Update CERNenv Space

space/training/app.py

@@ -269,7 +269,9 @@ def _training_pipeline(config: Dict[str, Any]) -> None:
                 log,
             )
             if rc != 0:
-                raise RuntimeError(f"pre-train eval failed (rc={rc})")
+                # don't abort — we still want training + post-eval evidence.
+                log.write(f"\n[warn] pre-train eval failed (rc={rc}); continuing without baseline\n")
+                log.flush()
 
             log.write(f"\n--- GRPO training ({config['num_gpus']} GPU process(es)) ---\n")
             log.flush()
@@ -293,7 +295,8 @@ def _training_pipeline(config: Dict[str, Any]) -> None:
                 log,
             )
             if rc != 0:
-                raise RuntimeError(f"post-train eval failed (rc={rc})")
+                log.write(f"\n[warn] post-train eval failed (rc={rc}); evidence will be partial\n")
+                log.flush()
 
             log.write("\n--- evidence: before/after summary, distribution, trajectories ---\n")
             log.flush()

space/training/requirements.txt

@@ -1,14 +1,20 @@
 --extra-index-url https://download.pytorch.org/whl/cu121
-torch==2.4.0
-unsloth
-unsloth_zoo
-transformers>=4.44.0
-trl>=0.9.0
-peft>=0.10.0
-accelerate>=1.0.0
-vllm>=0.5.0
-datasets>=2.18.0
-bitsandbytes>=0.43.0
+# Pin a torch + torchao + transformers triple known to work with Unsloth on
+# CUDA 12.1. We avoid newer transformers/torchao because they require
+# torch>=2.6 (torch.int1 added in 2.6); we pin torch 2.5.1 which is the
+# latest stable cu121 wheel that pairs cleanly with Unsloth's kernels.
+torch==2.5.1
+torchvision==0.20.1
+torchaudio==2.5.1
+torchao==0.7.0
+transformers==4.46.3
+trl==0.12.2
+peft==0.13.2
+accelerate==1.1.1
+datasets==3.1.0
+bitsandbytes==0.44.1
+unsloth==2024.12.4
+unsloth_zoo==2024.12.7
 matplotlib>=3.8.0
 numpy>=1.24.0
 scipy>=1.10.0