Round score
—
engine died before becoming healthy (last error: [Errno 111] Connection refused); container log tail: /src/python/sglang/launch_server.py:63: UserWarning: 'python -m sglang.launch_server' is still supported, but 'sglang serve' is the recommended entrypoint. Example: sglang serve --model-path <model> [options] warnings.warn( [2026-10-10 21:07:55] Attention backend not specified. Use flashinfer backend by default. [2026-10-10 21:07:55] Qwen3.5 GDN EAGLE topk=1 on SM120: defaulting to --enable-linear-replayssm-spec. [2026-10-10 21:07:55] Qwen3.5 GDN EAGLE topk=1 on SM120: defaulting to fp8_e4m3 KV. [2026-10-10 21:07:55] --enable-linear-replayssm-spec: setting --mamba-ssm-dtype float32 (cached replay uses compensated checkpoint projection and materialization). [2026-10-10 21:07:55] --mm-enable-dp-encoder is enabled with TP=1, so the encoder has no data-parallel work to distribute. Disable it unless you need to validate this configuration. [2026-10-10 21:07:55] Qwen3.5 GDN EAGLE topk=1 on SM120: defaulting to cost-aware adaptive draft depth [2, 3, 4, 5, 6, 7]. [2026-10-10 21:07:55] Adaptive spec up to 8 draft tokens: GDN ReplaySSM ring 32 -> 32. [2026-10-10 21:07:55] Rejection sampling is enabled for speculative decoding (speculative_use_rejection_sampling=True). [2026-10-10 21:07:57] kill_process_tree called: parent_pid=1, include_parent=False, pid=1 Traceback (most recent call last): File "<frozen runpy>", line 198, in _run_module_as_main File "<frozen runpy>", line 88, in _run_code File "/src/python/sglang/launch_server.py", line 76, in <module> run_server(server_args) File "/src/python/sglang/launch_server.py", line 59, in run_server launch_server(server_args) File "/src/python/sglang/srt/entrypoints/http_server.py", line 2825, in launch_server ) = Engine._launch_subprocesses( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/entrypoints/engine.py", line 1098, in _launch_subprocesses publish(server_args, role="tokenizer") File "/src/python/sglang/srt/runtime_context.py", line 1587, in publish _CONTEXT.set_server_args(server_args) File "/src/python/sglang/srt/runtime_context.py", line 1061, in set_server_args max_draft_tokens_of(server_args), ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/arg_groups/overrides.py", line 1968, in max_speculative_num_draft_tokens raise ValueError( ValueError: The speculative algorithm declared max_speculative_num_draft_tokens=8, below the configured speculative_num_draft_tokens=9.
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: /src/python/sglang/launch_server.py:63: UserWarning: 'python -m sglang.launch_server' is still supported, but 'sglang serve' is the recommended entrypoint. Example: sglang serve --model-path <model> [options] warnings.warn( [2026-10-10 21:07:55] Attention backend not specified. Use flashinfer backend by default. [2026-10-10 21:07:55] Qwen3.5 GDN EAGLE topk=1 on SM120: defaulting to --enable-linear-replayssm-spec. [2026-10-10 21:07:55] Qwen3.5 GDN EAGLE topk=1 on SM120: defaulting to fp8_e4m3 KV. [2026-10-10 21:07:55] --enable-linear-replayssm-spec: setting --mamba-ssm-dtype float32 (cached replay uses compensated checkpoint projection and materialization). [2026-10-10 21:07:55] --mm-enable-dp-encoder is enabled with TP=1, so the encoder has no data-parallel work to distribute. Disable it unless you need to validate this configuration. [2026-10-10 21:07:55] Qwen3.5 GDN EAGLE topk=1 on SM120: defaulting to cost-aware adaptive draft depth [2, 3, 4, 5, 6, 7]. [2026-10-10 21:07:55] Adaptive spec up to 8 draft tokens: GDN ReplaySSM ring 32 -> 32. [2026-10-10 21:07:55] Rejection sampling is enabled for speculative decoding (speculative_use_rejection_sampling=True). [2026-10-10 21:07:57] kill_process_tree called: parent_pid=1, include_parent=False, pid=1 Traceback (most recent call last): File "<frozen runpy>", line 198, in _run_module_as_main File "<frozen runpy>", line 88, in _run_code File "/src/python/sglang/launch_server.py", line 76, in <module> run_server(server_args) File "/src/python/sglang/launch_server.py", line 59, in run_server launch_server(server_args) File "/src/python/sglang/srt/entrypoints/http_server.py", line 2825, in launch_server ) = Engine._launch_subprocesses( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/entrypoints/engine.py", line 1098, in _launch_subprocesses publish(server_args, role="tokenizer") File "/src/python/sglang/srt/runtime_context.py", line 1587, in publish _CONTEXT.set_server_args(server_args) File "/src/python/sglang/srt/runtime_context.py", line 1061, in set_server_args max_draft_tokens_of(server_args), ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/src/python/sglang/srt/arg_groups/overrides.py", line 1968, in max_speculative_num_draft_tokens raise ValueError( ValueError: The speculative algorithm declared max_speculative_num_draft_tokens=8, below the configured speculative_num_draft_tokens=9.
Image digest
Patch hash
Engine