diff --git a/examples/puzzletron/configs/llama-3_1-8B_pruneffn_runtime/llama-3_1-8B_pruneffn_runtime.yaml b/examples/puzzletron/configs/llama-3_1-8B_pruneffn_runtime/llama-3_1-8B_pruneffn_runtime.yaml
index 2ca0f2c16cf..e788443defd 100644
--- a/examples/puzzletron/configs/llama-3_1-8B_pruneffn_runtime/llama-3_1-8B_pruneffn_runtime.yaml
+++ b/examples/puzzletron/configs/llama-3_1-8B_pruneffn_runtime/llama-3_1-8B_pruneffn_runtime.yaml
@@ -16,6 +16,12 @@ calc_subblock_stats:
     enabled: true
     num_warmup_iters: 2
     num_iters: 10
+    data_parallel_size: 1
+    tensor_parallel_size: 1
+    prefill_context_parallel_size: 1
+    decode_context_parallel_size: 1
+    enable_expert_parallel: false
+    extra_args: ""
 
 # FFN intermediate sizes to search over (heterogeneous architecture)
 pruning:
diff --git a/modelopt/torch/puzzletron/subblock_stats/calc_runtime_stats.py b/modelopt/torch/puzzletron/subblock_stats/calc_runtime_stats.py
index c7d926de18f..669b25e675d 100644
--- a/modelopt/torch/puzzletron/subblock_stats/calc_runtime_stats.py
+++ b/modelopt/torch/puzzletron/subblock_stats/calc_runtime_stats.py
@@ -186,9 +186,15 @@ def calc_runtime_for_subblocks(
         prefill_seq_len,
         generation_seq_len,
         batch_size,
+        runtime_stats_config.get("data_parallel_size", 1),
+        runtime_stats_config.get("tensor_parallel_size", 1),
+        runtime_stats_config.get("prefill_context_parallel_size", 1),
+        runtime_stats_config.get("decode_context_parallel_size", 1),
+        runtime_stats_config.get("enable_expert_parallel", False),
         runtime_stats_config.get("num_iters", 30),
         runtime_stats_config.get("num_warmup_iters", 10),
         runtime_stats_config.get("gpu_memory_utilization", 0.5),
+        runtime_stats_config.get("extra_args", ""),
     )
 
     runtime_by_subblock_dict = {}
diff --git a/modelopt/torch/puzzletron/subblock_stats/runtime_utils.py b/modelopt/torch/puzzletron/subblock_stats/runtime_utils.py
index 204c2a74305..c25083692b0 100644
--- a/modelopt/torch/puzzletron/subblock_stats/runtime_utils.py
+++ b/modelopt/torch/puzzletron/subblock_stats/runtime_utils.py
@@ -50,6 +50,11 @@ class RuntimeConfig:
     prefill_seq_len: int
     generation_seq_len: int
     batch_size: int
+    data_parallel_size: int
+    tensor_parallel_size: int
+    prefill_context_parallel_size: int
+    decode_context_parallel_size: int
+    enable_expert_parallel: bool
     num_iters: int
     num_warmup_iters: int
     # Fraction of total GPU memory vLLM may use. Kept well below the default
@@ -57,6 +62,7 @@ class RuntimeConfig:
     # GPU during benchmarking; requesting too much fails vLLM's startup
     # free-memory check.
     gpu_memory_utilization: float = 0.5
+    extra_args: str = ""
 
 
 def save_model(model: LlamaForCausalLM, tokenizer_path: Path, output_path: Path) -> None:
diff --git a/modelopt/torch/puzzletron/subblock_stats/runtime_vllm.py b/modelopt/torch/puzzletron/subblock_stats/runtime_vllm.py
index 386f3615d49..bd2493d6115 100644
--- a/modelopt/torch/puzzletron/subblock_stats/runtime_vllm.py
+++ b/modelopt/torch/puzzletron/subblock_stats/runtime_vllm.py
@@ -27,6 +27,7 @@
 """
 
 import json
+import os
 import subprocess  # nosec B404
 from pathlib import Path
 from types import SimpleNamespace
@@ -79,12 +80,19 @@ def run_vllm_latency_benchmark(model_path: Path, runtime_config: RuntimeConfig)
         str(runtime_config.num_iters),
         "--max-num-seqs",
         "1",
+        "--data-parallel-size",
+        str(runtime_config.data_parallel_size),
+        "--prefill-context-parallel-size",
+        str(runtime_config.prefill_context_parallel_size),
+        "--decode-context-parallel-size",
+        str(runtime_config.decode_context_parallel_size),
+        "--enable-expert-parallel"
+        if runtime_config.enable_expert_parallel
+        else "--no-enable-expert-parallel",
         "--tensor-parallel-size",
-        "1",
-        "--pipeline-parallel-size",
-        "1",
+        str(runtime_config.tensor_parallel_size),
         "--distributed-executor-backend",
-        "external_launcher",
+        "mp",  # "external_launcher",
         # Cap GPU memory so vLLM's startup free-memory check passes while the
         # parent puzzletron process is co-resident on the same GPU.
         "--gpu-memory-utilization",
@@ -93,6 +101,26 @@ def run_vllm_latency_benchmark(model_path: Path, runtime_config: RuntimeConfig)
         "--optimization-level",
         "0",
     ]
+    if runtime_config.extra_args:
+        cmd.extend(runtime_config.extra_args.split(" "))
+
+    # Strip torchrun-injected distributed vars so vLLM's mp workers don't
+    # inherit an outer process group's rendezvous state, which causes a
+    # deadlock when vLLM tries to create its own Gloo CPU group.
+    _TORCHRUN_VARS = {
+        "RANK",
+        "LOCAL_RANK",
+        "WORLD_SIZE",
+        "LOCAL_WORLD_SIZE",
+        "MASTER_ADDR",
+        "MASTER_PORT",
+        "TORCHELASTIC_RESTART_COUNT",
+        "TORCHELASTIC_MAX_RESTARTS",
+        "TORCHELASTIC_RUN_ID",
+        "TORCHELASTIC_USE_AGENT_STORE",
+        "TORCHELASTIC_ERROR_FILE",
+    }
+    env = {k: v for k, v in os.environ.items() if k not in _TORCHRUN_VARS}
 
     # cmd is a fixed list of strings (no shell, no untrusted input).
     try:
@@ -101,7 +129,8 @@ def run_vllm_latency_benchmark(model_path: Path, runtime_config: RuntimeConfig)
             check=True,
             capture_output=True,
             text=True,
-            timeout=1800,  # 30 minutes
+            timeout=3600,  # 1 hour
+            env=env,
         )  # nosec B603
     except subprocess.TimeoutExpired as exc:
         raise TimeoutError("vLLM latency benchmark timed out") from exc