TensorRT-LLMs/examples/layer_wise_benchmarks/config_ctx.yaml
2025-12-04 13:41:15 +08:00

20 lines
340 B
YAML

model: nvidia/DeepSeek-R1-0528-FP4-v2
layer_indices: [5]
run_type: CTX
scaled_from: null
# KV cache related args
tokens_per_block: 32
enable_attention_dp: true
# Model init args
moe_backend: CUTLASS
use_cuda_graph: false
# Per iteration args
batch_size: 1
seq_len_q: 8193
seq_len_kv_cache: 0
balance_method: Balanced
balance_ratio: null