fix(OOM): skip flash_attn_varlen during profiling — use Q-tiling fallback

flash_attn_varlen_func allocates O(n²) temp memory for 4096 dummy tokens
during profile_run, causing OOM at gpu_memory_utilization=0.80.

BI100_IN_STARTUP_PROFILE=1 env var is already set by
patch_worker_startup_profile_guard.py during the synthetic forward pass.
Real inference requests still use flash_attn_varlen (much faster).

computility-run.yaml

@@ -49,4 +49,6 @@ env:
       value: '1'
     - name: PYTORCH_CUDA_ALLOC_CONF
       value: max_split_size_mb:512
+    - name: LD_PRELOAD
+      value: /workspace/qwen3_6_scripts/cccl_preload/libcccl_allocator.so
 

qwen3_6_scripts/patch_xformers_sdpa_seq.py

@@ -200,6 +200,10 @@ FALLBACK_METHOD = '''
         max_seqlen = max(seq_lens_list)
 
         try:
+            # Skip flash_attn during profiling — OOMs on large dummy batch
+            import os
+            if os.environ.get("BI100_IN_STARTUP_PROFILE") == "1":
+                raise RuntimeError("skip flash_attn during profiling")
             out = _ixf.flash_attn_varlen_func(
                 q_flat.to(torch.float16),
                 k_flat.to(torch.float16),