[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/layers/pooler.py

@@ -0,0 +1,63 @@
+from enum import IntEnum
+
+import torch
+import torch.nn as nn
+
+from vllm.model_executor.pooling_metadata import (PoolingMetadata,
+                                                  PoolingTensors)
+from vllm.sequence import EmbeddingSequenceGroupOutput, PoolerOutput
+
+
+class PoolingType(IntEnum):
+    """Enumeration for different types of pooling methods."""
+    LAST = 0
+    ALL = 1
+
+
+class Pooler(nn.Module):
+    """A layer that pools specific information from hidden states.
+
+    This layer does the following:
+    1. Extracts specific tokens or aggregates data based on pooling method.
+    2. Normalizes output if specified.
+    3. Returns structured results as `PoolerOutput`.
+
+    Attributes:
+        pooling_type: The type of pooling to use (LAST, AVERAGE, MAX).
+        normalize: Whether to normalize the pooled data.
+    """
+
+    def __init__(self, pooling_type: PoolingType, normalize: bool):
+        super().__init__()
+        self.pooling_type = pooling_type
+        self.normalize = normalize
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        pooling_metadata: PoolingMetadata,
+    ) -> PoolerOutput:
+        """Pools specific information from hidden states based on metadata."""
+        prompt_lens = PoolingTensors.from_pooling_metadata(
+            pooling_metadata, hidden_states.device).prompt_lens
+
+        if self.pooling_type == PoolingType.LAST:
+            last_token_flat_indices = torch.cumsum(prompt_lens, dim=0) - 1
+            pooled_data = hidden_states[last_token_flat_indices]
+        elif self.pooling_type == PoolingType.ALL:
+            offset = 0
+            pooled_data = []
+            for prompt_len in prompt_lens:
+                pooled_data.append(hidden_states[offset:offset + prompt_len])
+                offset += prompt_len
+        else:
+            raise ValueError(f"Invalid pooling type: {self.pooling_type}")
+
+        if self.normalize:
+            pooled_data = nn.functional.normalize(pooled_data, p=2, dim=1)
+
+        pooled_outputs = [
+            EmbeddingSequenceGroupOutput(data.tolist()) for data in pooled_data
+        ]
+
+        return PoolerOutput(outputs=pooled_outputs)