Hardware Benchmarks - Step-3.7-Flash

StepFun Step-3.7-Flash — hybrid SWA/global-attention thinking model with MTP speculative decoding. Requires a full 8-GPU node (TP=8).

Results

Engine

Accelerators

Comparing 2 setups — same config and hardware.

vllm-step3p7-flash-8xh200 (on-demand)
vllm-step3p7-flash-8xh200 (spot)

Leaderboard

#	chart	setup	hardware	config	tok/s	$/1M completion
1		vllm-step3p7-flash-8xh200 (on-demand)	vllm · H200:8	TP=8+7 Model `trust-remote-code` `true` Trust remote code when downloading model and tokenizer. Parallelism `enable-expert-parallel` `true` Use expert parallelism instead of tensor parallelism for MoE layers. `tensor-parallel-size` `8` Number of tensor-parallel groups. Other `disable_cascade_attn` `true` `enable_auto_tool_choice` `true` `reasoning_parser` `step3p5` `speculative_config` `{"method": "mtp", "num_speculative_tokens": 3}` `tool_call_parser` `step3p5`	24,184.7	$0.9277
2		vllm-step3p7-flash-8xh200 (spot)	vllm · H200:8	TP=8+7 Model `trust-remote-code` `true` Trust remote code when downloading model and tokenizer. Parallelism `enable-expert-parallel` `true` Use expert parallelism instead of tensor parallelism for MoE layers. `tensor-parallel-size` `8` Number of tensor-parallel groups. Other `disable_cascade_attn` `true` `enable_auto_tool_choice` `true` `reasoning_parser` `step3p5` `speculative_config` `{"method": "mtp", "num_speculative_tokens": 3}` `tool_call_parser` `step3p5`	19,452.6	$0.4372