forgeenv source snapshot for training job

scripts/jobs/train_repair_agent.py

@@ -71,11 +71,13 @@ sys.path.insert(0, str(src_dir))
 step("1. pin torch (cu124) + install GPU-stable deps")
 # Force a CUDA 12.4 torch wheel BEFORE anything else so other packages'
 # resolvers don't pull a cu130 wheel that mismatches the host driver
-# (this is what causes "Error 802: system not yet initialized" on H200).
+# (Error 802 on some HF Job flavors).  TRL 1.2+ imports ``FSDPModule`` from
+# ``torch.distributed.fsdp``, which exists only in PyTorch >= 2.6 — do not
+# pin to 2.5.x.
 _sh([
     sys.executable, "-m", "pip", "install",
     "--index-url", "https://download.pytorch.org/whl/cu124",
-    "torch==2.5.1", "torchvision==0.20.1",
+    "torch==2.6.0", "torchvision==0.21.0",
 ])
 # `--no-deps` on openenv-core: it pins a different transformers/torch
 # stack that we don't want. We still need its *runtime* imports: