TensorRT-LLMs/tests/scripts/perf-sanity/l0_dgx_b200.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

294 lines
7.3 KiB
YAML

server_configs:
- name: "r1_fp8_dep8_mtp1_1k1k"
model_name: "deepseek_r1_0528_fp8"
gpus: 8
tensor_parallel_size: 8
moe_expert_parallel_size: 8
pipeline_parallel_size: 1
max_batch_size: 512
max_num_tokens: 8192
attn_backend: "TRTLLM"
enable_attention_dp: true
attention_dp_config:
batching_wait_iters: 0
enable_balance: true
timeout_iters: 60
moe_config:
backend: 'DEEPGEMM'
cuda_graph_config:
enable_padding: true
max_batch_size: 512
kv_cache_config:
dtype: 'fp8'
enable_block_reuse: false
free_gpu_memory_fraction: 0.8
speculative_config:
decoding_type: 'MTP'
num_nextn_predict_layers: 1
client_configs:
- name: "con4096_iter10_1k1k"
concurrency: 4096
iterations: 10
isl: 1024
osl: 1024
random_range_ratio: 0.2
backend: "openai"
- name: "r1_fp8_tep8_mtp3_1k1k"
model_name: "deepseek_r1_0528_fp8"
gpus: 8
tensor_parallel_size: 8
moe_expert_parallel_size: 8
pipeline_parallel_size: 1
max_batch_size: 64
max_num_tokens: 8192
attn_backend: "TRTLLM"
enable_attention_dp: false
moe_config:
backend: 'DEEPGEMM'
cuda_graph_config:
enable_padding: true
max_batch_size: 64
kv_cache_config:
dtype: 'fp8'
enable_block_reuse: false
free_gpu_memory_fraction: 0.8
speculative_config:
decoding_type: 'MTP'
num_nextn_predict_layers: 3
client_configs:
- name: "con64_iter10_1k1k"
concurrency: 64
iterations: 10
isl: 1024
osl: 1024
random_range_ratio: 0.2
backend: "openai"
- name: "r1_fp8_tp8_mtp3_1k1k"
model_name: "deepseek_r1_0528_fp8"
gpus: 8
tensor_parallel_size: 8
moe_expert_parallel_size: 1
pipeline_parallel_size: 1
max_batch_size: 8
max_num_tokens: 8192
attn_backend: "TRTLLM"
enable_attention_dp: false
moe_config:
backend: 'TRTLLM'
cuda_graph_config:
enable_padding: true
max_batch_size: 8
kv_cache_config:
dtype: 'fp8'
enable_block_reuse: false
free_gpu_memory_fraction: 0.8
speculative_config:
decoding_type: 'MTP'
num_nextn_predict_layers: 3
client_configs:
- name: "con8_iter10_1k1k"
concurrency: 8
iterations: 10
isl: 1024
osl: 1024
random_range_ratio: 0.2
backend: "openai"
- name: "r1_fp4_v2_dep4_mtp1_1k1k"
model_name: "deepseek_r1_0528_fp4_v2"
gpus: 4
tensor_parallel_size: 4
moe_expert_parallel_size: 4
pipeline_parallel_size: 1
max_batch_size: 512
max_num_tokens: 8192
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: 0.8
speculative_config:
decoding_type: 'MTP'
num_nextn_predict_layers: 1
client_configs:
- name: "con2048_iter10_1k1k"
concurrency: 2048
iterations: 10
isl: 1024
osl: 1024
random_range_ratio: 0.2
backend: "openai"
- name: "r1_fp4_v2_tep4_mtp3_1k1k"
model_name: "deepseek_r1_0528_fp4_v2"
gpus: 4
tensor_parallel_size: 4
moe_expert_parallel_size: 4
pipeline_parallel_size: 1
max_batch_size: 32
max_num_tokens: 8192
attn_backend: "TRTLLM"
enable_attention_dp: false
moe_config:
backend: 'TRTLLM'
cuda_graph_config:
enable_padding: true
max_batch_size: 32
kv_cache_config:
dtype: 'fp8'
enable_block_reuse: false
free_gpu_memory_fraction: 0.8
speculative_config:
decoding_type: 'MTP'
num_nextn_predict_layers: 3
client_configs:
- name: "con32_iter10_1k1k"
concurrency: 32
iterations: 10
isl: 1024
osl: 1024
random_range_ratio: 0.2
backend: "openai"
- name: "r1_fp4_v2_tp4_mtp3_1k1k"
model_name: "deepseek_r1_0528_fp4_v2"
gpus: 4
tensor_parallel_size: 4
moe_expert_parallel_size: 1
pipeline_parallel_size: 1
max_batch_size: 4
max_num_tokens: 8192
attn_backend: "TRTLLM"
enable_attention_dp: false
moe_config:
backend: 'TRTLLM'
cuda_graph_config:
enable_padding: true
max_batch_size: 4
kv_cache_config:
dtype: 'fp8'
enable_block_reuse: false
free_gpu_memory_fraction: 0.8
speculative_config:
decoding_type: 'MTP'
num_nextn_predict_layers: 3
client_configs:
- name: "con4_iter10_1k1k"
concurrency: 4
iterations: 10
isl: 1024
osl: 1024
random_range_ratio: 0.2
backend: "openai"
- name: "gpt_oss_fp4_dep2_1k1k"
model_name: "gpt_oss_120b_fp4"
gpus: 2
tensor_parallel_size: 2
moe_expert_parallel_size: 2
pipeline_parallel_size: 1
max_batch_size: 1024
max_num_tokens: 20000
attn_backend: "TRTLLM"
enable_attention_dp: true
attention_dp_config:
enable_balance: true
moe_config:
backend: 'TRTLLM'
cuda_graph_config:
enable_padding: true
max_batch_size: 1024
kv_cache_config:
dtype: 'fp8'
enable_block_reuse: false
free_gpu_memory_fraction: 0.8
num_postprocess_workers: 4
stream_interval: 20
client_configs:
- name: "con2048_iter5_1k1k"
concurrency: 2048
iterations: 5
isl: 1024
osl: 1024
random_range_ratio: 0.2
backend: "openai"
- name: "gpt_oss_fp4_dep4_1k1k"
model_name: "gpt_oss_120b_fp4"
gpus: 4
tensor_parallel_size: 4
moe_expert_parallel_size: 4
pipeline_parallel_size: 1
max_batch_size: 512
max_num_tokens: 20000
attn_backend: "TRTLLM"
enable_attention_dp: true
attention_dp_config:
enable_balance: true
moe_config:
backend: 'TRTLLM'
cuda_graph_config:
enable_padding: true
max_batch_size: 512
kv_cache_config:
dtype: 'fp8'
enable_block_reuse: false
free_gpu_memory_fraction: 0.8
num_postprocess_workers: 4
stream_interval: 20
client_configs:
- name: "con2048_iter5_1k1k"
concurrency: 2048
iterations: 5
isl: 1024
osl: 1024
random_range_ratio: 0.2
backend: "openai"
- name: "gpt_oss_fp4_tp4_eagle3_1k1k"
model_name: "gpt_oss_120b_fp4"
gpus: 4
tensor_parallel_size: 4
moe_expert_parallel_size: 1
pipeline_parallel_size: 1
max_batch_size: 1
max_num_tokens: 20000
attn_backend: "TRTLLM"
enable_attention_dp: false
moe_config:
backend: 'TRTLLM'
cuda_graph_config:
enable_padding: true
max_batch_size: 1
kv_cache_config:
dtype: 'fp8'
enable_block_reuse: false
free_gpu_memory_fraction: 0.8
speculative_config:
decoding_type: 'Eagle'
eagle3_layers_to_capture: [-1]
max_draft_len: 3
speculative_model_dir: "gpt_oss/gpt-oss-120b-Eagle3"
stream_interval: 20
num_postprocess_workers: 4
client_configs:
- name: "con1_iter32_1k1k"
concurrency: 1
iterations: 32
isl: 1024
osl: 1024
random_range_ratio: 0.2
backend: "openai"