try extra urls

generate_summaries_uv.py

@@ -2,6 +2,7 @@
 # requires-python = ">=3.10"
 # dependencies = [
 #     "datasets",
+#     "flashinfer",
 #     "hf_transfer",
 #     "huggingface-hub[hf_xet]",
 #     "polars",
@@ -11,6 +12,12 @@
 #     "tqdm",
 #     "setuptools",
 # ]
+#
+# [[tool.uv.index]]
+# url = "https://flashinfer.ai/whl/cu126/torch2.6"
+#
+# [[tool.uv.index]]
+# url = "https://wheels.vllm.ai/nightly"
 # ///
 import argparse
 import logging
@@ -20,7 +27,7 @@ from typing import Optional
 
 # Set environment variables to speed up model loading
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
-# Note: VLLM will auto-select the best available attention backend
+os.environ["VLLM_ATTENTION_BACKEND"] = "FLASHINFER"
 
 import polars as pl
 from datasets import Dataset, load_dataset