Sync from GitHub via hub-sync

hunyuan-ocr-1.5.py

@@ -4,14 +4,16 @@
 #     "datasets>=4.0.0",
 #     "huggingface-hub",
 #     "pillow",
-#     "vllm>=0.18.1",
-#     "transformers<5.13",  # vLLM ≤0.24.0's HunyuanVL processor breaks on transformers 5.13
-#                           # (string-key AutoImageProcessor.register; fixed in vllm#47872).
-#                           # Drop this cap once that fix ships in a stable vLLM release.
 #     "tqdm",
 #     "toolz",
-#     "torch",
 # ]
+#
+# [tool.hf-jobs]
+# image = "vllm/vllm-openai:v0.24.0"
+# python = "/usr/bin/python3"
+# env = { PYTHONPATH = "/usr/local/lib/python3.12/dist-packages" }
+# flavor = "a10g-small"
+# secrets = ["HF_TOKEN"]
 # ///
 
 """
@@ -42,9 +44,12 @@ Model: tencent/HunyuanOCR
 
 vLLM: 0.18.1 (release) is the first stable wheel with native
   `HunYuanVLForConditionalGeneration` support for autoregressive decoding — no
-  nightly or patch needed for batch OCR. The floor stays at 0.18.1; a bare
-  `vllm` resolves to the latest stable (0.24.0 as of 2026-07), which also works
-  once transformers is capped <5.13 (see the deps block for why). The DFlash
+  nightly or patch needed for batch OCR. The [tool.hf-jobs] header pins the
+  vllm/vllm-openai:v0.24.0 image (`hf` CLI 1.32+), which also sets the hardware
+  and the HF_TOKEN secret. Newer stacks fail: unpinned vLLM with transformers
+  >=5.13 does not recognise `hunyuan_vl`, and the v0.29.0 image fails at engine
+  start ("Expected 4 multimodal RoPE channels"). To run on your own GPU:
+  `uv run --with vllm==0.24.0 --with "transformers<5.13" ...`. The DFlash
   speculative-decoding draft (a per-request *latency* win that needs a vLLM
   nightly) is intentionally NOT implemented: it does not change offline batch
   throughput or output distribution.
@@ -500,6 +505,9 @@ def main(
         max_model_len=max_model_len,
         gpu_memory_utilization=gpu_memory_utilization,
         limit_mm_per_prompt={"image": 1},
+        # The encoder cache is sized from max_num_batched_tokens (8192 by default), but one
+        # image can reach img_max_token_num=16384 tokens; a 2000 px scan already needs ~8.6k.
+        max_num_batched_tokens=16384,
     )
 
     # Locked sampling per the model card (deterministic OCR); only repetition_penalty
@@ -698,10 +706,8 @@ if __name__ == "__main__":
             "   uv run hunyuan-ocr-1.5.py en-docs zh-docs --task-type doc_trans_en2zh"
         )
         print("\n6. Running on HF Jobs:")
-        print("   hf jobs uv run --flavor l4x1 \\")
-        print(
-            '     -e HF_TOKEN=$(python3 -c "from huggingface_hub import get_token; print(get_token())") \\'
-        )
+        print("   (image, hardware and HF_TOKEN come from the script's [tool.hf-jobs] header)")
+        print("   hf jobs uv run \\")
         print(
             "     https://huggingface.co/datasets/uv-scripts/ocr/raw/main/hunyuan-ocr-1.5.py \\"
         )

ocr-vllm-judge.py

@@ -5,11 +5,16 @@
 #     "datasets>=4.0.0",
 #     "huggingface-hub",
 #     "pillow",
-#     "vllm>=0.15.1",
-#     "torch",
 #     "rich",
 #     "tqdm",
 # ]
+#
+# [tool.hf-jobs]
+# image = "vllm/vllm-openai:v0.29.0"
+# python = "/usr/bin/python3"
+# env = { PYTHONPATH = "/usr/local/lib/python3.12/dist-packages" }
+# flavor = "a100-large"
+# secrets = ["HF_TOKEN"]
 # ///
 """
 Offline vLLM judge for OCR benchmark evaluation.
@@ -35,8 +40,10 @@ Usage:
         --judge-model Qwen/Qwen2.5-VL-7B-Instruct \\
         --max-samples 50
 
-    # Via HF Job
-    hf jobs uv run --flavor l4x1 -s HF_TOKEN \\
+    # Via HF Job. The [tool.hf-jobs] header pins vllm/vllm-openai:v0.29.0 (the
+    # default uv image has no nvcc, which vLLM needs at warmup) and a100-large
+    # (the default 7B judge does not leave room for the KV cache on a 24 GB L4).
+    hf jobs uv run --timeout 1h \\
         ocr-vllm-judge.py davanstrien/ocr-bench-nls-50 --from-prs \\
         --judge-model Qwen/Qwen3-VL-8B-Instruct --max-samples 50
 """
@@ -636,7 +643,7 @@ Examples:
       --max-samples 50
 
   # Via HF Job
-  hf jobs uv run --flavor l4x1 -s HF_TOKEN \\
+  hf jobs uv run --timeout 1h \\
       ocr-vllm-judge.py my-bench --from-prs \\
       --judge-model Qwen/Qwen3-VL-8B-Instruct
         """,
@@ -695,6 +702,9 @@ Examples:
         default=None,
         help="Push judge results to this HF dataset repo (e.g. davanstrien/ocr-bench-rubenstein-judge)",
     )
+    parser.add_argument(
+        "--private", action="store_true", help="Create the --save-results repo as private"
+    )
     args = parser.parse_args()
 
     # --- CUDA check ---
@@ -776,7 +786,7 @@ Examples:
 
         # Comparisons config
         comp_ds = Dataset.from_list(comparison_log)
-        comp_ds.push_to_hub(args.save_results, config_name="comparisons")
+        comp_ds.push_to_hub(args.save_results, config_name="comparisons", private=args.private)
         console.print(f"  Pushed [cyan]comparisons[/cyan] ({len(comparison_log)} rows)")
 
         # Leaderboard config
@@ -795,7 +805,7 @@ Examples:
                 }
             )
         Dataset.from_list(leaderboard_rows).push_to_hub(
-            args.save_results, config_name="leaderboard"
+            args.save_results, config_name="leaderboard", private=args.private
         )
         console.print(
             f"  Pushed [cyan]leaderboard[/cyan] ({len(leaderboard_rows)} rows)"
@@ -813,7 +823,7 @@ Examples:
             "from_prs": args.from_prs,
         }
         Dataset.from_list([metadata_row]).push_to_hub(
-            args.save_results, config_name="metadata"
+            args.save_results, config_name="metadata", private=args.private
         )
         console.print("  Pushed [cyan]metadata[/cyan]")
         console.print(f"  [green]Results saved to: {args.save_results}[/green]")