fix: remove wandb - click conflict with gradio causes resolution-too-deep

requirements.txt

@@ -1,9 +1,11 @@
 # ── Training dependencies ──────────────────────────────────────────────────────
 # Fully pinned to a pre-validated compatible set.
-# DO NOT add loose deps here — they cause pip backtracking (resolution-too-deep).
-# gradio is intentionally omitted: HF Spaces injects gradio[oauth,mcp]==6.13.0 itself.
+# NOTES:
+#   - gradio is injected by HF Spaces automatically — do NOT add it here
+#   - wandb is excluded: it conflicts with gradio over click versioning
+#     (wandb>=0.18 requires click!=8.0.0,>=7.1 but gradio requires click>=8.1)
+#     wandb is initialized at runtime by the training script if available
 
-wandb==0.18.7
 datasets==3.0.2
 transformers==4.46.3
 accelerate==1.0.1

training/train_grpo.py

@@ -49,8 +49,15 @@ if os.environ.get("FORCE_BOOTSTRAP_DEPS") == "1":
 
 # ── GPU/training imports (skipped in --test-local mode) ───────────────────────
 if not args.test_local:
+    # wandb is not in requirements.txt (conflicts with gradio over click versioning)
+    # Install it at runtime before importing
+    try:
+        import wandb
+    except ImportError:
+        os.system(f"{sys.executable} -m pip install -q 'wandb>=0.18.0'")
+        import wandb
+    
     import torch
-    import wandb
     from datasets import Dataset
     from transformers import (
         AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TrainerCallback