Round score
—
engine 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
Eligible requests
—
trusted baseline exclusions removed
Failed eligible requests
—
each request counted once
Raw weighted speedup
—
before failure cap and deduction
Temperature range
Request concurrency
Tier scheduling
Output policy
Thinking
Context limit
The same eligible inputs and output ceilings apply to every engine. Candidates must emit at least 90% of baseline tokens per request; natural output lengths can affect tier completion time. Slots refill until the group drains and stay occupied through protocol completion. Baseline exclusions and final drain can reduce actual concurrency below the configured cap.
The round score uses weighted completion times for all eligible work in each tier. Per-request speedups below are diagnostics, not score contributions. Only trusted baseline exclusions are removed; candidate failures remain in the eligible count and incur the configured failure penalty.
No prompt scores
engine 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
Image digest
Patch hash
Engine