fix: disable AMP for quantized models to avoid gradient scaler issues in GRPO training

scripts/gridmind_grpo_colab.ipynb

@@ -391,6 +391,7 @@
     ")\n",
     "\n",
     "# Configure GRPO training\n",
+    "# Note: Disable AMP (fp16/bf16) when using quantized models to avoid gradient scaler issues\n",
     "grpo_config_dict = {\n",
     "    \"output_dir\": \"./gridmind-grpo-output\",\n",
     "    \"num_train_epochs\": 1,\n",
@@ -401,7 +402,7 @@
     "    \"max_completion_length\": 80,\n",
     "    \"num_generations\": 4,\n",
     "    \"learning_rate\": 5e-5,\n",
-    "    \"fp16\": True,\n",
+    "    \"fp16\": False,\n",
     "    \"bf16\": False,\n",
     "    \"max_grad_norm\": 0.0,\n",
     "    \"logging_steps\": 1,\n",
@@ -424,7 +425,7 @@
     "print(f\"  Batch size: {getattr(grpo_config, 'per_device_train_batch_size', 1)}\")\n",
     "print(f\"  Generations: {getattr(grpo_config, 'num_generations', 4)}\")\n",
     "print(f\"  Learning rate: {getattr(grpo_config, 'learning_rate', 5e-5)}\")\n",
-    "print(f\"  Precision: FP16\")\n",
+    "print(f\"  Precision: Native (FP32, quantized to INT4)\")\n",
     "\n",
     "trainer = GRPOTrainer(\n",
     "    model=model,\n",

scripts/train_unsloth.py

@@ -687,8 +687,8 @@ def main():
         "log_completions": True,  # Enable completion metrics in table
         "num_completions_to_print": 1,  # Print 1 completion per step
         "save_steps": 100,
-        "fp16": not use_bf16,
-        "bf16": use_bf16,
+        "fp16": False,  # Disable AMP with quantized models (avoid grad scaler issues)
+        "bf16": False,
         "max_grad_norm": 0.0,
         "report_to": "none",
         "seed": 42,