     36     resolve_host_drafts (arbi_serve/spec_decode/device_drafts.py:105)
      4     select (selectors.py:468)
      1     synchronize (torch/cuda/streams.py:254)
      1     sample_batched (arbi_serve/spec_decode/rejection_sampler.py:648)
      1     resolved_eff_drafts (arbi_serve/engine/step_plan.py:140)
      1     release (arbi_serve/spec_decode/dflash_kv_slots.py:280)
      1     refresh_slate (arbi_serve/spec_decode/dflash_realize_capture.py:313)
      1     _mtp_sample_residual_philox_kernel (arbi_serve/_custom_ops/spec_decode.py:316)
      1     forward (arbi_serve/models/layers.py:83)
      1     _assemble_fixed_context_device (arbi_serve/spec_decode/_dflash_driver_context.py:468)
