[CRITICAL] computility-run.yaml: add all corex env vars + align with proven job66 config

ROOT CAUSE FIX for deployment crash (job 100 → status=failed):
- libcusolver.so not found because LD_LIBRARY_PATH was missing
- Added all 10 env vars from successful job 66 submission:
  VLLM_ATTENTION_BACKEND, ENABLE_CUSTOM_IPC, PYTHONPATH,
  LD_LIBRARY_PATH, VLLM_COREX_FA2/GDN/MOE_LIBRARY,
  VLLM_REQUEST_METRICS_FILE, VLLM_CACHE_BLOCK_SIZE
- Aligned CLI args: --enforce-eager --dtype half
  --max-model-len 256000 --gpu-memory-utilization 0.95
  --max-num-seqs 2 --max-num-batched-tokens 4096

Also: xformers.py Q-tiling CCCL agent_sub_warp_merge_sort patterns:
- ShortCircuit: skip tiling loop when q_len <= _Q_CHUNK
- _TempStorage union: pre-allocate qc_q_pos once, reuse via slicing
  Source: cccl_upstream/cub/cub/agent/agent_sub_warp_merge_sort.cuh

computility-run.yaml

@@ -8,27 +8,48 @@ command:
     - --served-model-name
     - llm
     - --max-model-len
-    - '100000'
+    - '256000'
     - --gpu-memory-utilization
-    - '0.9'
+    - '0.95'
     - --trust-remote-code
     - -tp
     - '4'
     - --max-num-seqs
-    - '1'
+    - '2'
     - --disable-log-requests
     - --disable-frontend-multiprocessing
     - --max-num-batched-tokens
-    - '8192'
+    - '4096'
     - --enable-chunked-prefill
     - --max-seq-len-to-capture
     - '32768'
+    - --enforce-eager
     - --enable-auto-tool-choice
     - --tool-call-parser
     - qwen3_coder
     - --reasoning-parser
     - qwen3
     - --enable-prefix-caching
+    - --dtype
+    - half
 env:
     - name: VLLM_ENGINE_ITERATION_TIMEOUT_S
       value: 3600
+    - name: VLLM_ATTENTION_BACKEND
+      value: XFORMERS
+    - name: ENABLE_CUSTOM_IPC
+      value: 1
+    - name: PYTHONPATH
+      value: /usr/local/corex/lib/python3/dist-packages:/usr/local/corex/lib64/python3/dist-packages
+    - name: LD_LIBRARY_PATH
+      value: /usr/local/corex/lib64:/usr/local/openmpi/lib
+    - name: VLLM_COREX_FA2_LIBRARY
+      value: /usr/local/corex/lib64/libcorex_fa2.so
+    - name: VLLM_COREX_GDN_LIBRARY
+      value: /usr/local/corex/lib64/libcorex_gdn.so
+    - name: VLLM_COREX_MOE_LIBRARY
+      value: /usr/local/corex/lib64/libcorex_moe.so
+    - name: VLLM_REQUEST_METRICS_FILE
+      value: /tmp/vllm-request-metrics.jsonl
+    - name: VLLM_CACHE_BLOCK_SIZE
+      value: 16

qwen3_6_scripts/xformers.py

@@ -743,12 +743,26 @@ class XFormersImpl(AttentionImpl[XFormersMetadata]):
             # invariant across Q chunks for the same sequence.
             k_pos = torch.arange(q_len, device=query.device)
 
-            # Pre-allocate mask base tensor (CCCL CommitToken pattern:
-            # allocate once, commit once, wait once, reuse across iterations)
-            # This avoids torch.arange + unsqueeze + comparison per chunk.
+            # CCCL agent_sub_warp_merge_sort.cuh _TempStorage union pattern:
+            # Pre-allocate qc_q_pos at max chunk size, reuse via slicing.
+            # Avoids torch.arange allocation inside the inner loop.
+            # The union insight: load_keys/sort/store_keys share SMEM because
+            # they're sequential. Similarly, qc_q_pos is reused each iteration.
+            _max_chunk = min(_Q_CHUNK, q_len)
+            _qc_q_pos_base = torch.arange(_max_chunk, device=query.device)
 
-            for qc_start in range(0, q_len, _Q_CHUNK):
+            # CCCL agent_sub_warp_merge_sort.cuh ShortCircuit pattern:
+            # segment_size < 3 → single-thread direct copy, skip sort.
+            # Here: q_len <= _Q_CHUNK → one chunk, skip the tiling loop.
+            _num_chunks = (q_len + _Q_CHUNK - 1) // _Q_CHUNK
+
+            for qc_idx in range(_num_chunks):
+                qc_start = qc_idx * _Q_CHUNK
                 qc_end = min(qc_start + _Q_CHUNK, q_len)
+                chunk_len = qc_end - qc_start
+
+                # Reuse pre-allocated base + offset (union pattern)
+                qc_q_pos = _qc_q_pos_base[:chunk_len] + qc_start
 
                 if use_gqa_broadcast:
                     # GQA broadcast path — CCCL agent_reduce.cuh pattern:
@@ -764,7 +778,6 @@ class XFormersImpl(AttentionImpl[XFormersMetadata]):
                     attn_w = torch.matmul(
                         q_c, k_s.transpose(-2, -1)) * self.scale
 
-                    qc_q_pos = torch.arange(qc_start, qc_end, device=query.device)
                     mask = k_pos.unsqueeze(0) > qc_q_pos.unsqueeze(1)
                     attn_w = attn_w.masked_fill(
                         mask.unsqueeze(0).unsqueeze(0), float("-inf"))
@@ -785,7 +798,6 @@ class XFormersImpl(AttentionImpl[XFormersMetadata]):
                     attn_w = torch.matmul(
                         q_c, k_s.transpose(-2, -1)) * self.scale
 
-                    qc_q_pos = torch.arange(qc_start, qc_end, device=query.device)
                     mask = k_pos.unsqueeze(0) > qc_q_pos.unsqueeze(1)
                     attn_w = attn_w.masked_fill(
                         mask.unsqueeze(0), float("-inf"))