feat: TensorRT-LLM Benchmark Command Generator (M113)

ROADMAP.md

@@ -1505,9 +1505,9 @@ Help users find the **optimal Prefill:Decode instance ratio** based on **real be
 - Programmatic `generate_sglang_commands()` API
 - 25 new tests
 
-### M112 🔄 TensorRT-LLM Benchmark Format Importer
+### M112 ✅ TensorRT-LLM Benchmark Format Importer
 
-*In progress — PR #TBD*
+*Completed — PR #248*
 
 - `TRTLLMImporter` module in `trtllm_import.py`
 - `TRTLLMRequest`, `TRTLLMBenchmarkData`, `TRTLLMImportConfig`, `TRTLLMImportResult` Pydantic models
@@ -1517,3 +1517,16 @@ Help users find the **optimal Prefill:Decode instance ratio** based on **real be
 - CLI `import --format trtllm` support
 - Programmatic `import_trtllm()` and `import_trtllm_data()` API
 - 25+ new tests
+
+### M113 🔄 TensorRT-LLM Benchmark Command Generator
+
+*In progress*
+
+- `TRTLLMCommandGenerator` class in `trtllm_commands.py`
+- `TRTLLMCommandConfig`, `TRTLLMServerCommand`, `TRTLLMBenchmarkCommand`, `TRTLLMCommandSet` Pydantic models
+- Generate TRT-LLM engine build (`trtllm-build`) and server launch commands for each P:D ratio
+- TRT-LLM specific options: max_batch_size, kv_cache_free_gpu_mem_fraction, pp_size, dtype, engine_dir
+- Shell script output with engine build + server + benchmark lifecycle
+- CLI `trtllm-commands` subcommand with table + JSON output
+- Programmatic `generate_trtllm_commands()` API
+- 29 new tests

docs/iterations/current.md

@@ -53,7 +53,7 @@ The project has completed **110 milestones**, covering the full feature chain fr
 - Closed-loop integration with xPyD-proxy auto-tuning
 - Web UI dashboard (replacing TUI)
 - Richer visualizations (interactive charts)
-- Support additional benchmark tool formats (TensorRT-LLM)
+- Support additional benchmark tool formats (TensorRT-LLM) ✅
 
 ## Iteration History
 
@@ -64,4 +64,5 @@ The project has completed **110 milestones**, covering the full feature chain fr
 | 3 | 2026-04-06 | M109 vLLM Benchmark Command Generator | ✅ merged | PR #242 |
 | 4 | 2026-04-06 | M110 SGLang Benchmark Format Importer | ✅ merged | PR #244 |
 | 5 | 2026-04-06 | M111 SGLang Benchmark Command Generator | ✅ merged | PR #246 |
-| 6 | 2026-04-06 | M112 TensorRT-LLM Benchmark Format Importer | ⏳ pending review | Issue #247 |
+| 6 | 2026-04-06 | M112 TensorRT-LLM Benchmark Format Importer | ✅ merged | PR #248, both bots approved |
+| 7 | 2026-04-06 | M113 TensorRT-LLM Benchmark Command Generator | ⏳ pending review | Issue #249 |

src/xpyd_plan/__init__.py

@@ -1412,6 +1412,10 @@
 from xpyd_plan.sglang_commands import SGLangCommandGenerator  # noqa: E402
 from xpyd_plan.sglang_commands import SGLangCommandSet  # noqa: E402
 from xpyd_plan.sglang_commands import generate_sglang_commands  # noqa: E402
+from xpyd_plan.trtllm_commands import TRTLLMCommandConfig  # noqa: E402, I001
+from xpyd_plan.trtllm_commands import TRTLLMCommandGenerator  # noqa: E402
+from xpyd_plan.trtllm_commands import TRTLLMCommandSet  # noqa: E402
+from xpyd_plan.trtllm_commands import generate_trtllm_commands  # noqa: E402
 from xpyd_plan.vllm_commands import BenchmarkCommand  # noqa: E402
 from xpyd_plan.vllm_commands import CommandGenerator  # noqa: E402
 from xpyd_plan.vllm_commands import CommandSet  # noqa: E402
@@ -1430,6 +1434,10 @@
     "SGLangCommandConfig",
     "SGLangCommandGenerator",
     "SGLangCommandSet",
+    "generate_trtllm_commands",
+    "TRTLLMCommandConfig",
+    "TRTLLMCommandGenerator",
+    "TRTLLMCommandSet",
 ]
 
 from xpyd_plan.vllm_import import (  # noqa: E402

src/xpyd_plan/cli/_main.py

@@ -96,6 +96,7 @@
 from xpyd_plan.cli._token_budget import add_token_budget_parser
 from xpyd_plan.cli._token_efficiency import add_token_efficiency_parser, handle_token_efficiency
 from xpyd_plan.cli._trend import _cmd_trend
+from xpyd_plan.cli._trtllm_commands import register_trtllm_commands
 from xpyd_plan.cli._validate import _cmd_validate
 from xpyd_plan.cli._variance import _cmd_variance, add_variance_parser
 from xpyd_plan.cli._vllm_commands import register_vllm_commands
@@ -963,6 +964,7 @@ def main(argv: list[str] | None = None) -> None:
     add_import_parser(subparsers)
     register_vllm_commands(subparsers)
     register_sglang_commands(subparsers)
+    register_trtllm_commands(subparsers)
     add_rate_limit_parser(subparsers)
     add_batch_analysis_parser(subparsers)
     add_stat_summary_parser(subparsers)
@@ -1308,6 +1310,10 @@ def main(argv: list[str] | None = None) -> None:
         from xpyd_plan.cli._sglang_commands import _cmd_sglang_commands
 
         _cmd_sglang_commands(args)
+    elif args.command == "trtllm-commands":
+        from xpyd_plan.cli._trtllm_commands import _cmd_trtllm_commands
+
+        _cmd_trtllm_commands(args)
     else:
         parser.print_help()
         sys.exit(1)

src/xpyd_plan/cli/_trtllm_commands.py

@@ -0,0 +1,162 @@
+"""CLI trtllm-commands subcommand."""
+
+from __future__ import annotations
+
+import argparse
+import json
+
+from rich.console import Console
+from rich.table import Table
+
+from xpyd_plan.trtllm_commands import TRTLLMCommandConfig, TRTLLMCommandGenerator
+
+
+def _cmd_trtllm_commands(args: argparse.Namespace) -> None:
+    """Handle the 'trtllm-commands' subcommand."""
+    console = Console()
+
+    qps_levels = [float(q) for q in args.qps.split(",")]
+
+    config = TRTLLMCommandConfig(
+        model=args.model,
+        total_instances=args.total_instances,
+        qps_levels=qps_levels,
+        tp_size=getattr(args, "tp_size", 1),
+        pp_size=getattr(args, "pp_size", 1),
+        max_batch_size=getattr(args, "max_batch_size", 256),
+        max_input_len=getattr(args, "max_input_len", 2048),
+        max_output_len=getattr(args, "max_output_len", 2048),
+        kv_cache_free_gpu_mem_fraction=getattr(
+            args, "kv_cache_free_gpu_mem_fraction", 0.9
+        ),
+        dtype=getattr(args, "dtype", "float16"),
+        dataset=getattr(args, "dataset", None),
+        num_prompts=getattr(args, "num_prompts", 1000),
+        host=getattr(args, "host", "localhost"),
+        port=getattr(args, "port", 8000),
+        engine_dir=getattr(args, "engine_dir", "./engines"),
+    )
+
+    gen = TRTLLMCommandGenerator(config)
+    result = gen.generate()
+
+    if args.output_script:
+        script = _to_shell_script(result, config)
+        with open(args.output_script, "w") as f:
+            f.write(script)
+        console.print(
+            f"[green]Shell script written to {args.output_script} "
+            f"({len(result)} ratios)[/green]"
+        )
+        return
+
+    output_format = getattr(args, "output_format", "table")
+
+    if output_format == "json":
+        print(json.dumps([cs.model_dump() for cs in result], indent=2, default=str))
+        return
+
+    # Table output
+    console.print("\n[bold]TensorRT-LLM Benchmark Commands[/bold]")
+    console.print(
+        f"Model: {config.model} | Instances: {config.total_instances} | "
+        f"Ratios: {len(result)}\n"
+    )
+
+    table = Table(title="Benchmark Runs")
+    table.add_column("P:D Ratio")
+    table.add_column("Prefill", justify="right")
+    table.add_column("Decode", justify="right")
+    table.add_column("QPS Levels")
+
+    for cs in result:
+        table.add_row(
+            cs.server.ratio,
+            str(cs.server.prefill_instances),
+            str(cs.server.decode_instances),
+            ", ".join(f"{b.qps}" for b in cs.benchmarks),
+        )
+
+    console.print(table)
+    console.print(
+        "\n[dim]Use --output-script to generate an executable shell script[/dim]"
+    )
+
+
+def _to_shell_script(
+    command_sets: list,
+    config: TRTLLMCommandConfig,
+) -> str:
+    """Build a complete shell script from command sets."""
+    lines = [
+        "#!/usr/bin/env bash",
+        "set -euo pipefail",
+        f"# TensorRT-LLM Benchmark Script — {config.model}",
+        f"# Total instances: {config.total_instances}",
+        f"# QPS levels: {', '.join(str(q) for q in config.qps_levels)}",
+        "",
+    ]
+    for cs in command_sets:
+        lines.append(cs.script_snippet)
+    lines.append("echo 'All benchmarks complete!'")
+    return "\n".join(lines) + "\n"
+
+
+def register_trtllm_commands(subparsers: argparse._SubParsersAction) -> None:
+    """Register the trtllm-commands subcommand."""
+    p = subparsers.add_parser(
+        "trtllm-commands",
+        help="Generate TensorRT-LLM benchmark commands for P:D ratio exploration",
+    )
+    p.add_argument("--model", type=str, required=True, help="HuggingFace model name")
+    p.add_argument(
+        "--total-instances",
+        type=int,
+        required=True,
+        help="Total instances (prefill + decode)",
+    )
+    p.add_argument(
+        "--qps",
+        type=str,
+        required=True,
+        help="Comma-separated QPS levels (e.g. 1,2,4)",
+    )
+    p.add_argument("--tp-size", type=int, default=1, help="Tensor parallel size")
+    p.add_argument("--pp-size", type=int, default=1, help="Pipeline parallel size")
+    p.add_argument("--max-batch-size", type=int, default=256, help="Max batch size")
+    p.add_argument(
+        "--max-input-len", type=int, default=2048, help="Max input length"
+    )
+    p.add_argument(
+        "--max-output-len", type=int, default=2048, help="Max output length"
+    )
+    p.add_argument(
+        "--kv-cache-free-gpu-mem-fraction",
+        type=float,
+        default=0.9,
+        help="KV cache GPU memory fraction",
+    )
+    p.add_argument(
+        "--dtype",
+        type=str,
+        default="float16",
+        choices=["float16", "bfloat16", "float32"],
+        help="Data type",
+    )
+    p.add_argument("--dataset", type=str, default=None, help="Dataset path")
+    p.add_argument("--num-prompts", type=int, default=1000, help="Prompts per run")
+    p.add_argument("--host", type=str, default="localhost", help="Server host")
+    p.add_argument("--port", type=int, default=8000, help="Server port")
+    p.add_argument(
+        "--engine-dir", type=str, default="./engines", help="Engine output directory"
+    )
+    p.add_argument(
+        "--output-script", type=str, default=None, help="Write shell script"
+    )
+    p.add_argument(
+        "--output-format",
+        choices=["table", "json"],
+        default="table",
+        help="Output format",
+    )
+    p.set_defaults(func=_cmd_trtllm_commands)