[None][test] Add Muse Spark Open NVFP4 core performance coverage

tests/integration/defs/perf/_model_paths.py

@@ -16,6 +16,7 @@
 
 # Model PATH of local dir synced from internal LLM models repo
 MODEL_PATH_DICT = {
+    "muse_spark_open_nvfp4": "Muse-Spark-Open-816B-A42B-NVFP4",
     "llama_v3.1_8b_instruct": "llama-3.1-model/Llama-3.1-8B-Instruct",
     "llama_v3.1_8b_instruct_fp8": "llama-3.1-model/Llama-3.1-8B-Instruct-FP8",
     "llama_v3.1_8b_instruct_fp4": "modelopt-hf-model-hub/Llama-3.1-8B-Instruct-fp4",

tests/integration/defs/perf/pytorch_model_config.py

@@ -107,6 +107,32 @@ def get_model_yaml_config(model_label: str,
     # Pattern-based configurations for models matching specific substrings
     # This allows for flexible configuration of models based on naming patterns
     pattern_configs = [
+        {
+            'patterns': ['muse_spark_open_nvfp4-serve-pytorch'],
+            'config': {
+                'disable_mm_encoder': True,
+                'enable_attention_dp': False,
+                'enable_chunked_prefill': True,
+                'disable_overlap_scheduler': False,
+                'attn_backend': 'TRTLLM',
+                'allreduce_strategy': 'NCCL',
+                'max_seq_len': 49152,
+                'cuda_graph_config': {
+                    'enable_padding': True,
+                    'max_batch_size': 256,
+                },
+                'kv_cache_config': {
+                    'dtype': 'fp8',
+                    'enable_block_reuse': True,
+                    'enable_partial_reuse': True,
+                    'free_gpu_memory_fraction': 0.8,
+                    'use_kv_cache_manager_v2': True,
+                },
+                'moe_config': {
+                    'backend': 'TRTLLM',
+                },
+            }
+        },
         {
             'patterns': ['deepseek_v41_flash_nvfp4'],
             'config': {

tests/integration/defs/perf/test_perf.py

@@ -1632,6 +1632,9 @@ def get_commands(self):
             elif self._config.model_name == "glm_5.3_fp8":
                 # Allow time for checkpoint loading and cold executor warmup.
                 server_timeout = 1800
+            elif self._config.model_name == "muse_spark_open_nvfp4":
+                # Allow time for NVFP4 checkpoint loading and executor warmup.
+                server_timeout = 3600
             elif self._config.model_name in {
                     "glm_5.3_flash_fp8", "glm_5.3_flash_fp8_mtp"
             }:

tests/integration/test_lists/qa/llm_perf_core.yml

@@ -154,6 +154,10 @@ llm_perf_core:
   tests:
   # Upper bound is 10.7 so Rubin shares this set: a VR200 NVL72 node exposes 4 GPUs,
   # the same shape these cases are sized for, so it runs them rather than a copy.
+  # Muse Spark Open NVFP4, TP4/EP4, FP8 KV cache; text-only serving.
+  - perf/test_perf.py::test_perf[muse_spark_open_nvfp4-serve-pytorch-streaming-float4-maxbs:256-maxnt:8192-kv_frac:0.8-input_output_len:8192,1024-reqs:5-con:1-ep:4-tp:4-gpus:4] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[muse_spark_open_nvfp4-serve-pytorch-streaming-float4-maxbs:256-maxnt:8192-kv_frac:0.8-input_output_len:8192,1024-reqs:160-con:32-ep:4-tp:4-gpus:4] TIMEOUT(120)
+  - perf/test_perf.py::test_perf[muse_spark_open_nvfp4-serve-pytorch-streaming-float4-maxbs:256-maxnt:8192-kv_frac:0.8-input_output_len:8192,1024-reqs:1280-con:256-ep:4-tp:4-gpus:4] TIMEOUT(120)
   # GLM-5.3-Flash FP8, TP4/EP4, BF16 KV cache, TRTLLM MoE, MTP3; honor EOS.
   - perf/test_perf.py::test_perf[glm_5.3_flash_fp8_mtp-serve-pytorch-streaming-float8-maxbs:64-maxnt:16384-kv_frac:0.5-input_output_len:1024,1024-reqs:5-con:1-ep:4-tp:4-gpus:4] TIMEOUT(120)
   - perf/test_perf.py::test_perf[glm_5.3_flash_fp8_mtp-serve-pytorch-streaming-float8-maxbs:64-maxnt:16384-kv_frac:0.5-input_output_len:1024,1024-reqs:320-con:64-ep:4-tp:4-gpus:4] TIMEOUT(120)