Upload folder using huggingface_hub

examples/prt14_qwen25vl/configs/config.yaml

@@ -3,7 +3,7 @@ defaults:
 
 model:
   model_id: "Qwen/Qwen2-VL-2B-Instruct"
-  bf16: true
+  bf16: false
   use_lora: true
 
 data:
@@ -19,7 +19,7 @@ generation:
 training:
   output_dir: "output_prt14"
   batch_size: 1
-  gradient_accumulation_steps: 8
+  gradient_accumulation_steps: 1
   learning_rate: 2.0e-4
   epochs: 1
   save_steps: 100

examples/prt14_qwen25vl/train_prt14.py

@@ -24,6 +24,13 @@ from transformers import (
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
 
+# Enforce T4-safe attention settings (Avoids OOM in SDPA)
+# T4 does not support Flash Attention 2, and mem_efficient can sometimes use more memory or fail.
+# Math (vanilla) Attention is the safest fallback.
+torch.backends.cuda.enable_flash_sdp(False)
+torch.backends.cuda.enable_mem_efficient_sdp(False)
+torch.backends.cuda.enable_math_sdp(True)
+
 @dataclass
 class PRTTrainingArguments(TrainingArguments):
     """PRT固有のパラメータを追加したTrainingArguments"""