Fix benchmark: use symlinked data root for subsampling, disable file size filter

Subsampled filelist needs its own data root (with symlinks to volume
data/label dirs) since `run_training` reads `mp_filelist.txt` from
the data root. Default `max_file_size=0` (no filter) since dataset_4
files fit in L4 VRAM.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: ryan-williams

modal/benchmark.py

@@ -62,7 +62,7 @@ def run_benchmark(
     channels: int = 32,
     residual_blocks: int = 16,
     samples: int = 50,
-    max_file_size: float = 25,
+    max_file_size: float = 0,
     seed: int = 42,
     wandb_project: str = "elf-net-ci",
     gpu_type: str = "L4",
@@ -80,14 +80,20 @@ def run_benchmark(
     # Subsample from volume filelist if requested
     filelist = Path(DATA_ROOT) / "mp_filelist.txt"
     all_ids = filelist.read_text().strip().splitlines()
-    if samples > 0 and samples < len(all_ids):
+    if 0 < samples < len(all_ids):
         import random
 
         random.seed(seed)
         subset = sorted(random.sample(all_ids, samples))
-        sub_filelist = Path(DATA_ROOT) / "mp_filelist_benchmark.txt"
-        sub_filelist.write_text("\n".join(subset) + "\n")
-        data_root = DATA_ROOT
+        # Create a temp data root with symlinks to volume data/label dirs
+        # and the subsampled filelist
+        data_root = "/tmp/benchmark_data"
+        Path(data_root).mkdir(parents=True, exist_ok=True)
+        for subdir in ["data", "label"]:
+            link = Path(data_root) / subdir
+            if not link.exists():
+                link.symlink_to(Path(DATA_ROOT) / subdir)
+        Path(data_root, "mp_filelist.txt").write_text("\n".join(subset) + "\n")
         log.info("Subsampled %d/%d samples", samples, len(all_ids))
     else:
         data_root = DATA_ROOT
@@ -130,7 +136,7 @@ def main(
     channels: int = 32,
     residual_blocks: int = 16,
     samples: int = 50,
-    max_file_size: float = 25,
+    max_file_size: float = 0,
     seed: int = 42,
     wandb_project: str = "elf-net-ci",
 ):