[TRTLLM-6541][test] Add NIM perf test cases (#7924)

Signed-off-by: FredricZ-2007 <[email protected]>

Author: fredricz-20070104

tests/integration/test_lists/qa/llm_perf_cluster_nim.yml

@@ -0,0 +1,141 @@
+version: 0.0.1
+llm_perf_cluster_nim:
+- condition:
+    ranges:
+      system_gpu_count:
+        gte: 1
+  tests:
+  - perf/test_perf.py::test_perf[llama_v3.1_8b_instruct-bench-bfloat16-input_output_len:128,128]
+  - perf/test_perf.py::test_perf[llama_v3.1_8b_instruct-bench-pytorch-bfloat16-input_output_len:128,128]
+  - perf/test_perf.py::test_perf[llama_v3.1_8b-bench-bfloat16-maxbs:256-input_output_len:128,128-quant:fp8]
+  - perf/test_perf.py::test_perf[llama_v3.1_8b-bench-bfloat16-maxbs:256-input_output_len:512,32-quant:fp8]
+  - perf/test_perf.py::test_perf[llama_v3.1_8b_instruct_fp8-bench-pytorch-float8-input_output_len:128,128]
+  - perf/test_perf.py::test_perf[llama_v3.1_8b_instruct_fp8-bench-pytorch-float8-input_output_len:2000,500]
+  - perf/test_perf.py::test_perf[llama_v3.1_8b_instruct_fp8-bench-pytorch-streaming-float8-input_output_len:2000,500]
+  - perf/test_perf.py::test_perf[llama_v3.1_8b_instruct_fp8-bench-pytorch-float8-input_output_len:500,2000]
+  - perf/test_perf.py::test_perf[mistral_7b_v0.1-bench-float16-input_output_len:1000,1000-quant:fp8]
+  - perf/test_perf.py::test_perf[mistral_7b_v0.1-bench-float16-input_output_len:500,2000-quant:fp8]
+  - perf/test_perf.py::test_perf[deepseek_v3_lite_nvfp4-bench-pytorch-streaming-float4-maxbs:2048-maxnt:8192-input_output_len:256,256-reqs:200]
+  # for chunked prefill cases
+  - perf/test_perf.py::test_perf[deepseek_v3_lite_nvfp4-bench-pytorch-float4-maxbs:512-maxnt:2048-kv_frac:0.85-input_output_len:5000,500-reqs:200]
+  # Phi-4-multimodal-instruct
+  - perf/test_perf.py::test_perf[phi_4_multimodal_instruct-bench-pytorch-bfloat16-input_output_len:500,2000-con:250]
+  - perf/test_perf.py::test_perf[phi_4_multimodal_instruct-bench-pytorch-bfloat16-input_output_len:1000,1000-con:250]
+  - perf/test_perf.py::test_perf[phi_4_multimodal_instruct-bench-pytorch-bfloat16-input_output_len:128,128]
+  - perf/test_perf.py::test_perf[phi_4_multimodal_instruct-bench-pytorch-bfloat16-input_output_len:512,32]
+  - perf/test_perf.py::test_perf[phi_4_multimodal_instruct_image-bench-pytorch-bfloat16-input_output_len:1000,1000-loras:1-con:250]
+  - perf/test_perf.py::test_perf[phi_4_multimodal_instruct_audio-bench-pytorch-bfloat16-input_output_len:1000,1000-loras:1-con:250]
+  #Mistral-Small-3.1-24B-Instruct-2503
+  - perf/test_perf.py::test_perf[mistral_small_v3.1_24b-bench-pytorch-bfloat16-maxbs:1-input_output_len:1000,2000-reqs:8-con:1]
+  - perf/test_perf.py::test_perf[mistral_small_v3.1_24b-bench-pytorch-bfloat16-input_output_len:1000,2000-reqs:500-con:200]
+  - perf/test_perf.py::test_perf[mistral_small_v3.1_24b-bench-pytorch-bfloat16-maxbs:1-maxnt:20000-input_output_len:20000,2000-reqs:8-con:1] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[mistral_small_v3.1_24b-bench-pytorch-bfloat16-maxbs:4096-maxnt:20000-input_output_len:20000,2000-reqs:300-con:200] TIMEOUT(120)
+
+
+- condition:
+    ranges:
+      system_gpu_count:
+        gte: 2
+  tests:
+  - perf/test_perf.py::test_perf[llama_v3.1_8b-bench-bfloat16-input_output_len:128,128-quant:nvfp4-gpus:2]
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp8-bench-float8-maxbs:256-input_output_len:128,128-gpus:2]
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp8-bench-float8-maxbs:256-input_output_len:512,32-gpus:2]
+  - perf/test_perf.py::test_perf[llama_v3.1_70b_instruct_fp8-bench-pytorch-float8-maxbs:256-input_output_len:128,128-gpus:2]
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp8-bench-pytorch-float8-maxbs:256-input_output_len:512,32-gpus:2]
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp8-bench-pytorch-streaming-float8-maxbs:256-input_output_len:512,32-gpus:2]
+  - perf/test_perf.py::test_perf[llama_v2_13b-bench-float16-input_output_len:128,128-loras:8-gpus:2]
+  #Mistral-Small-3.1-24B-Instruct-2503
+  - perf/test_perf.py::test_perf[mistral_small_v3.1_24b-bench-pytorch-bfloat16-maxbs:1-input_output_len:1000,2000-reqs:8-con:1-gpus:2]
+  - perf/test_perf.py::test_perf[mistral_small_v3.1_24b-bench-pytorch-bfloat16-input_output_len:1000,2000-reqs:500-con:200-gpus:2]
+  - perf/test_perf.py::test_perf[mistral_small_v3.1_24b-bench-pytorch-bfloat16-maxbs:1-maxnt:20000-input_output_len:20000,2000-reqs:8-con:1-gpus:2] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[mistral_small_v3.1_24b-bench-pytorch-bfloat16-maxbs:4096-maxnt:20000-input_output_len:20000,2000-reqs:300-con:200-gpus:2] TIMEOUT(120)
+
+# Tests for systems with 4+ GPUs
+- condition:
+    ranges:
+      system_gpu_count:
+        gte: 4
+  tests:
+  - perf/test_perf.py::test_perf[starcoder_15b-bench-float16-input_output_len:512,200-gpus:4]
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-float4-maxbs:512-input_output_len:128,128-ep:4-tp:4-gpus:4]
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-streaming-float4-maxbs:512-input_output_len:128,128-ep:4-tp:4-gpus:4]
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-float4-kv_frac:0.85-input_output_len:1000,1000-reqs:2000-ep:4-tp:4-gpus:4] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-float4-kv_frac:0.85-input_output_len:1000,2000-reqs:3000-ep:4-tp:4-gpus:4] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-float4-maxbs:1000-maxnt:5000-kv_frac:0.85-input_output_len:5000,500-reqs:20000-ep:4-tp:4-gpus:4] TIMEOUT(120)
+  # for chunked prefill cases
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-float4-maxbs:512-maxnt:2048-kv_frac:0.85-input_output_len:5000,500-reqs:200-ep:4-tp:4-gpus:4] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-float4-maxbs:256-maxnt:1024-kv_frac:0.85-input_output_len:2000,2000-reqs:200-ep:4-tp:4-gpus:4] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-float4-maxbs:1000-maxnt:5000-kv_frac:0.85-input_output_len:5000,500-reqs:2000-ep:4-tp:4-gpus:4] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[qwen3_235b_a22b_fp4-bench-pytorch-float4-input_output_len:1000,2000-con:512-ep:4-gpus:4]
+  #llama_v3.1_405b_instruct_fp4
+  - perf/test_perf.py::test_perf[llama_v3.1_405b_instruct_fp4-bench-pytorch-float4-kv_frac:0.85-input_output_len:128,128-gpus:4]
+  - perf/test_perf.py::test_perf[llama_v3.1_405b_instruct_fp4-bench-pytorch-float4-kv_frac:0.85-input_output_len:1000,1000-gpus:4]
+  - perf/test_perf.py::test_perf[llama_v3.1_405b_instruct_fp4-bench-pytorch-float4-maxbs:1024-maxnt:4096-kv_frac:0.85-input_output_len:1024,2048-gpus:4]
+  - perf/test_perf.py::test_perf[llama_v3.1_405b_instruct_fp4-bench-pytorch-float4-maxbs:4096-maxnt:20000-kv_frac:0.85-input_output_len:20000,2000-gpus:4] TIMEOUT(120)
+  #llama_v3.3_70b_instruct_fp4
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp4-bench-pytorch-float4-input_output_len:128,128-gpus:4]
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp4-bench-pytorch-float4-input_output_len:512,32-gpus:4]
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp4-bench-pytorch-float4-maxbs:1024-maxnt:4096-kv_frac:0.85-input_output_len:1000,1000-reqs:1000-gpus:4]
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp4-bench-pytorch-float4-maxbs:4096-maxnt:20000-kv_frac:0.85-input_output_len:20000,2000-reqs:200-gpus:4] TIMEOUT(120)
+  #llama_v4_scout_17b_16e_instruct_fp4
+  - perf/test_perf.py::test_perf[llama_v4_scout_17b_16e_instruct_fp4-bench-pytorch-float4-input_output_len:128,128-gpus:4]
+  - perf/test_perf.py::test_perf[llama_v4_scout_17b_16e_instruct_fp4-bench-pytorch-float4-maxbs:1024-maxnt:4096-kv_frac:0.85-input_output_len:500,2000-reqs:500-gpus:4]
+  - perf/test_perf.py::test_perf[llama_v4_scout_17b_16e_instruct_fp4-bench-pytorch-float4-maxbs:1024-maxnt:4096-kv_frac:0.85-input_output_len:1000,1000-reqs:500-gpus:4]
+  - perf/test_perf.py::test_perf[llama_v4_scout_17b_16e_instruct_fp4-bench-pytorch-float4-maxbs:4096-maxnt:20000-kv_frac:0.85-input_output_len:20000,2000-reqs:200-gpus:4] TIMEOUT(120)
+
+
+# Tests for systems with 8+ GPUs
+- condition:
+    ranges:
+      system_gpu_count:
+        gte: 8
+  tests:
+  #llama_v3.3_nemotron_super_49b
+  - perf/test_perf.py::test_perf[llama_v3.3_nemotron_super_49b-bench-bfloat16-input_output_len:500,2000-quant:fp8-con:250-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v3.3_nemotron_super_49b_fp8-bench-pytorch-bfloat16-input_output_len:500,2000-con:250-gpus:8]
+  #llama_v3.3_70b_instruct_fp4
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp4-bench-pytorch-float4-maxbs:1024-maxnt:4096-kv_frac:0.85-input_output_len:500,2000-reqs:3000-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp4-bench-pytorch-float4-maxbs:1024-maxnt:4096-kv_frac:0.85-input_output_len:1000,1000-reqs:3000-tp:8-gpus:8]
+
+  #llama_v4_scout_17b_16e_instruct_fp4
+  - perf/test_perf.py::test_perf[llama_v4_scout_17b_16e_instruct_fp4-bench-pytorch-float4-kv_frac:0.85-input_output_len:128,128-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v4_scout_17b_16e_instruct_fp4-bench-pytorch-float4-kv_frac:0.85-input_output_len:512,32-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v4_scout_17b_16e_instruct_fp4-bench-pytorch-float4-maxbs:1024-maxnt:4096-kv_frac:0.85-input_output_len:500,2000-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v4_scout_17b_16e_instruct_fp4-bench-pytorch-float4-maxbs:1024-maxnt:4096-kv_frac:0.85-input_output_len:1000,1000-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[gpt_20b-bench-float16-maxbs:8-input_output_len:128,128-reqs:80-gpus:8]
+  - perf/test_perf.py::test_perf[gpt_20b-bench-float16-maxbs:8-input_output_len:512,32-reqs:80-gpus:8]
+  #deepseek_r1_fp8
+  - perf/test_perf.py::test_perf[deepseek_r1_fp8-bench-pytorch-float8-maxbs:512-input_output_len:128,128-ep:8-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[deepseek_r1_fp8-bench-pytorch-float8-maxbs:1-input_output_len:1000,2000-reqs:10-ep:4-tp:8-gpus:8] #min latency test
+  - perf/test_perf.py::test_perf[deepseek_r1_fp8-bench-pytorch-float8-maxbs:384-maxnt:1536-input_output_len:1000,2000-reqs:49152-con:3072-ep:8-tp:8-gpus:8] #max throughput test
+  #deepseek_r1_nvfp4
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-float4-maxbs:512-input_output_len:128,128-ep:8-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-float4-maxbs:1-input_output_len:1000,2000-reqs:10-ep:4-tp:8-gpus:8] #min latency test
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-streaming-float4-maxbs:1-input_output_len:1000,2000-reqs:10-ep:4-tp:8-gpus:8] #min latency test
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-float4-maxbs:384-maxnt:1536-input_output_len:1000,2000-reqs:49152-con:3072-ep:8-tp:8-gpus:8] TIMEOUT (120) #max throughput test
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-streaming-float4-maxbs:384-maxnt:1536-input_output_len:1000,2000-reqs:49152-con:3072-ep:8-tp:8-gpus:8] #max throughput test
+  # for chunked prefill cases
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-float4-maxbs:512-maxnt:2048-kv_frac:0.85-input_output_len:5000,500-reqs:200-ep:8-tp:8-gpus:8] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[deepseek_r1_nvfp4-bench-pytorch-float4-maxbs:256-maxnt:1024-kv_frac:0.85-input_output_len:2000,2000-reqs:200-ep:8-tp:8-gpus:8] TIMEOUT(120)
+  #deepseek_r1_0528_fp4
+  - perf/test_perf.py::test_perf[deepseek_r1_0528_fp4-bench-pytorch-float4-kv_frac:0.85-input_output_len:1000,1000-reqs:20000-ep:8-tp:8-gpus:8] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[deepseek_r1_0528_fp4-bench-pytorch-float4-kv_frac:0.85-input_output_len:1000,2000-reqs:3000-ep:8-tp:8-gpus:8] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[deepseek_r1_0528_fp4-bench-pytorch-float4-maxbs:1000-maxnt:5000-kv_frac:0.85-input_output_len:5000,500-reqs:20000-ep:4-tp:4-gpus:4] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[llama_v4_maverick_17b_128e_instruct-bench-pytorch-bfloat16-input_output_len:128,128-ep:8-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v4_maverick_17b_128e_instruct-bench-pytorch-bfloat16-input_output_len:500,2000-ep:8-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v4_maverick_17b_128e_instruct-bench-pytorch-bfloat16-input_output_len:2000,500-ep:8-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v4_maverick_17b_128e_instruct-bench-pytorch-streaming-bfloat16-input_output_len:2000,500-ep:8-tp:8-gpus:8] TIMEOUT (40)
+  - perf/test_perf.py::test_perf[llama_v4_maverick_17b_128e_instruct_fp8-bench-pytorch-float8-input_output_len:128,128-ep:8-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v4_maverick_17b_128e_instruct_fp8-bench-pytorch-float8-input_output_len:500,2000-ep:8-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v4_maverick_17b_128e_instruct_fp8-bench-pytorch-float8-input_output_len:2000,500-ep:8-tp:8-gpus:8] TIMEOUT (40)
+  - perf/test_perf.py::test_perf[llama_v4_maverick_17b_128e_instruct_fp8-bench-pytorch-streaming-float8-input_output_len:2000,500-ep:8-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[qwen3_235b_a22b_fp4-bench-pytorch-float4-input_output_len:1000,2000-con:8-ep:8-tp:8-gpus:8]
+  #gpt_oss_120b
+  # max throughput test
+  - perf/test_perf.py::test_perf[gpt_oss_120b_fp4-bench-pytorch-float4-maxbs:720-maxnt:16384-input_output_len:1024,1024-reqs:1280-con:256-ep:8-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[gpt_oss_120b_fp4-bench-pytorch-float4-maxbs:720-maxnt:16384-input_output_len:1024,1024-reqs:2560-con:512-ep:8-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[gpt_oss_120b_fp4-bench-pytorch-float4-maxbs:720-maxnt:16384-input_output_len:1024,1024-reqs:5120-con:1024-ep:8-tp:8-gpus:8] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[gpt_oss_120b_fp4-bench-pytorch-float4-maxbs:720-maxnt:16384-input_output_len:1024,1024-reqs:20480-con:4096-ep:8-tp:8-gpus:8] TIMEOUT(180)
+  # min latency test
+  - perf/test_perf.py::test_perf[gpt_oss_120b_fp4-bench-pytorch-float4-maxbs:720-maxnt:16384-input_output_len:1024,1024-reqs:8-con:1-ep:8-tp:8-gpus:8]
+  - perf/test_perf.py::test_perf[gpt_oss_120b_fp4-bench-pytorch-float4-maxbs:720-maxnt:16384-input_output_len:1024,1024-reqs:100-con:32-ep:8-tp:8-gpus:8]

tests/integration/test_lists/qa/llm_perf_nim.yml

@@ -150,7 +150,6 @@ llm_perf_nim:
   - perf/test_perf.py::test_perf[llama_v3.2_1b-bench-bfloat16-input_output_len:128,128-gpus:2]
   #t5
   - perf/test_perf.py::test_perf[t5-bench-float16-input_output_len:128,20-gpus:2]
-  - perf/test_perf.py::test_perf[flan_t5_large-bench-float16-input_output_len:128,20-gpus:2]
 
 - condition:
     ranges:
@@ -168,7 +167,6 @@ llm_perf_nim:
   #llama_v3.1_70b
   #trt backend
   - perf/test_perf.py::test_perf[llama_v3.1_70b-bench-bfloat16-input_output_len:1024,1024-tp:2-gpus:2]
-  - perf/test_perf.py::test_perf[llama_70b_sq_per_tensor-cppmanager-exe-plugin_ifb-float16-input_output_len:128,128+512,32-gpus:2]
   #mixtral_8x7b_v0.1
   #trt backend
   - perf/test_perf.py::test_perf[mixtral_8x7b-cppmanager-exe-plugin_ifb-float16-mp-input_output_len:128,128+512,32-gpus:2]
@@ -199,6 +197,8 @@ llm_perf_nim:
   #trt backend
   - perf/test_perf.py::test_perf[mistral_7b_v0.1-bench-float16-input_output_len:1000,1000-quant:fp8-tp:2]
   - perf/test_perf.py::test_perf[mistral_7b_v0.1-bench-float16-input_output_len:500,2000-quant:fp8-tp:2]
+  # torch backend
+  - perf/test_perf.py::test_perf[mistral_7b_v0.1-bench-pytorch-float16-input_output_len:128,128]
   #phi_3_mini_128k_instruct
   #trt backend
   - perf/test_perf.py::test_perf[phi_3_mini_128k_instruct-bench-float16-maxbs:128-input_output_len:1000,1000-quant:fp8-tp:2]