fix(training): force FP32 precision for PII LoRA training to fix uniform 0.9 confidence

Issue vllm-project#647 reported uniform 0.9 confidence scores in PII detection.

Root cause: Training with FP16 (torch.float16) compresses confidence score
distributions due to limited mantissa precision (~10-11 significant bits).
Token classification requires precise per-token probability distributions.

Fix: Force torch.float32 for all PII token classification training, ensuring
proper confidence score variance and accurate entity detection probabilities.

This fix complements PR vllm-project#648 which enables LoRA PII model auto-detection.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>
Signed-off-by: Yossi Ovadia <[email protected]>

Author: yossiovadia

src/training/training_lora/pii_model_fine_tuning_lora/pii_bert_finetuning_lora.py

@@ -149,10 +149,12 @@ def create_lora_token_model(model_name: str, num_labels: int, lora_config: dict)
         tokenizer.pad_token = tokenizer.eos_token
 
     # Load base model for token classification
+    # Force FP32 for proper confidence score distribution (FP16 causes uniform ~0.9 scores)
+    # Issue #647: FP16 limited precision (10-11 bits) compresses confidence distributions
     base_model = AutoModelForTokenClassification.from_pretrained(
         model_name,
         num_labels=num_labels,
-        dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
+        torch_dtype=torch.float32,  # Always use FP32 for stable token classification
     )
 
     # Create LoRA configuration for token classification