and waiting for the server to answer.
Recorded
Not recorded
Since previous step
—
Phase
starting_engine
Candidate 2 of 5, running SLA15:03:46

Candidate 3 of 5

2/2

Candidate 3 of 5, starting the engine
Loading weights and waiting for the server to answer.
Recorded
Not recorded
Since previous step
—
Phase
starting_engine
Candidate 3 of 5, running SLA15:03:46

Candidate 4 of 5

2/2

Candidate 4 of 5, starting the engine
Loading weights and waiting for the server to answer.
Recorded
Not recorded
Since previous step
—
Phase
starting_engine
Candidate 4 of 5, running SLA15:03:46

Candidate 5 of 5

2/2

Candidate 5 of 5, starting the engine
Loading weights and waiting for the server to answer.
Recorded
Not recorded
Since previous step
—
Phase
starting_engine
Candidate 5 of 5, running SLA15:03:46

Scorer

2/2

Scorer, starting the engine
Loading weights and waiting for the server to answer.
Recorded
Not recorded
Since previous step
—
Phase
starting_engine
Scorer, running the shared scorer

Drift baseline

2/2

Drift baseline, starting the engine
Loading weights and waiting for the server to answer.
Recorded
Not recorded
Since previous step
—
Phase
starting_engine
Drift baseline, running SLA

Teardown

1/1

Releasing the pod15:03:46
Destroying the pod and its volume.
Recorded
Sep 17, 2026, 15:03 UTC
Since previous step
0ms
Phase
teardown