Update requirements.txt

requirements.txt

@@ -1,5 +1,10 @@
-gradio==4.21.0
-transformers==4.47.1        # any ≥4.44 works; 4.47.1 tested May-2025
-accelerate==0.30.0
-torch==2.5.1                # CPU wheel, inside ZeroGPU’s allowed range
-sentencepiece               # tokenizer dep for Llama models
+--extra-index-url https://download.pytorch.org/whl/cu124   # grab a CUDA Torch wheel
+torch==2.5.1+cu124                                         # keep before flash-attn
+
+# FlashAttention pre-built wheel that matches:  Torch 2.5  •  CUDA 12  •  cp310
+https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.0.post2/flash_attn-2.8.0.post2+cu12torch2.5cxx11abiFALSE-cp310-cp310-linux_x86_64.whl  # <- 240 MB wheel:contentReference[oaicite:2]{index=2}
+
+transformers>=4.52.0
+accelerate>=0.30.2          # bug-fix for device_map edge case:contentReference[oaicite:3]{index=3}
+gradio>=4.44.0              # Zero-GPU queue fix PR #5698:contentReference[oaicite:4]{index=4}
+sentencepiece