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

# Generation (decode) worker. Pair with qwen3.8-flash-next-disagg-ctx.yaml.
# Add speculative_config here, not on the context worker, to run MTP.
trust_remote_code: true
max_batch_size: 16
cuda_graph_config:
  enable_padding: true
  max_batch_size: 16
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