feat: add reproduce.md + rename reproducibility.lock.yaml to REPRO.yaml

Author: AIwork4me

reproducibility.lock.yaml REPRO.yamlreproducibility.lock.yaml renamed to REPRO.yaml

@@ -0,0 +1,69 @@
+---
+model_id: hunyuan-ocr
+backend: vllm
+hardware:
+  gpu: "AMD gfx1100"
+  vram_min_gb: 48
+environment:
+  type: docker
+  rocm: "7.2"
+command: |
+  # 1. Start vLLM server:
+  vllm serve tencent/HunyuanOCR --host 0.0.0.0 --port 10000 --max-model-len 32768
+
+  # 2. Run adapter:
+  python adapter/run_adapter.py \
+    --platform linux-rocm --backend vllm \
+    --server-url http://127.0.0.1:10000/v1 \
+    --model tencent/HunyuanOCR \
+    --img-dir /root/datasets/OmniDocBench_data/images \
+    --out-dir /tmp/hunyuanocr-predictions
+
+  # 3. Score + publish (via omnidocbench-rocm):
+  omnidocbench-rocm run --stage score --platform linux-rocm --cdm \
+    --predictions-dir /tmp/hunyuanocr-predictions \
+    --run-stats /tmp/hunyuanocr-predictions/_run_stats.json --version v16
+expected_overall:
+  value: 93.64
+  tolerance: 0.5
+---
+
+# Reproduce HunyuanOCR 93.64 on AMD ROCm
+
+## Prerequisites
+
+```bash
+rocminfo | grep -E "Name:|VRAM"    # must show gfx1100 + ≥48 GB
+ls -la /dev/kfd                     # must exist
+```
+
+## Quickstart
+
+Full 1,651-page evaluation with CDM formula scoring.
+
+```bash
+# Start vLLM server on AMD GPU
+vllm serve tencent/HunyuanOCR --host 0.0.0.0 --port 10000 --max-model-len 32768
+
+# In another terminal, run adapter
+cd /path/to/HunyuanOCR-ROCm
+python adapter/run_adapter.py \
+  --platform linux-rocm --backend vllm \
+  --server-url http://127.0.0.1:10000/v1 \
+  --model tencent/HunyuanOCR \
+  --img-dir /root/datasets/OmniDocBench_data/images \
+  --out-dir /tmp/hunyuanocr-predictions
+
+# Score
+omnidocbench-rocm run --stage score --platform linux-rocm --cdm \
+  --predictions-dir /tmp/hunyuanocr-predictions \
+  --run-stats /tmp/hunyuanocr-predictions/_run_stats.json --version v16
+```
+
+## Expected output
+
+Overall **93.64** (±0.5). Text 95.48, Table TEDS 92.97%, Formula CDM 92.46%.
+
+## If it fails
+
+See [OmniDocBench-ROCm pitfalls](https://github.com/AIwork4me/OmniDocBench-ROCm/docs/pitfalls.md).