# Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

# Context (prefill) worker. Pair with qwen3.8-flash-next-disagg-gen.yaml.
# The handoff carries recurrent and PLE state, so both workers must set
# cache_transceiver_config.transceiver_runtime to PYTHON.
trust_remote_code: true
max_batch_size: 16
disable_overlap_scheduler: true
moe_config:
  backend: TRTLLM
kv_cache_config:
  free_gpu_memory_fraction: 0.5
  mamba_ssm_cache_dtype: bfloat16
cache_transceiver_config:
  backend: NIXL
  transceiver_runtime: PYTHON
  max_tokens_in_buffer: 8192