fix(OOM): cap GPU blocks at 5000 via BI100_MAX_GPU_BLOCKS env var

Profiling zeros-out attention → vllm overestimates free memory → 7942 blocks
→ first real request OOMs. Cap at 5000 (80K tokens / 16 block_size).

patch_block_major_worker_capacity.py reads BI100_MAX_GPU_BLOCKS from env,
caps num_gpu_blocks after reserve_block_major_gpu_blocks.

computility-run.yaml

@@ -49,3 +49,5 @@ env:
       value: '1'
     - name: PYTORCH_CUDA_ALLOC_CONF
       value: max_split_size_mb:512
+    - name: BI100_MAX_GPU_BLOCKS
+      value: '5000'

qwen3_6_scripts/patch_block_major_worker_capacity.py

@@ -20,6 +20,14 @@ CAPACITY_ANCHOR = """\
 CAPACITY_REPLACEMENT = """\
         num_gpu_blocks = reserve_block_major_gpu_blocks(
             num_gpu_blocks, cache_block_size)
+        # BI100: cap GPU blocks — profiling with zero-tensor attention
+        # underestimates memory, causing runtime OOM if uncapped.
+        _bi100_max = int(os.environ.get("BI100_MAX_GPU_BLOCKS", "0"))
+        if _bi100_max > 0 and num_gpu_blocks > _bi100_max:
+            logger.warning(
+                "[BI100] capping num_gpu_blocks: %d -> %d (BI100_MAX_GPU_BLOCKS)",
+                num_gpu_blocks, _bi100_max)
+            num_gpu_blocks = _bi100_max
         num_gpu_blocks = max(num_gpu_blocks, 0)
         num_cpu_blocks = max(num_cpu_blocks, 0)
 """