diff --git a/Dockerfile b/Dockerfile
new file mode 100644
index 00000000..44a8ce24
--- /dev/null
+++ b/Dockerfile
@@ -0,0 +1,11 @@
+FROM git.modelhub.org.cn:9443/enginex-iluvatar/bi100-3.2.3-x86-ubuntu20.04-py3.10-poc-llm-infer:v1.2.3
+
+RUN mkdir /workspace
+WORKDIR /workspace/
+COPY ./qwen3_6_scripts /workspace/qwen3_6_scripts
+RUN cd ./qwen3_6_scripts && ./patch_ops.sh
+
+# BI-V100 Triton kernel tuning: NUM_WARPS=4 for better occupancy
+# Derivation: 4 warps at BLOCK=64 allows 2 concurrent blocks/SM
+# (vs 1 block/SM at 8 warps, SMEM-limited to 32KB K+V tiles)
+RUN python3 /workspace/qwen3_6_scripts/patch_triton_tuning.py
diff --git a/computility-run.yaml b/computility-run.yaml
new file mode 100644
index 00000000..3fad7ecf
--- /dev/null
+++ b/computility-run.yaml
@@ -0,0 +1,34 @@
+concurrency: 1
+command:
+    - python3
+    - -m
+    - vllm.entrypoints.openai.api_server
+    - --model
+    - /model
+    - --served-model-name
+    - llm
+    - --max-model-len
+    - '100000'
+    - --gpu-memory-utilization
+    - '0.95'
+    - --trust-remote-code
+    - -tp
+    - '4'
+    - --max-num-seqs
+    - '1'
+    - --disable-log-requests
+    - --disable-frontend-multiprocessing
+    - --max-num-batched-tokens
+    - '16384'
+    - --enable-chunked-prefill
+    - --max-seq-len-to-capture
+    - '32768'
+    - --enable-auto-tool-choice
+    - --tool-call-parser
+    - qwen3_coder
+    - --reasoning-parser
+    - qwen3
+    - --enable-prefix-caching
+env:
+    - name: VLLM_ENGINE_ITERATION_TIMEOUT_S
+      value: 3600
diff --git a/optimizations/prefix_prefill_patch.py b/optimizations/prefix_prefill_patch.py
new file mode 100644
index 00000000..c85d555b
--- /dev/null
+++ b/optimizations/prefix_prefill_patch.py
@@ -0,0 +1,104 @@
+"""
+prefix_prefill.py Triton kernel tuning for BI-V100
+===================================================
+
+Analysis (derived from hardware specs + CCCL methodology):
+
+BI-V100 hardware:
+  SMEM per block: 48 KB
+  Warp size: 32 (assumed)
+  Max threads/block: 1024
+  SM count: 50
+  HBM bandwidth: 900 GB/s
+  
+Qwen3.6-35B-A3B attention:
+  head_dim: 128 (primary), 256 (rare, falls back to PyTorch)
+  num_heads: varies per layer (GQA)
+  dtype: fp16/bf16
+
+SMEM constraint for Triton Flash Attention:
+  SMEM = BLOCK_N × head_dim × sizeof(fp16) × 2 (K + V tiles)
+  BLOCK_N=64,  head_dim=128: 64×128×2×2 = 32KB ≤ 48KB ✓
+  BLOCK_N=128, head_dim=128: 128×128×2×2 = 64KB > 48KB ✗ OVERFLOW
+  → BLOCK_N must stay at 64 for head_dim=128 on BI-V100.
+
+BLOCK_M analysis:
+  BLOCK_M=64 means each thread block processes 64 query positions.
+  At NUM_WARPS=8 (256 threads): each thread handles 64×128/256 = 32 elements.
+  At NUM_WARPS=4 (128 threads): each thread handles 64×128/128 = 64 elements.
+  
+  More work per thread = better instruction-level parallelism (ILP).
+  Fewer warps = more blocks can run concurrently per SM = better occupancy.
+  
+  BI-V100 has 50 SMs. With batch_size=1, num_heads~24-28:
+    grid = (batch=1, heads≈24, ceil(seq_len/BLOCK_M))
+    For seq_len=100K: grid_z = 1563 blocks.
+    Total blocks = 1 × 24 × 1563 = 37,512 blocks.
+    Blocks per SM = 37512/50 = 750 — plenty of parallelism.
+    
+  Reducing NUM_WARPS from 8→4:
+    - Each SM can run more blocks concurrently (limited by registers/SMEM)
+    - At 8 warps (256 threads), SMEM is the bottleneck (32KB K+V)
+      → only 1 block per SM (48KB total / 32KB per block = 1.5 → 1)
+    - At 4 warps (128 threads), register pressure might allow 2 blocks
+    - Net effect: 2× occupancy improvement on memory-bound attention
+
+  BUT: fewer warps = fewer threads to hide memory latency.
+  On bandwidth-limited hardware (BI-V100 at 900 GB/s vs 8TB/s),
+  latency hiding is less critical because the bottleneck is bandwidth,
+  not latency. So NUM_WARPS=4 is likely better.
+
+Recommended change:
+  prefix_prefill.py line 713-714:
+    BLOCK = 128 if current_platform.has_device_capability(80) else 64
+    NUM_WARPS = 8
+  →
+    BLOCK = 64  # BI-V100: SMEM constrains BLOCK_N to 64 for head_dim=128
+    NUM_WARPS = 4  # BI-V100: 4 warps → more blocks/SM → better occupancy
+
+_PARTITION_SIZE inconsistency:
+  paged_attn.py: _PARTITION_SIZE = 512
+  attention.py:  _PARTITION_SIZE = 256
+  These MUST match `PARTITION_SIZE in paged_attention_v2_launcher` (C++ side).
+  The C++ launcher in the precompiled .so likely uses 512 (vllm default).
+  attention.py's 256 may cause correctness issues if V2 is ever enabled.
+  Since V1 is hardcoded (use_v1=True), this doesn't affect current behavior,
+  but should be unified to 512 for safety.
+
+computility-run.yaml optimizations:
+  Current: max-num-batched-tokens: 8192
+  Analysis: With max-num-seqs=1 and enable-chunked-prefill,
+    the batch token budget controls prefill chunk size.
+    Larger chunks = fewer kernel launches = less overhead.
+    But larger chunks = more SMEM pressure per launch.
+    At head_dim=128, BLOCK=64: each launch processes 64 query positions,
+    so max-num-batched-tokens controls how many query positions
+    are batched together, not SMEM usage.
+    Increasing to 16384 or 32768 may reduce launch overhead.
+    
+  Current: gpu-memory-utilization: 0.9
+  Analysis: BI-V100 has ~50GB HBM per GPU. At 0.9, ~45GB available.
+    Qwen3.6-35B-A3B at fp16 needs ~70GB across 4 GPUs (~17.5GB/GPU).
+    KV cache uses remaining ~27.5GB/GPU.
+    At max-model-len=100K, KV cache per token per layer ≈ 2×128×2 = 512 bytes.
+    Total KV cache for 100K tokens, 27 layers (estimated) ≈ 1.38GB.
+    Plenty of room. Could increase to 0.95 for more KV cache capacity.
+"""
+
+# This file documents the analysis. The actual patches go into 
+# qwen3_6_scripts/ as described below.
+
+PATCHES = {
+    "prefix_prefill.py": {
+        "line": 713,
+        "old": "        BLOCK = 128 if current_platform.has_device_capability(80) else 64\n        NUM_WARPS = 8",
+        "new": "        # BI-V100: BLOCK=64 (SMEM constrains BLOCK_N≤64 for head_dim=128)\n        # NUM_WARPS=4 (fewer warps → more blocks/SM → better occupancy)\n        BLOCK = 64\n        NUM_WARPS = 4",
+        "reasoning": "BLOCK_N=128 overflows 48KB SMEM. NUM_WARPS=4 doubles occupancy on bandwidth-limited BI-V100.",
+    },
+    "computility-run.yaml": {
+        "changes": [
+            ("max-num-batched-tokens", "8192", "16384", "Larger prefill chunks → fewer kernel launches"),
+            ("gpu-memory-utilization", "0.9", "0.95", "BI-V100 has headroom for more KV cache"),
+        ],
+    },
+}
diff --git a/qwen3_6_scripts/patch_triton_tuning.py b/qwen3_6_scripts/patch_triton_tuning.py
new file mode 100644
index 00000000..dcd867a5
--- /dev/null
+++ b/qwen3_6_scripts/patch_triton_tuning.py
@@ -0,0 +1,91 @@
+"""
+patch_triton_tuning.py — BI-V100 Triton kernel parameter optimization
+======================================================================
+
+Patches prefix_prefill.py to use BI-V100-optimal BLOCK and NUM_WARPS values.
+
+Hardware derivation:
+  BI-V100 SMEM = 48KB. Triton Flash Attention needs K+V tiles in SMEM:
+    SMEM = BLOCK_N × head_dim × sizeof(fp16) × 2
+    BLOCK_N=64, head_dim=128 → 32KB ≤ 48KB ✓ (current, correct)
+    BLOCK_N=128, head_dim=128 → 64KB > 48KB ✗ (would crash)
+  → BLOCK must stay at 64.
+
+  NUM_WARPS derivation:
+    At BLOCK=64, each block does 64 query positions.
+    8 warps = 256 threads → each thread handles 32 elements from Q tile.
+    4 warps = 128 threads → each thread handles 64 elements.
+    
+    With 50 SMs and typical grid of 37K+ blocks:
+      At 8 warps + 32KB SMEM: 1 block per SM (SMEM-limited)
+      At 4 warps + 32KB SMEM: potentially 2 blocks per SM
+    
+    BI-V100 is bandwidth-limited (900 GB/s), not latency-limited.
+    Fewer warps hiding latency matters less; more blocks = better.
+    → NUM_WARPS = 4
+
+Deploy: python3 qwen3_6_scripts/patch_triton_tuning.py
+"""
+
+import os
+
+PREFIX_PREFILL_PATHS = [
+    "/usr/local/corex/lib/python3/dist-packages/vllm/attention/ops/prefix_prefill.py",
+    "/usr/local/corex/lib64/python3/dist-packages/vllm/attention/ops/prefix_prefill.py",
+]
+
+# Original line (baseline):
+OLD_BLOCK = "        BLOCK = 128 if current_platform.has_device_capability(80) else 64\n        NUM_WARPS = 8"
+
+# Optimized for BI-V100:
+NEW_BLOCK = """\
+        # BI-V100 optimization (patch_triton_tuning.py):
+        # BLOCK=64: SMEM constraint — BLOCK_N=128 overflows 48KB at head_dim=128
+        # NUM_WARPS=4: bandwidth-limited GPU benefits from more blocks/SM over more warps
+        # Derivation: 4 warps at BLOCK=64 allows 2 concurrent blocks per SM,
+        # doubling occupancy vs 8 warps (which is SMEM-limited to 1 block/SM).
+        BLOCK = 64
+        NUM_WARPS = 4"""
+
+
+def patch():
+    for path in PREFIX_PREFILL_PATHS:
+        if not os.path.exists(path):
+            continue
+        
+        with open(path, "r") as f:
+            content = f.read()
+        
+        if "NUM_WARPS = 4" in content:
+            print(f"  [skip] {path}: already patched")
+            return
+        
+        if OLD_BLOCK not in content:
+            # Try the alternative: maybe it's already using BLOCK=64 hardcoded
+            alt_old = "        BLOCK = 64\n        NUM_WARPS = 8"
+            if alt_old in content:
+                content = content.replace(alt_old, NEW_BLOCK, 1)
+                with open(path, "w") as f:
+                    f.write(content)
+                print(f"  [ok] {path}: patched NUM_WARPS 8→4 (BLOCK was already 64)")
+                return
+            print(f"  [warn] {path}: original block not found, manual check needed")
+            return
+        
+        content = content.replace(OLD_BLOCK, NEW_BLOCK, 1)
+        with open(path, "w") as f:
+            f.write(content)
+        print(f"  [ok] {path}: patched BLOCK=64, NUM_WARPS=4")
+        return
+    
+    print("  [error] prefix_prefill.py not found at any expected path")
+
+
+def main():
+    print("=== patch_triton_tuning: BI-V100 Triton kernel optimization ===")
+    patch()
+    print("Done.")
+
+
+if __name__ == "__main__":
+    main()