[None][test] add kv cache size in bench metric and fix failed cases (#7211)

Signed-off-by: ruodil <[email protected]>

Author: ruodil

tests/integration/defs/perf/test_perf.py

@@ -266,6 +266,8 @@ def import_allowed_perf_config():
     re.compile(r"Average time-to-first-token \[TTFT\] \(ms\):\s+([\d\.]+)"),
     PerfMetricType.OUTPUT_TOKEN_TIME:
     re.compile(r"Average time-per-output-token \[TPOT\] \(ms\):\s+([\d\.]+)"),
+    PerfMetricType.KV_CACHE_SIZE:
+    re.compile(r".*Allocated ([\d\.]+) GiB for max tokens in paged KV cache.*"),
 }
 # (Relative threshold, Absolute threshold) for all metric types
 PERF_METRIC_THRESHOLD = {
@@ -325,6 +327,7 @@ def import_allowed_perf_config():
     PerfMetricType.INFERENCE_TIME,
     PerfMetricType.TOKEN_THROUGHPUT,
     PerfMetricType.SEQ_THROUGHPUT,
+    PerfMetricType.KV_CACHE_SIZE,
 ]
 
 

tests/integration/test_lists/qa/llm_perf_cluster.yml

@@ -13,7 +13,6 @@ llm_perf_cluster:
   - 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[t5-bench-float16-input_output_len:128,20]
   - 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]
@@ -46,7 +45,7 @@ llm_perf_cluster:
   - 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:500-con:200] 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:
@@ -61,42 +60,40 @@ llm_perf_cluster:
   - 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]
-  - perf/test_perf.py::test_perf[mixtral_8x7b_v0.1-bench-float16-input_output_len:128,128-quant:fp8-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:500-con:200-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[mixtral_8x22b_v0.1-bench-float16-input_output_len:512,512-quant:fp8-tp:4]
   - 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:20000-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:1000,1000-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)
+  - 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-reqs:1000-gpus:4] TIMEOUT(120)
+  - 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:3000-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:1000-gpus:4] TIMEOUT(120)
+  - 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:1500-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:1500-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:1000-gpus:4] TIMEOUT(120)
+  - 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
@@ -105,8 +102,6 @@ llm_perf_cluster:
       system_gpu_count:
         gte: 8
   tests:
-  #- perf/test_perf.py::test_perf[llama_v3.1_405b_instruct_fp4-bench-pytorch-float4-input_output_len:128,128-gpus:8]
-  #- perf/test_perf.py::test_perf[llama_v3.1_405b_instruct_fp4-bench-pytorch-float4-input_output_len:512,32-gpus:8]
   #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]
@@ -124,8 +119,6 @@ llm_perf_cluster:
   - 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]
-  #mixtral_8x22b_v0.1
-  - perf/test_perf.py::test_perf[mixtral_8x22b_v0.1-bench-float16-input_output_len:512,512-quant:fp8-tp: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

tests/integration/test_lists/qa/llm_perf_full.yml

@@ -52,8 +52,6 @@ llm_perf_full:
   #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:500-con:200] TIMEOUT(120)
 
   # Test list validation
   - test_list_validation.py::test_list_validation
@@ -76,6 +74,8 @@ llm_perf_full:
   - perf/test_perf.py::test_perf[deepseek_v3_lite_fp8-bench-pytorch-float8-input_output_len:2000,500]
   - perf/test_perf.py::test_perf[deepseek_v3_lite_fp8-bench-pytorch-streaming-float8-input_output_len:2000,500]
   - perf/test_perf.py::test_perf[deepseek_v3_lite_fp8-bench-pytorch-float8-input_output_len:500,2000]
+  - 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:500-con:200] TIMEOUT(120)
 
 # FP8 specific tests
 - condition:
@@ -179,8 +179,8 @@ llm_perf_full:
   - perf/test_perf.py::test_perf[mixtral_8x7b_v0.1_instruct_fp8-bench-pytorch-float8-input_output_len:512,32-gpus:2]
   #phi_3_mini_128k_instruct
   #pytorch backend
-  - perf/test_perf.py::test_perf[phi_3_mini_128k_instruct-bench-pytorch-float16-maxbs:128-input_output_len:1000,1000-quant:fp8-tp:2]
-  - perf/test_perf.py::test_perf[phi_3_mini_128k_instruct-bench-pytorch-float16-maxbs:128-input_output_len:500,2000-quant:fp8-tp:2]
+  - perf/test_perf.py::test_perf[phi_3_mini_128k_instruct-bench-pytorch-float16-maxbs:128-input_output_len:1000,1000-tp:2]
+  - perf/test_perf.py::test_perf[phi_3_mini_128k_instruct-bench-pytorch-float16-maxbs:128-input_output_len:500,2000-tp:2]
 
 - condition:
     terms:
@@ -262,8 +262,8 @@ llm_perf_full:
   #pytorch backend
   - perf/test_perf.py::test_perf[llama_v3.1_70b_instruct-bench-pytorch-bfloat16-input_output_len:2000,200-reqs:64-gpus:8]
   - perf/test_perf.py::test_perf[llama_v3.1_70b_instruct-bench-pytorch-streaming-bfloat16-input_output_len:2000,200-reqs:64-gpus:8]
-  - perf/test_perf.py::test_perf[llama_v3.3_70b-bench-pytorch-bfloat16-input_output_len:500,2000-gpus:8]
-  - perf/test_perf.py::test_perf[llama_v3.3_70b-bench-pytorch-bfloat16-input_output_len:2000,500-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct-bench-pytorch-bfloat16-input_output_len:500,2000-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct-bench-pytorch-bfloat16-input_output_len:2000,500-gpus:8]
   - perf/test_perf.py::test_perf[llama_v3.1_70b_instruct-bench-pytorch-bfloat16-input_output_len:200,2000-reqs:64-con:200-gpus:8]