[TRTLLM-8991][test] Add Llama 3.3 70B model with different performance config (#8753)

Signed-off-by: yufeiwu-nv <230315618+yufeiwu-nv@users.noreply.github.com>
Co-authored-by: Larry Xu <197874197+LarryXFly@users.noreply.github.com>

tensorrt_llm/bench/benchmark/__init__.py

@@ -105,7 +105,7 @@ def get_llm(runtime_config: RuntimeConfig, kwargs: dict):
     """
     llm_cls = LLM
 
-    if runtime_config.backend != "tensorrt":
+    if runtime_config.backend != None:
         ignore_trt_only_args(kwargs, runtime_config.backend)
 
     if runtime_config.backend == 'pytorch':

tests/integration/defs/perf/pytorch_model_config.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 # -*- coding: utf-8 -*-
 """
-Model pytorch yaml config for trtllm-bench perf tests
+Model pytorch/TRT yaml config for trtllm-bench perf tests
 """
 
 
@@ -36,12 +36,18 @@ def get_model_yaml_config(model_label: str,
         Returns:
             dict: yaml config
         """
-    base_config = {
-        'print_iter_log': True,
-        'cuda_graph_config': {
-            'enable_padding': True,
-        },
-    }
+    if 'pytorch' in model_label:
+        # Pytorch backend config
+        base_config = {
+            'print_iter_log': True,
+            'cuda_graph_config': {
+                'enable_padding': True,
+            },
+        }
+    else:
+        # TRT backend config
+        base_config = {}
+
     if 'kv_cache_dtype' in model_label:
         base_config.update({
             'kv_cache_dtype':
@@ -241,6 +247,19 @@ def get_model_yaml_config(model_label: str,
             'config': {
                 'enable_chunked_prefill': True,
             }
+        },
+        # Llama-v3.3 models with xgrammar guided decoding
+        {
+            'patterns': [
+                "llama_v3.3_70b_instruct_fp8-bench-float8-maxbs:512-maxnt:2048-input_output_len:500,2000-reqs:400-con:200-gpus:8-extra"
+            ],
+            'config': {
+                'extended_runtime_perf_knob_config': {
+                    'cuda_graph_cache_size': 1.0,
+                    'cuda_graph_mode': True,
+                },
+                'guided_decoding_backend': 'xgrammar'
+            }
         }
     ]
 
@@ -251,7 +270,8 @@ def get_model_yaml_config(model_label: str,
             patterns = [patterns]
         for pattern in patterns:
             if pattern in model_label.lower():
-                recursive_update(base_config, pattern_config['config'])
+                if pattern_config.get('config'):
+                    recursive_update(base_config, pattern_config['config'])
                 break  # Stop checking other patterns for this config once we find a match
 
     # lora-specific change for pytorch

tests/integration/defs/perf/test_perf.py

@@ -57,7 +57,6 @@ MODEL_PATH_DICT = {
     "modelopt-hf-model-hub/Llama-3.3-70B-Instruct-fp8",
     "llama_v3.3_70b_instruct_fp4":
     "modelopt-hf-model-hub/Llama-3.3-70B-Instruct-fp4",
-    "llama_v3.3_70b_instruct": "llama-3.3-models/Llama-3.3-70B-Instruct",
     "llama_v3.1_405b_instruct_fp8":
     "llama-3.1-model/Llama-3.1-405B-Instruct-FP8",
     "llama_v3.1_405b_instruct_fp4":
@@ -783,6 +782,8 @@ class PerfTestConfig:
         tp_size: int = 1,
         pp_size: int = 1,
         num_gpus: int = 1,
+        # only for torch-backend currently
+        extra: bool = False,
         # _autodeploy backend specific parameters
         ad_compile_backend: str = "torch-opt",
         free_mem_ratio: float = 0.9,
@@ -841,6 +842,8 @@ class PerfTestConfig:
         self.pp_size = pp_size
         # Number of GPUs.
         self.num_gpus = num_gpus
+        # Extra flag to enable pytorch_model_config reading for TRT backend
+        self.extra = extra
         # _autodeploy backend specific parameters
         self.ad_compile_backend = ad_compile_backend
         self.free_mem_ratio = free_mem_ratio
@@ -1016,6 +1019,10 @@ class PerfTestConfig:
         if self.num_gpus > 1:
             entries.append(f"gpus:{self.num_gpus}")
 
+        # Add extra flag for llm-api-config.yml.
+        if self.extra:
+            entries.append("extra")
+
         # Concatenate labels with "-".
         return "-".join(entries)
 
@@ -1180,6 +1187,11 @@ class PerfTestConfig:
             self.num_gpus = 1 if not labels[0].startswith("gpus:") else int(
                 labels.pop(0).replace("gpus:", ""))
 
+        if len(labels) > 0:
+            self.extra = True if labels[0] == "extra" else False
+            if self.extra:
+                labels.pop(0)
+
         assert len(
             labels
         ) == 0, f"Invalid test name! Some labels cannot be parsed: {labels}"
@@ -1644,18 +1656,26 @@ class MultiMetricPerfTest(AbstractPerfScriptTestClass):
             benchmark_cmd += [f"--pp={self._config.pp_size}"]
         if self._config.streaming == "streaming":
             benchmark_cmd += [f"--streaming"]
-        #use default yaml config
-        if self._config.backend == "pytorch":
+
+        #Add extra-llm-api-config.yml for pytorch backend and tensorrt backend with extra flag
+        if self._config.backend == "pytorch" or (self._config.backend == ""
+                                                 and self._config.extra):
             pytorch_config_path = os.path.join(engine_dir,
                                                "extra-llm-api-config.yml")
             if not os.path.exists(pytorch_config_path):
                 os.makedirs(os.path.dirname(pytorch_config_path), exist_ok=True)
             config = get_model_yaml_config(self._config.to_string(),
                                            lora_dirs=self.lora_dirs)
-            print_info(f"pytorch model config: {config}")
-            with open(pytorch_config_path, 'w') as f:
-                yaml.dump(config, f, default_flow_style=False)
-            benchmark_cmd += [f"--extra_llm_api_options={pytorch_config_path}"]
+            if config:
+                print_info(f"pytorch/TRT model config: {config}")
+                with open(pytorch_config_path, 'w') as f:
+                    yaml.dump(config, f, default_flow_style=False)
+                benchmark_cmd += [
+                    f"--extra_llm_api_options={pytorch_config_path}"
+                ]
+                # If guided_decoding_backend is set, we need to initialize tokenizer
+                if config.get('guided_decoding_backend') is not None:
+                    benchmark_cmd += ["--no_skip_tokenizer_init"]
         elif self._config.backend == "_autodeploy":
             autodeploy_config_path = os.path.join(engine_dir,
                                                   "extra_llm_api_options.yaml")

tests/integration/test_lists/qa/llm_perf_nim.yml

@@ -392,6 +392,8 @@ llm_perf_nim:
   #trt backend
   - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp8-bench-float8-input_output_len:128,128-gpus:8]
   - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp8-bench-float8-input_output_len:512,32-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp8-bench-float8-maxbs:512-maxnt:2048-input_output_len:500,2000-reqs:400-con:200-gpus:8]
+  - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp8-bench-float8-maxbs:512-maxnt:2048-input_output_len:500,2000-reqs:400-con:200-gpus:8-extra]
   - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp8-bench-streaming-float8-maxbs:16-input_output_len:512,32-gpus:8]
   - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct-bench-bfloat16-maxbs:16-maxnt:5000-input_output_len:5000,500-quant:fp8-reqs:8-con:1-gpus:8]
   - perf/test_perf.py::test_perf[llama_v3.3_70b_instruct-bench-bfloat16-maxbs:16-maxnt:5000-input_output_len:5000,500-quant:fp8-reqs:64-con:250-gpus:8]