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

# Aggregated serving on one Blackwell GPU, block-FP8 or NVFP4 checkpoint.
# For the BF16 checkpoint use tensor_parallel_size: 2 and moe_config.backend: CUTLASS.
trust_remote_code: true
max_batch_size: 16
enable_chunked_prefill: true
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
speculative_config:
  decoding_type: MTP
  max_draft_len: 3