refactor(profiling): remove cProfile, keep torch profiler only

Remove cProfile wrapping from the training loop and profiling utilities.
The torch profiler already captures fine-grained timing and operator
breakdowns; cProfile added redundant overhead without actionable
insight for GPU-bound models.

- Remove render_cprofile_summary, run_with_cprofile from profiling_utils
- Replace cProfile-wrapped calls in lerobot_train with direct calls
- Remove cprofile_summaries from artifact index in run_model_profiling
- Update tests to match

Made-with: Cursor

src/lerobot/scripts/lerobot_train.py

@@ -54,7 +54,6 @@ from lerobot.utils.profiling_utils import (
     StepTimingCollector,
     ensure_dir,
     make_torch_profiler,
-    run_with_cprofile,
     write_deterministic_forward_artifacts,
     write_torch_profiler_outputs,
 )
@@ -231,10 +230,8 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
 
     profiling_enabled = cfg.profile_mode != "off"
     profile_output_dir = None
-    cprofile_dir = None
     if profiling_enabled and is_main_process and cfg.profile_output_dir is not None:
         profile_output_dir = ensure_dir(Path(cfg.profile_output_dir))
-        cprofile_dir = ensure_dir(profile_output_dir / "cprofile")
         logging.info("Profiling enabled. Artifacts will be written to %s", profile_output_dir)
 
     # Initialize wandb only on main process
@@ -260,10 +257,7 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
     # Dataset loading synchronization: main process downloads first to avoid race conditions
     if is_main_process:
         logging.info("Creating dataset")
-        if cprofile_dir is not None:
-            dataset = run_with_cprofile("dataset_setup", cprofile_dir, make_dataset, cfg)
-        else:
-            dataset = make_dataset(cfg)
+        dataset = make_dataset(cfg)
 
     accelerator.wait_for_everyone()
 
@@ -281,21 +275,11 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
 
     if is_main_process:
         logging.info("Creating policy")
-    if is_main_process and cprofile_dir is not None:
-        policy = run_with_cprofile(
-            "policy_setup",
-            cprofile_dir,
-            make_policy,
-            cfg=cfg.policy,
-            ds_meta=dataset.meta,
-            rename_map=cfg.rename_map,
-        )
-    else:
-        policy = make_policy(
-            cfg=cfg.policy,
-            ds_meta=dataset.meta,
-            rename_map=cfg.rename_map,
-        )
+    policy = make_policy(
+        cfg=cfg.policy,
+        ds_meta=dataset.meta,
+        rename_map=cfg.rename_map,
+    )
 
     if cfg.peft is not None:
         logging.info("Using PEFT! Wrapping model.")
@@ -349,36 +333,16 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
             },
         }
 
-    if is_main_process and cprofile_dir is not None:
-        preprocessor, postprocessor = run_with_cprofile(
-            "processor_setup",
-            cprofile_dir,
-            make_pre_post_processors,
-            policy_cfg=cfg.policy,
-            pretrained_path=processor_pretrained_path,
-            **processor_kwargs,
-            **postprocessor_kwargs,
-        )
-    else:
-        preprocessor, postprocessor = make_pre_post_processors(
-            policy_cfg=cfg.policy,
-            pretrained_path=processor_pretrained_path,
-            **processor_kwargs,
-            **postprocessor_kwargs,
-        )
+    preprocessor, postprocessor = make_pre_post_processors(
+        policy_cfg=cfg.policy,
+        pretrained_path=processor_pretrained_path,
+        **processor_kwargs,
+        **postprocessor_kwargs,
+    )
 
     if is_main_process:
         logging.info("Creating optimizer and scheduler")
-    if is_main_process and cprofile_dir is not None:
-        optimizer, lr_scheduler = run_with_cprofile(
-            "optimizer_setup",
-            cprofile_dir,
-            make_optimizer_and_scheduler,
-            cfg,
-            policy,
-        )
-    else:
-        optimizer, lr_scheduler = make_optimizer_and_scheduler(cfg, policy)
+    optimizer, lr_scheduler = make_optimizer_and_scheduler(cfg, policy)
 
     if profiling_enabled and is_main_process and profile_output_dir is not None:
         logging.info("Recording deterministic forward-pass artifacts")

src/lerobot/utils/profiling_utils.py

@@ -16,13 +16,9 @@
 
 from __future__ import annotations
 
-import cProfile
 import hashlib
-import io
 import json
-import pstats
 import statistics
-from collections.abc import Callable
 from dataclasses import dataclass, field
 from numbers import Real
 from pathlib import Path
@@ -37,15 +33,6 @@ def ensure_dir(path: Path) -> Path:
     return path
 
 
-def render_cprofile_summary(
-    profile: cProfile.Profile, *, sort_by: str = "cumulative", limit: int = 40
-) -> str:
-    output = io.StringIO()
-    stats = pstats.Stats(profile, stream=output).strip_dirs().sort_stats(sort_by)
-    stats.print_stats(limit)
-    return output.getvalue()
-
-
 def write_profiler_table(
     profiler: Any,
     output_path: Path,
@@ -103,26 +90,6 @@ def write_torch_profiler_outputs(
     write_profiler_table(profiler, tables_dir / "flops.txt", sort_by="flops")
 
 
-def run_with_cprofile[T](
-    label: str,
-    output_dir: Path,
-    fn: Callable[..., T],
-    *args: Any,
-    sort_by: str = "cumulative",
-    limit: int = 40,
-    **kwargs: Any,
-) -> T:
-    ensure_dir(output_dir)
-    profile = cProfile.Profile()
-    profile.enable()
-    try:
-        return fn(*args, **kwargs)
-    finally:
-        profile.disable()
-        summary = render_cprofile_summary(profile, sort_by=sort_by, limit=limit)
-        (output_dir / f"{label}.txt").write_text(summary)
-
-
 def _stable_float(value: float | int | None) -> float | None:
     if value is None:
         return None