nano-vllm/nanovllm/kvcache/sparse/xattn_bsa.py

"""
XAttention Block Sparse Attention (BSA) Policy for nano-vllm.

This module implements XAttention-inspired block sparse attention for chunked prefill.

Key design:
1. Use xattn_estimate_chunked to estimate sparse block mask
2. Use BSA kernel for efficient sparse attention computation
3. Support chunked prefill with q_start_pos for correct position handling

Note: Decode phase is not supported - use FullAttentionPolicy for decode.
"""

import logging
import torch
import torch.cuda.nvtx as nvtx
from typing import List, Tuple, TYPE_CHECKING

from nanovllm.kvcache.sparse.policy import SparsePolicy, PolicyContext
from nanovllm.utils.density_observer import DensityObserver

if TYPE_CHECKING:
    from nanovllm.kvcache.offload_engine import OffloadEngine
    from nanovllm.kvcache.manager import KVCacheManager
    from nanovllm.engine.sequence import Sequence

logger = logging.getLogger(__name__)

# Check BSA availability
try:
    from block_sparse_attn import block_sparse_attn_func
    BSA_AVAILABLE = True
except ImportError:
    BSA_AVAILABLE = False
    logger.warning("block_sparse_attn not available, XAttentionBSAPolicy will fallback to dense")

# Check xattn_estimate_chunked availability
try:
    from nanovllm.ops.xattn import xattn_estimate_chunked
    XATTN_AVAILABLE = True
except ImportError:
    XATTN_AVAILABLE = False
    logger.warning("xattn_estimate_chunked not available")


def expand_kv_for_gqa(
    key_states: torch.Tensor,
    value_states: torch.Tensor,
    num_heads: int,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Expand KV for Grouped Query Attention.

    Args:
        key_states: [B, num_kv_heads, seq_len, head_dim]
        value_states: [B, num_kv_heads, seq_len, head_dim]
        num_heads: Number of query heads

    Returns:
        Expanded (key, value) with shape [B, num_heads, seq_len, head_dim]
    """
    num_kv_heads = key_states.shape[1]
    if num_heads == num_kv_heads:
        return key_states, value_states
    num_groups = num_heads // num_kv_heads
    return (
        key_states.repeat_interleave(num_groups, dim=1),
        value_states.repeat_interleave(num_groups, dim=1),
    )


class XAttentionBSAPolicy(SparsePolicy):
    """
    XAttention Block Sparse Attention policy for chunked prefill.

    Uses xattn_estimate_chunked to estimate sparse mask, then BSA kernel
    for efficient sparse attention computation.

    Note:
        - Only supports prefill phase (decode uses FullAttentionPolicy)
        - BSA block size is fixed at 128 tokens
    """

    supports_prefill = True
    supports_decode = False  # Decode uses FullAttentionPolicy
    requires_block_selection = False  # Selection happens internally

    # BSA requires 128-token blocks
    BSA_BLOCK_SIZE = 128

    def __init__(
        self,
        threshold: float = 0.95,  # High threshold for accuracy testing
        stride: int = 8,
        chunk_size: int = 16384,
        block_size: int = 128,
        samples_per_chunk: int = 128,
        use_triton: bool = True,
        estimate_block_size: int = 1024,  # Optimized block size for softmax_fuse_block_sum
    ):
        """
        Initialize XAttention BSA policy.

        Args:
            threshold: Cumulative attention threshold for block selection (0-1)
                       Higher values = more blocks selected = less sparse
            stride: Stride for Q/K reshape in estimation (typically 8)
            chunk_size: Processing chunk size for xattn_estimate (Triton alignment)
            block_size: BSA block size (must be 128)
            samples_per_chunk: Samples per chunk for estimation (unused)
            use_triton: Whether to use Triton kernels
            estimate_block_size: Block size for softmax_fuse_block_sum in select_blocks.
                                 Default 1024 is optimal (15x faster than 4096).
                                 Must be a factor of cpu_block_size (e.g., 4096/1024=4).
        """
        self.threshold = threshold
        self.stride = stride
        self.chunk_size = chunk_size
        self.use_triton = use_triton
        self.estimate_block_size = estimate_block_size
        self._num_heads = None  # Set during first forward

        # Sparse metadata: stores attention scores per layer
        # Dict[layer_id, Tensor[num_q_blocks, num_k_blocks]]
        self.sparse_metadata: dict = {}

        # Statistics for density tracking
        self._stats_total_available_blocks = 0
        self._stats_total_selected_blocks = 0
        self._stats_num_chunks = 0

        # Pre-allocated GQA expansion buffers (GPU-only mode)
        # Set by alloc_policy_metadata(), None if not pre-allocated
        self._k_expanded: torch.Tensor | None = None
        self._v_expanded: torch.Tensor | None = None
        self._max_seq_len: int = 0

        # Pre-allocated mask buffer for chunked prefill (offload mode)
        # Stores BSA-level mask from select_blocks for use in compute_chunked_prefill
        # Shape: [1, num_heads, max_q_bsa_blocks, max_k_bsa_blocks]
        self._prefill_mask_buffer: torch.Tensor | None = None
        self._current_mask_q_bsa: int = 0  # Current Q BSA blocks in buffer
        self._current_mask_k_bsa: int = 0  # Current K BSA blocks in buffer

        # Selected block indices for mask extraction in compute_chunked_prefill
        # Stores the indices of selected CPU blocks in available_blocks
        self._selected_cpu_indices: List[int] = []
        self._bsa_per_cpu: int = 0  # BSA blocks per CPU block
        
        #> Debug: store all K cache and density counts
        self._debug_k_full: torch.Tensor | None = None
        self._debug_selected: int = 0  # 累积的 selected blocks
        self._debug_total: int = 0     # 累积的 total blocks

    def alloc_policy_metadata(
        self,
        num_heads: int,
        num_kv_heads: int,
        head_dim: int,
        max_seq_len: int,
        dtype: torch.dtype,
        device: torch.device,
    ) -> None:
        """
        Pre-allocate GQA expansion buffers for GPU-only mode.

        These buffers are used by compute_prefill() to avoid dynamic allocation
        during forward pass. The buffers are sized for max_seq_len and sliced
        to actual seq_len during use.

        Memory usage: 2 * num_heads * max_seq_len * head_dim * dtype_size
        For 64K seq, 32 heads, 128 dim, fp16: 2 * 32 * 65536 * 128 * 2 = 1 GB

        Args:
            num_heads: Number of query heads
            num_kv_heads: Number of KV heads (for GQA)
            head_dim: Dimension per head
            max_seq_len: Maximum sequence length
            dtype: Data type
            device: Target device
        """
        # Pre-allocate mask buffer for chunked prefill (offload mode)
        # mask shape: [1, num_heads, max_q_bsa_blocks, max_k_bsa_blocks]
        # This is needed regardless of GQA
        max_q_bsa_blocks = self.chunk_size // self.BSA_BLOCK_SIZE
        max_k_bsa_blocks = max_seq_len // self.BSA_BLOCK_SIZE
        mask_shape = (1, num_heads, max_q_bsa_blocks, max_k_bsa_blocks)
        self._prefill_mask_buffer = torch.empty(mask_shape, dtype=torch.bool, device=device)
        mask_memory_mb = num_heads * max_q_bsa_blocks * max_k_bsa_blocks / (1024 * 1024)
        logger.info(f"[XAttn] Pre-allocated mask buffer: shape={mask_shape}, memory={mask_memory_mb:.1f} MB")

        # Only allocate GQA expansion buffers if GQA (num_heads != num_kv_heads)
        if num_heads == num_kv_heads:
            logger.info(f"[XAttn] No GQA expansion needed (num_heads == num_kv_heads = {num_heads})")
            return

        # Shape: [1, num_heads, max_seq_len, head_dim] for xattn_estimate format
        # Also used for BSA which expects [seq_len, num_heads, head_dim]
        shape = (1, num_heads, max_seq_len, head_dim)
        self._k_expanded = torch.empty(shape, dtype=dtype, device=device)
        self._v_expanded = torch.empty(shape, dtype=dtype, device=device)
        self._max_seq_len = max_seq_len

        memory_mb = 2 * num_heads * max_seq_len * head_dim * dtype.itemsize / (1024 * 1024)
        logger.info(f"[XAttn] Pre-allocated GQA buffers: shape={shape}, memory={memory_mb:.1f} MB")
        
        #DEBUG : buffer for save all K cache
        self._debug_k_full = torch.empty((1, num_heads, max_seq_len, head_dim), dtype=dtype, device=device)
        self._debug_selected = 0
        self._debug_total = 0

    # =========================================================================
    # GPU-only methods (non-chunked)
    # =========================================================================

    def compute_prefill(
        self,
        q: torch.Tensor,
        k: torch.Tensor,
        v: torch.Tensor,
        cu_seqlens_q: torch.Tensor,
        cu_seqlens_k: torch.Tensor,
        max_seqlen_q: int,
        max_seqlen_k: int,
        softmax_scale: float,
        layer_id: int,
        block_tables: torch.Tensor = None,
    ) -> torch.Tensor:
        """
        GPU-only prefill attention using XAttention + BSA.

        This method implements sparse attention for GPU-only mode:
        1. Estimate block importance using xattn_estimate
        2. Compute sparse attention using block_sparse_attn_func

        Args:
            q: Query tensor [total_q, num_heads, head_dim] (varlen packed)
            k: Key tensor [total_kv, num_kv_heads, head_dim] (varlen packed)
            v: Value tensor [total_kv, num_kv_heads, head_dim] (varlen packed)
            cu_seqlens_q: Cumulative sequence lengths for Q [batch+1]
            cu_seqlens_k: Cumulative sequence lengths for K [batch+1]
            max_seqlen_q: Maximum Q sequence length
            max_seqlen_k: Maximum K sequence length
            softmax_scale: Softmax scaling factor
            layer_id: Transformer layer index
            block_tables: Paged attention block tables (not used for XAttention)

        Returns:
            Attention output [total_q, num_heads, head_dim]
        """
        # When block_tables is provided (paged KV cache / prefix cache),
        # fallback to flash_attn as XAttention expects contiguous K, V
        if block_tables is not None:
            from flash_attn import flash_attn_varlen_func
            return flash_attn_varlen_func(
                q, k, v,
                cu_seqlens_q=cu_seqlens_q,
                cu_seqlens_k=cu_seqlens_k,
                max_seqlen_q=max_seqlen_q,
                max_seqlen_k=max_seqlen_k,
                softmax_scale=softmax_scale,
                causal=True,
                block_table=block_tables,
            )

        if not BSA_AVAILABLE:
            # Fallback to flash attention if BSA not available
            from flash_attn import flash_attn_varlen_func
            return flash_attn_varlen_func(
                q, k, v,
                cu_seqlens_q=cu_seqlens_q,
                cu_seqlens_k=cu_seqlens_k,
                max_seqlen_q=max_seqlen_q,
                max_seqlen_k=max_seqlen_k,
                softmax_scale=softmax_scale,
                causal=True,
            )

        if not XATTN_AVAILABLE:
            # Fallback to flash attention if xattn not available
            from flash_attn import flash_attn_varlen_func
            return flash_attn_varlen_func(
                q, k, v,
                cu_seqlens_q=cu_seqlens_q,
                cu_seqlens_k=cu_seqlens_k,
                max_seqlen_q=max_seqlen_q,
                max_seqlen_k=max_seqlen_k,
                softmax_scale=softmax_scale,
                causal=True,
            )

        from nanovllm.ops.xattn import xattn_estimate

        # Set DensityObserver mode on first layer
        if layer_id == 0:
            DensityObserver.set_mode("gpu_only")

        # Get dimensions
        total_q, num_heads, head_dim = q.shape
        total_kv, num_kv_heads, _ = k.shape

        # For now, assume batch_size = 1 (single sequence)
        # TODO: Support batched varlen format
        batch_size = cu_seqlens_q.shape[0] - 1
        if batch_size != 1:
            # Fallback to flash attention for batched input
            from flash_attn import flash_attn_varlen_func
            logger.warning(f"[XAttn] batch_size={batch_size} > 1, falling back to flash attention")
            return flash_attn_varlen_func(
                q, k, v,
                cu_seqlens_q=cu_seqlens_q,
                cu_seqlens_k=cu_seqlens_k,
                max_seqlen_q=max_seqlen_q,
                max_seqlen_k=max_seqlen_k,
                softmax_scale=softmax_scale,
                causal=True,
            )

        q_len = max_seqlen_q
        k_len = max_seqlen_k

        # Convert from varlen format [total, heads, dim] to [batch, heads, seq, dim]
        # q: [q_len, num_heads, head_dim] -> [1, num_heads, q_len, head_dim]
        Q = q.unsqueeze(0).transpose(1, 2)  # [1, num_heads, q_len, head_dim]
        K = k.unsqueeze(0).transpose(1, 2)  # [1, num_kv_heads, k_len, head_dim]
        V = v.unsqueeze(0).transpose(1, 2)  # [1, num_kv_heads, k_len, head_dim]

        # Expand KV for GQA - use pre-allocated buffers if available
        if num_heads != num_kv_heads:
            num_groups = num_heads // num_kv_heads
            if self._k_expanded is not None and k_len <= self._max_seq_len:
                # Use pre-allocated buffers with in-place expansion
                K_exp = self._k_expanded[:, :, :k_len, :]
                V_exp = self._v_expanded[:, :, :k_len, :]
                # In-place GQA expansion: [1, num_kv_heads, k_len, head_dim] -> [1, num_heads, k_len, head_dim]
                # Reshape K to [1, num_kv_heads, 1, k_len, head_dim] and broadcast to [1, num_kv_heads, num_groups, k_len, head_dim]
                K_exp.view(1, num_kv_heads, num_groups, k_len, head_dim).copy_(
                    K.unsqueeze(2).expand(-1, -1, num_groups, -1, -1)
                )
                V_exp.view(1, num_kv_heads, num_groups, k_len, head_dim).copy_(
                    V.unsqueeze(2).expand(-1, -1, num_groups, -1, -1)
                )
            else:
                # Fallback: dynamic allocation (when buffers not pre-allocated or seq too long)
                K_exp, V_exp = expand_kv_for_gqa(K, V, num_heads)
        else:
            K_exp, V_exp = K, V

        # Estimate block importance and get sparse mask
        with nvtx.range("xattn_estimate"):
            _, mask = xattn_estimate(
                Q, K_exp,
                chunk_size=self.chunk_size,
                block_size=self.BSA_BLOCK_SIZE,
                stride=self.stride,
                threshold=self.threshold,
                use_triton=self.use_triton,
                causal=True,
            )

        # Compute block counts
        q_block_num = (q_len + self.BSA_BLOCK_SIZE - 1) // self.BSA_BLOCK_SIZE
        k_block_num = (k_len + self.BSA_BLOCK_SIZE - 1) // self.BSA_BLOCK_SIZE

        # Prepare tensors for BSA
        # q, k, v need to be [seq_len, num_heads, head_dim]
        q_bsa = q  # Already [q_len, num_heads, head_dim]

        # For GQA with BSA, reuse the expanded K_exp, V_exp (convert to BSA format)
        # K_exp: [1, num_heads, k_len, head_dim] -> [k_len, num_heads, head_dim]
        if num_heads != num_kv_heads:
            k_bsa = K_exp.squeeze(0).transpose(0, 1)  # [k_len, num_heads, head_dim]
            v_bsa = V_exp.squeeze(0).transpose(0, 1)  # [k_len, num_heads, head_dim]
        else:
            k_bsa = k
            v_bsa = v

        # Prepare BSA inputs
        cu_seqlens_q_bsa = torch.tensor([0, q_len], dtype=torch.int32, device=q.device)
        cu_seqlens_k_bsa = torch.tensor([0, k_len], dtype=torch.int32, device=k.device)
        head_groups = torch.ones(num_heads, dtype=torch.int32, device=q.device)

        # Trim mask to actual block counts
        mask_trimmed = mask[:, :, :q_block_num, :k_block_num].contiguous()

        # Compute sparse attention using BSA
        with nvtx.range("xattn_bsa_compute"):
            output = block_sparse_attn_func(
                q_bsa, k_bsa, v_bsa,
                cu_seqlens_q_bsa,
                cu_seqlens_k_bsa,
                head_groups,
                None,  # key_padding_mask
                mask_trimmed,
                q_len, k_len,
                p_dropout=0.0,
                deterministic=True,
                is_causal=True,
            )

        # Record density for all layers via DensityObserver
        if layer_id == 0:
            # DEBUG: 打印 GPU-only Layer 0 的 mask 详情
            q_bk = mask_trimmed.shape[2]
            k_bk = mask_trimmed.shape[3]
            causal_total = q_bk * (q_bk + 1) // 2 * mask_trimmed.shape[0] * mask_trimmed.shape[1]
            causal_mask = torch.tril(torch.ones(q_bk, k_bk, device=mask_trimmed.device, dtype=torch.bool))
            selected = (mask_trimmed & causal_mask.unsqueeze(0).unsqueeze(0)).sum().item()
            logger.info(f"[DEBUG GPU-only Layer0] mask_shape={mask_trimmed.shape}, "
                       f"density={selected/causal_total:.6f}, selected={selected}, total={causal_total}")
        DensityObserver.record(layer_id, mask_trimmed, causal=True)

        return output

    def compute_decode(
        self,
        q: torch.Tensor,
        k_cache: torch.Tensor,
        v_cache: torch.Tensor,
        cache_seqlens: torch.Tensor,
        softmax_scale: float,
        layer_id: int,
        block_tables: torch.Tensor = None,
    ) -> torch.Tensor:
        """
        GPU-only decode attention - delegates to FullAttentionPolicy.

        XAttention is designed for long prefill sequences. For decode (single token),
        we use FullAttentionPolicy which calls flash_attn_with_kvcache.
        """
        from nanovllm.kvcache.sparse.full_policy import FullAttentionPolicy
        return FullAttentionPolicy().compute_decode(
            q, k_cache, v_cache, cache_seqlens, softmax_scale, layer_id, block_tables
        )

    # =========================================================================
    # Chunked offload methods
    # =========================================================================

    def select_blocks(
        self,
        available_blocks: List[int],
        offload_engine: "OffloadEngine",
        ctx: PolicyContext,
        q: torch.Tensor,
        k: torch.Tensor,
    ) -> List[int]:
        """
        Compute attention scores for all available blocks using flat_group_gemm,
        then use softmax_fuse_block_sum and find_blocks_chunked to select important blocks.

        This method aligns with GPU-only xattn_estimate_chunked:
        1. Loads each K block from CPU (historical blocks)
        2. Gets current chunk K from prefill buffer
        3. Concatenates [historical K, current chunk K] for correct softmax normalization
        4. Uses causal=True with correct chunk_start for position-aware masking
        5. Only selects from historical blocks (current chunk is always full attention)

        Args:
            available_blocks: List of CPU block IDs (historical blocks only)
            offload_engine: OffloadEngine for loading blocks
            ctx: PolicyContext with metadata
            q: Query tensor [seq_len, num_heads, head_dim] for current chunk
            k: Key tensor [seq_len, num_kv_heads, head_dim] for current chunk (used for estimation)

        Returns:
            Selected block IDs based on attention threshold
        """
        if q is None:
            return available_blocks

        from nanovllm.ops.xattn import flat_group_gemm_fuse_reshape, softmax_fuse_block_sum, find_blocks_chunked
        import math

        layer_id = ctx.layer_id
        # Use passed q parameter instead of ctx.query

        # Set DensityObserver mode on first layer
        if layer_id == 0:
            DensityObserver.set_mode("offload")

        # Convert Q to [batch, heads, seq_len, head_dim]
        # q: [seq_len, num_heads, head_dim] -> [1, num_heads, seq_len, head_dim]
        Q = q.unsqueeze(0).transpose(1, 2)  # [1, num_heads, seq_len, head_dim]

        num_heads = Q.shape[1]
        head_dim = Q.shape[3]
        q_len = Q.shape[2]

        # flat_group_gemm requires q_len to be divisible by stride * BLOCK_M (typically 8 * 128 = 1024)
        # Pad Q if necessary
        BLOCK_M = 128  # Triton block size
        alignment = self.stride * BLOCK_M
        if q_len < alignment:
            # Q too short, skip estimation and return all blocks
            logger.debug(f"[XAttn] select_blocks: q_len={q_len} < alignment={alignment}, skipping estimation")
            return available_blocks

        # Pad Q to alignment
        padded_q_len = ((q_len + alignment - 1) // alignment) * alignment
        if padded_q_len != q_len:
            pad_size = padded_q_len - q_len
            Q = torch.nn.functional.pad(Q, (0, 0, 0, pad_size), value=0)

        q_reshaped_len = padded_q_len // self.stride

        # Get block size from context
        block_size = ctx.block_size  # tokens per CPU block (e.g., 4096)
        reshaped_block_size = block_size // self.stride  # e.g., 4096/8 = 512

        # ============================================================
        # Step 1: Compute chunk_start and related parameters
        # ============================================================
        # chunk_start = Q's global position in reshaped space
        # Q starts at position: num_historical_blocks * block_size
        num_historical_blocks = len(available_blocks)
        historical_k_len = num_historical_blocks * block_size
        chunk_start = historical_k_len // self.stride  # Q's position in reshaped space
        chunk_end = chunk_start + q_reshaped_len

        # For valid Q length tracking (excluding padding)
        valid_q_reshaped = (q_len + self.stride - 1) // self.stride
        real_q_len = chunk_start + valid_q_reshaped

        # ============================================================
        # Step 2: Pipeline load historical K blocks and compute attn_scores
        # ============================================================
        # Key design: Load each block, compute immediately, then release
        # This avoids storing all K in GPU memory at once (offload-friendly)
        slot = 0
        attn_scores_list = []
        BLOCK_N = 128
        k_alignment = self.stride * BLOCK_N

        with nvtx.range("xattn_estimate_historical"):
            for cpu_block_id in available_blocks:
                # Load only K from CPU to GPU (V not needed for estimate)
                offload_engine.load_k_only_to_slot_layer(slot, layer_id, cpu_block_id, chunk_idx=cpu_block_id)
                offload_engine.wait_slot_layer(slot)

                # Get K only: [1, block_size, num_kv_heads, head_dim]
                k_block = offload_engine.get_k_for_slot(slot)

                # Convert K to [batch, heads, k_len, head_dim]
                K_chunk = k_block.transpose(1, 2)  # [1, num_kv_heads, block_size, head_dim]
                
                # Handle GQA: expand K heads to match Q heads
                num_kv_heads = K_chunk.shape[1]
                if num_heads != num_kv_heads:
                    num_groups = num_heads // num_kv_heads
                    K_chunk = K_chunk.repeat_interleave(num_groups, dim=1)
                                
                #> DEBUG: save all K cache
                start_pos = cpu_block_id * block_size
                self._debug_k_full[:, :, start_pos:start_pos + block_size, :].copy_(K_chunk)
                
                # # Pad K if necessary
                # k_len = K_chunk.shape[2]
                # if k_len < k_alignment:
                #     pad_size = k_alignment - k_len
                #     K_chunk = torch.nn.functional.pad(K_chunk, (0, 0, 0, pad_size), value=0)

                # # Compute attention scores for this historical block
                # # Historical blocks: all positions < Q, so Q always sees them (full attention)
                # # Use LOCAL chunk_start=0 to match test_xattn_k_chunked.py behavior
                # attn_chunk = flat_group_gemm_fuse_reshape(
                #     Q, K_chunk, self.stride,
                #     chunk_start=0,  # Local: same as test
                #     chunk_end=q_reshaped_len,
                #     is_causal=False,  # Historical K: all visible to Q
                # )
                # attn_scores_list.append(attn_chunk)

                # Mark slot as done for reuse
                offload_engine.record_slot_compute_done(slot)
        
        num_kv_heads = k.shape[1]
        if num_heads != num_kv_heads:
            num_groups = num_heads // num_kv_heads
        k_repeated = k.repeat_interleave(num_groups, dim=1).unsqueeze(0).transpose(1, 2)  # [1, num_heads, historical_k_len, head_dim]
        
        self._debug_k_full[:, :, historical_k_len:historical_k_len + q_len, :].copy_(k_repeated)

        # ============================================================
        # DEBUG: 累积 selected/total counts (仅 layer 0)
        # 使用完整 K 调用 xattn_estimate，与 GPU-only 逻辑一致
        # ============================================================
        if layer_id == 0:
            from nanovllm.ops.xattn import xattn_estimate

            total_k_len = historical_k_len + q_len
            K_full = self._debug_k_full[:, :, :total_k_len, :]

            # 用当前 Q chunk 和累积的 K 调用 xattn_estimate
            # 设置 chunk_size 为 q_len 的最小对齐值 (stride * BLOCK_M = 8 * 128 = 1024)
            alignment = self.stride * 128
            aligned_chunk_size = ((q_len + alignment - 1) // alignment) * alignment
            # DEBUG: 使用固定 threshold 测试
            _, mask_chunk = xattn_estimate(
                Q[:, :, :q_len, :],  # 当前 Q chunk
                K_full,              # 累积的 K
                block_size=self.BSA_BLOCK_SIZE,
                stride=self.stride,
                threshold=self.threshold,  # DEBUG: 使用传入的 threshold
                chunk_size=aligned_chunk_size,  # 对齐的 chunk_size
                causal=True,
            )

            # 计算有效的 block 数量（排除 padding）
            valid_q_blocks = (q_len + self.BSA_BLOCK_SIZE - 1) // self.BSA_BLOCK_SIZE
            valid_k_blocks = (total_k_len + self.BSA_BLOCK_SIZE - 1) // self.BSA_BLOCK_SIZE

            # 裁剪 mask 到有效区域
            mask_valid = mask_chunk[:, :, :valid_q_blocks, :valid_k_blocks]

            # 计算当前 chunk 的 selected/total (考虑 causal，考虑 Q 偏移量)
            q_blocks = valid_q_blocks
            k_blocks = valid_k_blocks
            # Q 从位置 (k_blocks - q_blocks) 开始，所以 Q block i 实际位置是 i + offset
            # Q block i (实际位置 i+offset) 可以看到 K block 0 到 i+offset
            q_offset_blocks = k_blocks - q_blocks
            indices = torch.arange(k_blocks, device=mask_valid.device).unsqueeze(0)  # [1, k_blocks]
            q_indices = torch.arange(q_blocks, device=mask_valid.device).unsqueeze(1)  # [q_blocks, 1]
            causal_mask = indices <= (q_indices + q_offset_blocks)  # [q_blocks, k_blocks]
            chunk_total = causal_mask.sum().item() * mask_valid.shape[0] * mask_valid.shape[1]
            chunk_selected = (mask_valid & causal_mask.unsqueeze(0).unsqueeze(0)).sum().item()

            # 累积
            self._debug_selected += chunk_selected
            self._debug_total += chunk_total

            # 打印当前累积的 density
            if self._debug_total > 0:
                density = self._debug_selected / self._debug_total
                logger.info(f"[DEBUG Offload Layer0] 累积 density: {density:.4f} "
                           f"(selected={self._debug_selected}, total={self._debug_total}, k_len={total_k_len}, "
                           f"mask_shape={mask_chunk.shape}, q_offset={q_offset_blocks})")

            # DEBUG: 跳过正常 offload 逻辑，直接返回所有 blocks
            return available_blocks
        else:
            # DEBUG: 非 Layer 0 也跳过正常 offload 逻辑
            return available_blocks

        # ============================================================
        # Step 3: Get current chunk K and compute its attn_scores
        # ============================================================
        with nvtx.range("xattn_estimate_current"):
            # Current chunk K is in prefill buffer (already on GPU)
            k_curr, _ = offload_engine.get_prefill_buffer_slice(layer_id, q_len)
            # k_curr: [1, q_len, num_kv_heads, head_dim] -> [1, num_kv_heads, q_len, head_dim]
            K_current = k_curr.transpose(1, 2)

            # Handle GQA for current chunk K
            num_kv_heads = K_current.shape[1]
            if num_heads != num_kv_heads:
                num_groups = num_heads // num_kv_heads
                K_current = K_current.repeat_interleave(num_groups, dim=1)

            # Pad current K if necessary
            curr_k_len = K_current.shape[2]
            padded_curr_k_len = ((curr_k_len + k_alignment - 1) // k_alignment) * k_alignment
            if padded_curr_k_len != curr_k_len:
                pad_size = padded_curr_k_len - curr_k_len
                K_current = torch.nn.functional.pad(K_current, (0, 0, 0, pad_size), value=0)

            # Compute attention scores for current chunk
            # IMPORTANT: Use LOCAL coordinates (0 to q_reshaped_len) for current chunk!
            # Because K_current only contains current chunk K (not full sequence),
            # block_n in kernel starts from 0. Using global chunk_start would cause
            # incorrect causal mask (Q would see K blocks it shouldn't).
            attn_current = flat_group_gemm_fuse_reshape(
                Q, K_current, self.stride,
                chunk_start=0,  # Local: Q starts at 0 relative to K_current
                chunk_end=q_reshaped_len,  # Local: Q ends at q_reshaped_len
                is_causal=True,  # Current chunk: apply causal mask
            )
            attn_scores_list.append(attn_current)
            del K_current

        # ============================================================
        # Step 4: Concatenate all attn_scores
        # ============================================================
        if not attn_scores_list:
            return available_blocks

        attn_scores = torch.cat(attn_scores_list, dim=-1)
        del attn_scores_list

        # Calculate padded K length for later use
        padded_k_len = historical_k_len + padded_curr_k_len

        # ============================================================
        # Step 5: Apply softmax_fuse_block_sum with causal=True
        # ============================================================
        cpu_block_size = block_size  # e.g., 4096
        bsa_per_cpu = cpu_block_size // self.BSA_BLOCK_SIZE  # e.g., 4096/128 = 32

        # Use BSA_BLOCK_SIZE for block aggregation (aligned with GPU-only)
        reshaped_bsa_bs = self.BSA_BLOCK_SIZE // self.stride  # e.g., 128/8 = 16
        norm = 1.0
        scale = 1.4426950408889634 / math.sqrt(head_dim) / self.stride / norm
        segment_size = min(4096, reshaped_bsa_bs)

        with nvtx.range("xattn_estimate_softmax"):
            block_sums = softmax_fuse_block_sum(
                attn_scores,
                reshaped_bsa_bs,
                segment_size,
                chunk_start=chunk_start,
                chunk_end=chunk_end,
                real_q_len=real_q_len,
                scale=scale,
                is_causal=True,  # Causal for consistent with GPU-only
            )
        # block_sums shape: [batch, heads, q_bsa_blocks, total_k_bsa_blocks]

        # ============================================================
        # Step 6: Use find_blocks_chunked to generate BSA-level mask
        # ============================================================
        # Calculate BSA block indices
        q_bsa_blocks = (padded_q_len + self.BSA_BLOCK_SIZE - 1) // self.BSA_BLOCK_SIZE
        total_k_bsa_blocks = (padded_k_len + self.BSA_BLOCK_SIZE - 1) // self.BSA_BLOCK_SIZE
        historical_k_bsa_blocks = num_historical_blocks * bsa_per_cpu

        # current_index for find_blocks_chunked: Q's block offset
        q_start_bsa_block = historical_k_bsa_blocks  # Q starts after historical K

        with nvtx.range("xattn_find_blocks"):
            # 对于历史 K 的选择，使用 causal=False 因为历史 K 都在当前 Q 之前
            # current_index=0 避免超出 block_sums 的 K 维度
            mask = find_blocks_chunked(
                block_sums,
                current_index=0,
                threshold=self.threshold,
                num_to_choose=None,
                decoding=False,
                mode="both",
                causal=False,
            )
        # mask shape: [batch, heads, q_bsa_blocks, total_k_bsa_blocks]

        # ============================================================
        # Step 7: Extract mask portions and record density
        # ============================================================
        B, H, Q_bsa, K_bsa_total = mask.shape

        # Calculate valid Q blocks (excluding padding)
        valid_q_bsa = (q_len + self.BSA_BLOCK_SIZE - 1) // self.BSA_BLOCK_SIZE
        valid_curr_k_bsa = (curr_k_len + self.BSA_BLOCK_SIZE - 1) // self.BSA_BLOCK_SIZE

        # 7a: Record historical blocks density (暂时禁用，使用 DEBUG 输出代替)
        # if historical_k_bsa_blocks > 0:
        #     ... DensityObserver.record_counts ...

        # 7b: Record current chunk density (暂时禁用)
        # if valid_curr_k_bsa > 0:
        #     ... DensityObserver.record_counts ...

        # Step 7.5: Save historical mask to pre-allocated buffer for compute_chunked_prefill
        # Use full Q_bsa (padded) for buffer, not valid_q_bsa
        mask_historical_full = mask[:, :, :, :historical_k_bsa_blocks]
        if self._prefill_mask_buffer is not None:
            # Only save historical portion of mask
            self._prefill_mask_buffer[:, :, :Q_bsa, :historical_k_bsa_blocks].copy_(mask_historical_full)
            self._current_mask_q_bsa = Q_bsa
            self._current_mask_k_bsa = historical_k_bsa_blocks

        # ============================================================
        # Step 8: Aggregate mask to CPU block level (union of heads)
        # ============================================================
        # Only aggregate historical blocks (current chunk is always full attention)
        num_cpu_blocks = num_historical_blocks

        with nvtx.range("xattn_aggregate_mask"):
            # Reshape historical mask: [B, H, Q_bsa, historical_k_bsa] -> [B, H, Q_bsa, num_cpu, bsa_per_cpu]
            # Use full Q_bsa (not valid_q_bsa) for aggregation
            mask_per_cpu = mask_historical_full.view(B, H, Q_bsa, num_cpu_blocks, bsa_per_cpu)

            # Union across: bsa_per_cpu, Q_bsa, heads -> [B, num_cpu]
            cpu_needed = mask_per_cpu.any(dim=-1).any(dim=2).any(dim=1)  # [B, num_cpu]

            # Get selected indices
            selected_indices = cpu_needed[0].nonzero().squeeze(-1).tolist()
            if isinstance(selected_indices, int):
                selected_indices = [selected_indices]

        # Handle empty available_blocks case (first chunk)
        if available_blocks:
            selected_block_ids = [available_blocks[i] for i in selected_indices]
        else:
            selected_block_ids = []

        # Always include first block (sink) and last block for safety
        if available_blocks and available_blocks[0] not in selected_block_ids:
            selected_block_ids.insert(0, available_blocks[0])
        if available_blocks and available_blocks[-1] not in selected_block_ids:
            selected_block_ids.append(available_blocks[-1])

        # Record communication density (CPU block granularity) - only if there are historical blocks
        if available_blocks:
            DensityObserver.record_comm_density(
                layer_id,
                selected_cpu_blocks=len(selected_block_ids),
                total_cpu_blocks=len(available_blocks),
            )

        # Update statistics (only for layer 0 to avoid overcounting)
        if layer_id == 0 and available_blocks:
            self._stats_total_available_blocks += len(available_blocks)
            self._stats_total_selected_blocks += len(selected_block_ids)
            self._stats_num_chunks += 1

            # Log per-chunk density
            chunk_density = len(selected_block_ids) / len(available_blocks)
            logger.debug(f"[XAttn] chunk={ctx.query_chunk_idx}, available={len(available_blocks)}, "
                        f"selected={len(selected_block_ids)}, chunk_density={chunk_density:.1%}")

        # Free intermediate tensors to prevent memory leak
        del attn_scores, block_sums, mask, mask_historical_full

        return selected_block_ids

    def compute_chunked_prefill(
        self,
        q: torch.Tensor,
        k: torch.Tensor,
        v: torch.Tensor,
        layer_id: int,
        softmax_scale: float,
        offload_engine: "OffloadEngine",
        kvcache_manager: "KVCacheManager",
        current_chunk_idx: int,
        seq: "Sequence",
        num_tokens: int,
        selected_blocks: List[int],
    ) -> torch.Tensor:
        """
        Compute attention for chunked prefill using XAttention sparse block selection.

        This method handles the chunked prefill computation:
        1. Load and compute attention to historical chunks (using selected_blocks)
        2. Compute attention to current chunk
        3. Merge all results

        Note: The BSA-level mask is saved in self._prefill_mask_buffer by select_blocks().
        Currently we use flash_attn_with_lse for computation (supports LSE merge).
        TODO: Optimize to use BSA kernel with the saved mask for per-head sparse attention.

        Args:
            q: Query tensor [seq_len, num_heads, head_dim]
            k: Key tensor [seq_len, num_kv_heads, head_dim] (unused, from prefill buffer)
            v: Value tensor [seq_len, num_kv_heads, head_dim] (unused, from prefill buffer)
            layer_id: Current layer index
            softmax_scale: Softmax scaling factor
            offload_engine: OffloadEngine for loading blocks
            kvcache_manager: KVCacheManager for block management
            current_chunk_idx: Current chunk index
            seq: Sequence object
            num_tokens: Number of tokens in current chunk
            selected_blocks: List of CPU block IDs selected by select_blocks

        Returns:
            Attention output [seq_len, num_heads, head_dim]
        """
        # Use FlashInfer-based implementations (more optimized)
        from nanovllm.ops.chunked_attention import (
            flash_attn_with_lse_flashinfer as flash_attn_with_lse,
            merge_attention_outputs_flashinfer as merge_attention_outputs,
        )

        q_batched = q.unsqueeze(0)  # [1, seq_len, num_heads, head_dim]
        o_acc = None
        lse_acc = None
        compute_stream = offload_engine.compute_stream

        # Use the pre-selected blocks directly
        cpu_block_table = selected_blocks

        # Note: BSA mask is available in self._prefill_mask_buffer (saved by select_blocks)
        # Mask shape: [1, num_heads, Q_bsa, K_bsa] where Q_bsa = self._current_mask_q_bsa
        # Selected indices: self._selected_cpu_indices, bsa_per_cpu: self._bsa_per_cpu
        # TODO: Use this mask with BSA kernel for per-head sparse attention optimization

        if cpu_block_table:
            with nvtx.range("xattn_compute_historical"):
                load_slots = list(range(offload_engine.num_ring_slots))
                num_blocks = len(cpu_block_table)

                if len(load_slots) == 1:
                    # Only 1 slot - use synchronous mode
                    slot = load_slots[0]
                    for block_idx in range(num_blocks):
                        cpu_block_id = cpu_block_table[block_idx]
                        offload_engine.load_to_slot_layer(slot, layer_id, cpu_block_id, chunk_idx=cpu_block_id)
                        offload_engine.wait_slot_layer(slot)

                        with torch.cuda.stream(compute_stream):
                            prev_k, prev_v = offload_engine.get_kv_for_slot(slot)
                            prev_o, prev_lse = flash_attn_with_lse(
                                q_batched, prev_k, prev_v,
                                softmax_scale=softmax_scale,
                                causal=False,
                            )
                            if o_acc is None:
                                o_acc, lse_acc = prev_o, prev_lse
                            else:
                                o_acc, lse_acc = merge_attention_outputs(o_acc, lse_acc, prev_o, prev_lse)
                            offload_engine.record_slot_compute_done(slot)
                else:
                    # Multiple slots - use pipeline
                    num_slots = len(load_slots)
                    num_preload = min(num_slots, num_blocks)
                    for i in range(num_preload):
                        cpu_block_id = cpu_block_table[i]
                        offload_engine.load_to_slot_layer(load_slots[i], layer_id, cpu_block_id, chunk_idx=cpu_block_id)

                    for block_idx in range(num_blocks):
                        current_slot = load_slots[block_idx % num_slots]

                        offload_engine.wait_slot_layer(current_slot)

                        with torch.cuda.stream(compute_stream):
                            prev_k, prev_v = offload_engine.get_kv_for_slot(current_slot)
                            prev_o, prev_lse = flash_attn_with_lse(
                                q_batched, prev_k, prev_v,
                                softmax_scale=softmax_scale,
                                causal=False,
                            )
                            offload_engine.record_slot_compute_done(current_slot)

                            if o_acc is None:
                                o_acc, lse_acc = prev_o, prev_lse
                            else:
                                o_acc, lse_acc = merge_attention_outputs(o_acc, lse_acc, prev_o, prev_lse)

                        # Issue next transfer
                        next_block_idx = block_idx + num_slots
                        if next_block_idx < num_blocks:
                            next_slot = load_slots[next_block_idx % num_slots]
                            next_cpu_block_id = cpu_block_table[next_block_idx]
                            offload_engine.load_to_slot_layer(next_slot, layer_id, next_cpu_block_id, chunk_idx=next_cpu_block_id)

        # Compute attention to current chunk (causal mask)
        with nvtx.range("xattn_compute_current"):
            with torch.cuda.stream(compute_stream):
                k_curr, v_curr = offload_engine.get_prefill_buffer_slice(layer_id, num_tokens)
                current_o, current_lse = flash_attn_with_lse(
                    q_batched, k_curr, v_curr,
                    softmax_scale=softmax_scale,
                    causal=True,
                )

        # Merge historical and current attention
        with nvtx.range("xattn_compute_merge"):
            with torch.cuda.stream(compute_stream):
                if o_acc is None:
                    final_o = current_o
                else:
                    final_o, _ = merge_attention_outputs(o_acc, lse_acc, current_o, current_lse)

        # Sync default stream with compute_stream before returning
        torch.cuda.default_stream().wait_stream(compute_stream)

        # Remove batch dimension: [1, seq_len, num_heads, head_dim] -> [seq_len, num_heads, head_dim]
        return final_o.squeeze(0)

    def compute_chunked_decode(
        self,
        q: torch.Tensor,
        layer_id: int,
        softmax_scale: float,
        offload_engine: "OffloadEngine",
        kvcache_manager: "KVCacheManager",
        seq: "Sequence",
        selected_blocks: List[int],
    ) -> torch.Tensor:
        """
        XAttention does not support decode phase.
        """
        raise NotImplementedError(
            "XAttentionBSAPolicy does not support decode phase. "
            "Use FullAttentionPolicy for decode."
        )

    def reset(self) -> None:
        """Reset policy state and clear sparse metadata."""
        self.sparse_metadata.clear()
        # Don't reset statistics here - they accumulate across the entire prefill

    def reset_stats(self) -> None:
        """Reset density statistics."""
        self._stats_total_available_blocks = 0
        self._stats_total_selected_blocks = 0
        self._stats_num_chunks = 0

    def get_density_stats(self) -> dict:
        """Get density statistics."""
        if self._stats_total_available_blocks == 0:
            return {
                "total_available_blocks": 0,
                "total_selected_blocks": 0,
                "num_chunks": 0,
                "overall_density": 0.0,
            }
        return {
            "total_available_blocks": self._stats_total_available_blocks,
            "total_selected_blocks": self._stats_total_selected_blocks,
            "num_chunks": self._stats_num_chunks,
            "overall_density": self._stats_total_selected_blocks / self._stats_total_available_blocks,
        }

    def print_density_stats(self) -> None:
        """Print density statistics summary."""
        stats = self.get_density_stats()
        logger.info(f"[XAttn BSA] Density Stats: chunks={stats['num_chunks']}, "
                   f"available={stats['total_available_blocks']}, "
                   f"selected={stats['total_selected_blocks']}, "
                   f"density={stats['overall_density']:.1%}")

    def __repr__(self) -> str:
        return f"XAttentionBSAPolicy(threshold={self.threshold}, stride={self.stride})"