[DEPLOY] Complete submission: baseline + all optimizations

Adds ALL files needed for Dockerfile build:
  - qwen3_6_scripts/ (baseline patches + our optimizations)
  - vllm/ (full vllm package)
  - paged_attention_v2_pytorch.py (V2 with single-bmm optimization)
  - Dockerfile + computility-run.yaml

Our optimizations vs baseline:
  1. paged_attn.py: pre-gathered context KV (eliminates 194 gather calls),
     Triton try/fallback, V2 heuristic, threshold 32K→64K
  2. paged_attention_v2_pytorch.py: fills NotImplementedError,
     single-bmm Phase 1 (195 launches → 3)
  3. patch_enable_triton.py: HAS_TRITON=True with safety fallback
  4. patch_triton_tuning.py: BLOCK=64, NUM_WARPS=4 for BI-V100
  5. computility-run.yaml: gpu-memory-utilization 0.9→0.95,
     max-num-batched-tokens 8192→16384

This repo can now be submitted to dev.modelhub.org.cn as-is.

vllm/model_executor/pooling_metadata.py

@@ -0,0 +1,69 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Tuple
+
+import torch
+
+from vllm.pooling_params import PoolingParams
+from vllm.utils import is_pin_memory_available
+
+
+class PoolingMetadata:
+    """Metadata for pooling operations in the Pooler layer.
+
+    This class holds the necessary information for pooling operations,
+    providing context for how to perform pooling and other related operations.
+
+    Attributes:
+        seq_groups: List of (seq_ids, pooling_params).
+        seq_data: A mapping of sequence ID to additional sequence data.
+        prompt_lens: List of the lengths of each prompt.
+    """
+
+    def __init__(
+        self,
+        seq_groups: List[Tuple[List[int], PoolingParams]],
+        seq_data: Dict[int, Any],  # Specific data related to sequences
+        prompt_lens: List[int],
+    ) -> None:
+        self.seq_groups = seq_groups
+        self.seq_data = seq_data
+        self.prompt_lens = prompt_lens
+
+    def __repr__(self) -> str:
+        return ("PoolingMetadata("
+                f"seq_groups={self.seq_groups}, "
+                f"seq_data={self.seq_data}, "
+                f"prompt_lens={self.prompt_lens})")
+
+
+@dataclass
+class PoolingTensors:
+    """Tensors for pooling."""
+
+    prompt_lens: torch.Tensor
+
+    @classmethod
+    def from_pooling_metadata(
+        cls,
+        pooling_metadata: "PoolingMetadata",
+        device: torch.device,
+    ) -> "PoolingTensors":
+        """
+        Create PoolingTensors from PoolingMetadata.
+
+        Args:
+            pooling_metadata: PoolingMetadata instance to convert.
+            device: Device to store the tensors.
+        """
+        # Convert prompt lengths to tensor
+        pin_memory = is_pin_memory_available()
+
+        prompt_lens_t = torch.tensor(
+            pooling_metadata.prompt_lens,
+            device="cpu",
+            dtype=torch.long,
+            pin_memory=pin_memory,
+        )
+
+        return cls(prompt_lens=prompt_lens_t.to(device=device,
+                                                non_blocking=True), )