TensorRT-LLMs/tests/scripts/perf-sanity/l0_gb200_multi_nodes.yaml
chenfeiz0326 61745f034a
[https://nvbugs/5727481][ci] Fix Port Conflict in Perf-Sanity CI Test (#9896)
Signed-off-by: Chenfei Zhang <chenfeiz@nvidia.com>
2025-12-12 17:16:50 +08:00

72 lines
1.7 KiB
YAML

# Hardware Config
hardware:
gpus_per_node: 4
gpus_per_server: 8
server_configs:
- name: "r1_fp4_v2_dep8_mtp1"
model_name: "deepseek_r1_0528_fp4_v2"
gpus: 8
gpus_per_node: 4
trust_remote_code: true
tensor_parallel_size: 8
moe_expert_parallel_size: 8
pipeline_parallel_size: 1
max_batch_size: 512
max_num_tokens: 3136
attn_backend: "TRTLLM"
enable_attention_dp: true
attention_dp_config:
batching_wait_iters: 0
enable_balance: true
timeout_iters: 60
moe_config:
backend: 'CUTLASS'
cuda_graph_config:
enable_padding: true
max_batch_size: 512
kv_cache_config:
dtype: 'fp8'
enable_block_reuse: false
free_gpu_memory_fraction: 0.5
client_configs:
- name: "con32_iter12_1k1k"
concurrency: 32
iterations: 12
isl: 1024
osl: 1024
random_range_ratio: 0.2
backend: "openai"
- name: "r1_fp4_v2_tep8_mtp3"
model_name: "deepseek_r1_0528_fp4_v2"
gpus: 8
gpus_per_node: 4
trust_remote_code: true
tensor_parallel_size: 8
moe_expert_parallel_size: 8
pipeline_parallel_size: 1
max_batch_size: 512
max_num_tokens: 3136
attn_backend: "TRTLLM"
enable_attention_dp: false
moe_config:
backend: "TRTLLM"
cuda_graph_config:
enable_padding: true
max_batch_size: 512
kv_cache_config:
dtype: 'fp8'
free_gpu_memory_fraction: 0.5
speculative_config:
decoding_type: 'MTP'
num_nextn_predict_layers: 3
client_configs:
- name: "con32_iter12_1k1k"
concurrency: 32
iterations: 12
isl: 1024
osl: 1024
random_range_ratio: 0.2
backend: "openai"