Update grpo_train.py

grpo_train.py

@@ -351,8 +351,8 @@ else:
     _name = "CPU"
     _cc = (0, 0)
 
-USE_4BIT = _vram < 40 * 1024**3   # T4 (15 GB), L4 (24 GB) → 4-bit; A100 (80 GB) → bf16
-USE_BF16 = _cc >= (8, 0)          # Ampere+ (A100, L4) support bf16; Turing (T4) does not
+USE_4BIT = _vram < 40 * 1024**3   # T4 (15 GB), L4 (24 GB) → 4-bit; A100 (80 GB) → full
+USE_BF16 = _cc >= (8, 0) and not USE_4BIT  # bf16 only when full-precision; 4-bit LoRA uses fp16 internally
 
 # #region agent log
 _dlog("A", "grpo_train.py:gpu_detect", "GPU config resolved", {"name":_name,"vram_gb":round(_vram/1024**3,1),"cc":list(_cc),"USE_4BIT":USE_4BIT,"USE_BF16":USE_BF16})