sglang 0.5.2rc2__py3-none-any.whl → 0.5.3rc0__py3-none-any.whl

sglang/srt/layers/attention/fla/chunk_delta_h.py

@@ -0,0 +1,314 @@
+# Adapted from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/common/chunk_delta_h.py
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from typing import Optional, Tuple
+
+import torch
+import triton
+import triton.language as tl
+
+from sglang.srt.layers.attention.fla.index import (
+    prepare_chunk_indices,
+    prepare_chunk_offsets,
+)
+from sglang.srt.layers.attention.fla.op import exp, safe_exp
+from sglang.srt.layers.attention.fla.utils import is_nvidia_hopper
+
+NUM_WARPS = [2, 4] if is_nvidia_hopper else [2, 4, 8, 16]
+
+
+@triton.heuristics(
+    {
+        "USE_G": lambda args: args["g"] is not None,
+        "USE_INITIAL_STATE": lambda args: args["h0"] is not None,
+        "STORE_FINAL_STATE": lambda args: args["ht"] is not None,
+        "SAVE_NEW_VALUE": lambda args: args["v_new"] is not None,
+        "IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
+    }
+)
+# @triton.autotune(
+#     configs=[
+#         triton.Config({"BV": BV}, num_warps=num_warps, num_stages=num_stages)
+#         for num_warps in [2, 4]
+#         for num_stages in [2, 3, 4]
+#         for BV in [32, 64]
+#     ],
+#     key=["H", "K", "V", "BT", "USE_G"],
+#     use_cuda_graph=use_cuda_graph,
+# )
+@triton.jit(do_not_specialize=["T"])
+def chunk_gated_delta_rule_fwd_kernel_h_blockdim64(
+    k,
+    v,
+    w,
+    v_new,
+    g,
+    h,
+    h0,
+    ht,
+    cu_seqlens,
+    chunk_offsets,
+    T,
+    H: tl.constexpr,
+    Hg: tl.constexpr,
+    K: tl.constexpr,
+    V: tl.constexpr,
+    BT: tl.constexpr,
+    BV: tl.constexpr,
+    USE_G: tl.constexpr,
+    USE_INITIAL_STATE: tl.constexpr,
+    STORE_FINAL_STATE: tl.constexpr,
+    SAVE_NEW_VALUE: tl.constexpr,
+    IS_VARLEN: tl.constexpr,
+):
+    i_v, i_nh = tl.program_id(0), tl.program_id(1)
+    i_n, i_h = i_nh // H, i_nh % H
+    if IS_VARLEN:
+        bos, eos = tl.load(cu_seqlens + i_n).to(tl.int32), tl.load(
+            cu_seqlens + i_n + 1
+        ).to(tl.int32)
+        T = eos - bos
+        NT = tl.cdiv(T, BT)
+        boh = tl.load(chunk_offsets + i_n).to(tl.int32)
+    else:
+        bos, eos = i_n * T, i_n * T + T
+        NT = tl.cdiv(T, BT)
+        boh = i_n * NT
+
+    # [BK, BV]
+    b_h1 = tl.zeros([64, BV], dtype=tl.float32)
+    if K > 64:
+        b_h2 = tl.zeros([64, BV], dtype=tl.float32)
+    if K > 128:
+        b_h3 = tl.zeros([64, BV], dtype=tl.float32)
+    if K > 192:
+        b_h4 = tl.zeros([64, BV], dtype=tl.float32)
+
+    # calculate offset
+    h += (boh * H + i_h) * K * V
+    v += (bos * H + i_h) * V
+    k += (bos * Hg + i_h // (H // Hg)) * K
+    w += (bos * H + i_h) * K
+    if SAVE_NEW_VALUE:
+        v_new += (bos * H + i_h) * V
+    stride_v = H * V
+    stride_h = H * K * V
+    stride_k = Hg * K
+    stride_w = H * K
+    if USE_INITIAL_STATE:
+        h0 = h0 + i_nh * K * V
+    if STORE_FINAL_STATE:
+        ht = ht + i_nh * K * V
+
+    # load initial state
+    if USE_INITIAL_STATE:
+        p_h0_1 = tl.make_block_ptr(h0, (K, V), (V, 1), (0, i_v * BV), (64, BV), (1, 0))
+        b_h1 += tl.load(p_h0_1, boundary_check=(0, 1)).to(tl.float32)
+        if K > 64:
+            p_h0_2 = tl.make_block_ptr(
+                h0, (K, V), (V, 1), (64, i_v * BV), (64, BV), (1, 0)
+            )
+            b_h2 += tl.load(p_h0_2, boundary_check=(0, 1)).to(tl.float32)
+        if K > 128:
+            p_h0_3 = tl.make_block_ptr(
+                h0, (K, V), (V, 1), (128, i_v * BV), (64, BV), (1, 0)
+            )
+            b_h3 += tl.load(p_h0_3, boundary_check=(0, 1)).to(tl.float32)
+        if K > 192:
+            p_h0_4 = tl.make_block_ptr(
+                h0, (K, V), (V, 1), (192, i_v * BV), (64, BV), (1, 0)
+            )
+            b_h4 += tl.load(p_h0_4, boundary_check=(0, 1)).to(tl.float32)
+
+    # main recurrence
+    for i_t in range(NT):
+        p_h1 = tl.make_block_ptr(
+            h + i_t * stride_h, (K, V), (V, 1), (0, i_v * BV), (64, BV), (1, 0)
+        )
+        tl.store(p_h1, b_h1.to(p_h1.dtype.element_ty), boundary_check=(0, 1))
+        if K > 64:
+            p_h2 = tl.make_block_ptr(
+                h + i_t * stride_h, (K, V), (V, 1), (64, i_v * BV), (64, BV), (1, 0)
+            )
+            tl.store(p_h2, b_h2.to(p_h2.dtype.element_ty), boundary_check=(0, 1))
+        if K > 128:
+            p_h3 = tl.make_block_ptr(
+                h + i_t * stride_h, (K, V), (V, 1), (128, i_v * BV), (64, BV), (1, 0)
+            )
+            tl.store(p_h3, b_h3.to(p_h3.dtype.element_ty), boundary_check=(0, 1))
+        if K > 192:
+            p_h4 = tl.make_block_ptr(
+                h + i_t * stride_h, (K, V), (V, 1), (192, i_v * BV), (64, BV), (1, 0)
+            )
+            tl.store(p_h4, b_h4.to(p_h4.dtype.element_ty), boundary_check=(0, 1))
+
+        p_v = tl.make_block_ptr(
+            v, (T, V), (stride_v, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
+        )
+        p_v_new = (
+            tl.make_block_ptr(
+                v_new, (T, V), (stride_v, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
+            )
+            if SAVE_NEW_VALUE
+            else None
+        )
+        b_v_new = tl.zeros([BT, BV], dtype=tl.float32)
+        p_w = tl.make_block_ptr(
+            w, (T, K), (stride_w, 1), (i_t * BT, 0), (BT, 64), (1, 0)
+        )
+        b_w = tl.load(p_w, boundary_check=(0, 1))
+        b_v_new += tl.dot(b_w, b_h1.to(b_w.dtype))
+        if K > 64:
+            p_w = tl.make_block_ptr(
+                w, (T, K), (stride_w, 1), (i_t * BT, 64), (BT, 64), (1, 0)
+            )
+            b_w = tl.load(p_w, boundary_check=(0, 1))
+            b_v_new += tl.dot(b_w, b_h2.to(b_w.dtype))
+        if K > 128:
+            p_w = tl.make_block_ptr(
+                w, (T, K), (stride_w, 1), (i_t * BT, 128), (BT, 64), (1, 0)
+            )
+            b_w = tl.load(p_w, boundary_check=(0, 1))
+            b_v_new += tl.dot(b_w, b_h3.to(b_w.dtype))
+        if K > 192:
+            p_w = tl.make_block_ptr(
+                w, (T, K), (stride_w, 1), (i_t * BT, 192), (BT, 64), (1, 0)
+            )
+            b_w = tl.load(p_w, boundary_check=(0, 1))
+            b_v_new += tl.dot(b_w, b_h4.to(b_w.dtype))
+        b_v_new = -b_v_new + tl.load(p_v, boundary_check=(0, 1))
+
+        if SAVE_NEW_VALUE:
+            p_v_new = tl.make_block_ptr(
+                v_new, (T, V), (stride_v, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
+            )
+            tl.store(
+                p_v_new, b_v_new.to(p_v_new.dtype.element_ty), boundary_check=(0, 1)
+            )
+
+        if USE_G:
+            last_idx = min((i_t + 1) * BT, T) - 1
+            b_g_last = tl.load(g + bos * H + last_idx * H + i_h)
+            p_g = tl.make_block_ptr(
+                g + bos * H + i_h, (T,), (H,), (i_t * BT,), (BT,), (0,)
+            )
+            b_g = tl.load(p_g, boundary_check=(0,))
+            b_v_new = b_v_new * safe_exp(b_g_last - b_g)[:, None]
+            b_g_last = exp(b_g_last)
+            b_h1 = b_h1 * b_g_last
+            if K > 64:
+                b_h2 = b_h2 * b_g_last
+            if K > 128:
+                b_h3 = b_h3 * b_g_last
+            if K > 192:
+                b_h4 = b_h4 * b_g_last
+        b_v_new = b_v_new.to(k.dtype.element_ty)
+        p_k = tl.make_block_ptr(
+            k, (K, T), (1, stride_k), (0, i_t * BT), (64, BT), (0, 1)
+        )
+        b_k = tl.load(p_k, boundary_check=(0, 1))
+        b_h1 += tl.dot(b_k, b_v_new)
+        if K > 64:
+            p_k = tl.make_block_ptr(
+                k, (K, T), (1, stride_k), (64, i_t * BT), (64, BT), (0, 1)
+            )
+            b_k = tl.load(p_k, boundary_check=(0, 1))
+            b_h2 += tl.dot(b_k, b_v_new)
+        if K > 128:
+            p_k = tl.make_block_ptr(
+                k, (K, T), (1, stride_k), (128, i_t * BT), (64, BT), (0, 1)
+            )
+            b_k = tl.load(p_k, boundary_check=(0, 1))
+            b_h3 += tl.dot(b_k, b_v_new)
+        if K > 192:
+            p_k = tl.make_block_ptr(
+                k, (K, T), (1, stride_k), (192, i_t * BT), (64, BT), (0, 1)
+            )
+            b_k = tl.load(p_k, boundary_check=(0, 1))
+            b_h4 += tl.dot(b_k, b_v_new)
+
+    # epilogue
+    if STORE_FINAL_STATE:
+        p_ht = tl.make_block_ptr(ht, (K, V), (V, 1), (0, i_v * BV), (64, BV), (1, 0))
+        tl.store(p_ht, b_h1.to(p_ht.dtype.element_ty), boundary_check=(0, 1))
+        if K > 64:
+            p_ht = tl.make_block_ptr(
+                ht, (K, V), (V, 1), (64, i_v * BV), (64, BV), (1, 0)
+            )
+            tl.store(p_ht, b_h2.to(p_ht.dtype.element_ty), boundary_check=(0, 1))
+        if K > 128:
+            p_ht = tl.make_block_ptr(
+                ht, (K, V), (V, 1), (128, i_v * BV), (64, BV), (1, 0)
+            )
+            tl.store(p_ht, b_h3.to(p_ht.dtype.element_ty), boundary_check=(0, 1))
+        if K > 192:
+            p_ht = tl.make_block_ptr(
+                ht, (K, V), (V, 1), (192, i_v * BV), (64, BV), (1, 0)
+            )
+            tl.store(p_ht, b_h4.to(p_ht.dtype.element_ty), boundary_check=(0, 1))
+
+
+def chunk_gated_delta_rule_fwd_h(
+    k: torch.Tensor,
+    w: torch.Tensor,
+    u: torch.Tensor,
+    g: Optional[torch.Tensor] = None,
+    initial_state: Optional[torch.Tensor] = None,
+    output_final_state: bool = False,
+    chunk_size: int = 64,  # SY: remove this argument and force chunk size 64?
+    save_new_value: bool = True,
+    cu_seqlens: Optional[torch.LongTensor] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    B, T, Hg, K, V = *k.shape, u.shape[-1]
+    H = u.shape[-2]
+    BT = chunk_size
+
+    chunk_indices = (
+        prepare_chunk_indices(cu_seqlens, chunk_size)
+        if cu_seqlens is not None
+        else None
+    )
+    # N: the actual number of sequences in the batch with either equal or variable lengths
+    if cu_seqlens is None:
+        N, NT, chunk_offsets = B, triton.cdiv(T, BT), None
+    else:
+        N, NT, chunk_offsets = (
+            len(cu_seqlens) - 1,
+            len(chunk_indices),
+            prepare_chunk_offsets(cu_seqlens, BT),
+        )
+    assert K <= 256, "current kernel does not support head dimension larger than 256."
+
+    h = k.new_empty(B, NT, H, K, V)
+    final_state = (
+        k.new_empty(N, H, K, V, dtype=torch.float32) if output_final_state else None
+    )
+
+    v_new = torch.empty_like(u) if save_new_value else None
+
+    def grid(meta):
+        return (triton.cdiv(V, meta["BV"]), N * H)
+
+    chunk_gated_delta_rule_fwd_kernel_h_blockdim64[grid](
+        k=k,
+        v=u,
+        w=w,
+        v_new=v_new,
+        g=g,
+        h=h,
+        h0=initial_state,
+        ht=final_state,
+        cu_seqlens=cu_seqlens,
+        chunk_offsets=chunk_offsets,
+        T=T,
+        H=H,
+        Hg=Hg,
+        K=K,
+        V=V,
+        BT=BT,
+        BV=32,
+        num_warps=4,
+        num_stages=2,
+    )
+    return h, v_new, final_state

sglang/srt/layers/attention/fla/chunk_o.py

@@ -0,0 +1,178 @@
+# Adapted from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/common/chunk_o.py
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from typing import Optional, Tuple
+
+import torch
+import triton
+import triton.language as tl
+
+from sglang.srt.layers.attention.fla.index import prepare_chunk_indices
+from sglang.srt.layers.attention.fla.op import exp, safe_exp
+from sglang.srt.layers.attention.fla.utils import check_shared_mem, is_nvidia_hopper
+
+BKV_LIST = [64, 128] if check_shared_mem() else [32, 64]
+NUM_WARPS = [2, 4] if is_nvidia_hopper else [2, 4, 8]
+
+
+@triton.heuristics(
+    {
+        "USE_G": lambda args: args["g"] is not None,
+        "IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
+    }
+)
+# @triton.autotune(
+#     configs=[
+#         triton.Config({"BK": BK, "BV": BV}, num_warps=num_warps, num_stages=num_stages)
+#         for BK in BKV_LIST
+#         for BV in BKV_LIST
+#         for num_warps in NUM_WARPS
+#         for num_stages in [2, 3, 4]
+#     ],
+#     key=["H", "K", "V", "BT"],
+# )
+@triton.jit(do_not_specialize=["T"])
+def chunk_fwd_kernel_o(
+    q,
+    k,
+    v,
+    h,
+    g,
+    o,
+    cu_seqlens,
+    chunk_indices,
+    scale,
+    T,
+    H: tl.constexpr,
+    Hg: tl.constexpr,
+    K: tl.constexpr,
+    V: tl.constexpr,
+    BT: tl.constexpr,
+    BK: tl.constexpr,
+    BV: tl.constexpr,
+    USE_G: tl.constexpr,
+    IS_VARLEN: tl.constexpr,
+):
+    i_v, i_t, i_bh = tl.program_id(0), tl.program_id(1), tl.program_id(2)
+    i_b, i_h = i_bh // H, i_bh % H
+
+    if IS_VARLEN:
+        i_tg = i_t
+        i_n, i_t = tl.load(chunk_indices + i_t * 2).to(tl.int32), tl.load(
+            chunk_indices + i_t * 2 + 1
+        ).to(tl.int32)
+        bos, eos = tl.load(cu_seqlens + i_n).to(tl.int32), tl.load(
+            cu_seqlens + i_n + 1
+        ).to(tl.int32)
+        T = eos - bos
+        NT = tl.cdiv(T, BT)
+    else:
+        NT = tl.cdiv(T, BT)
+        i_tg = i_b * NT + i_t
+        bos, eos = i_b * T, i_b * T + T
+
+    # offset calculation
+    q += (bos * Hg + i_h // (H // Hg)) * K
+    k += (bos * Hg + i_h // (H // Hg)) * K
+    v += (bos * H + i_h) * V
+    o += (bos * H + i_h) * V
+    h += (i_tg * H + i_h).to(tl.int64) * K * V
+
+    b_o = tl.zeros([BT, BV], dtype=tl.float32)
+    b_A = tl.zeros([BT, BT], dtype=tl.float32)
+
+    for i_k in range(tl.cdiv(K, BK)):
+        p_q = tl.make_block_ptr(
+            q, (T, K), (Hg * K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0)
+        )
+        p_k = tl.make_block_ptr(
+            k, (K, T), (1, Hg * K), (i_k * BK, i_t * BT), (BK, BT), (0, 1)
+        )
+        p_h = tl.make_block_ptr(
+            h, (K, V), (V, 1), (i_k * BK, i_v * BV), (BK, BV), (1, 0)
+        )
+        # [BT, BK]
+        b_q = tl.load(p_q, boundary_check=(0, 1))
+        # [BK, BT]
+        b_k = tl.load(p_k, boundary_check=(0, 1))
+        # [BK, BV]
+        b_h = tl.load(p_h, boundary_check=(0, 1))
+
+        # [BT, BK] @ [BK, BV] -> [BT, BV]
+        b_o += tl.dot(b_q, b_h)
+        # [BT, BK] @ [BK, BT] -> [BT, BT]
+        b_A += tl.dot(b_q, b_k)
+
+    if USE_G:
+        g += bos * H + i_h
+        p_g = tl.make_block_ptr(g, (T,), (H,), (i_t * BT,), (BT,), (0,))
+        b_g = tl.load(p_g, boundary_check=(0,))
+        b_o = b_o * exp(b_g)[:, None]
+        b_A = b_A * safe_exp(b_g[:, None] - b_g[None, :])
+
+    o_i = tl.arange(0, BT)
+    m_A = o_i[:, None] >= o_i[None, :]
+    b_A = tl.where(m_A, b_A, 0)
+
+    p_v = tl.make_block_ptr(
+        v, (T, V), (H * V, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
+    )
+    p_o = tl.make_block_ptr(
+        o, (T, V), (H * V, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
+    )
+    b_v = tl.load(p_v, boundary_check=(0, 1))
+
+    # to fix mma -> mma layout conversion
+    # already solved by triton v3.2 or higher
+    b_o = b_o * scale + tl.dot(b_A.to(b_v.dtype), b_v) * scale
+    tl.store(p_o, b_o.to(p_o.dtype.element_ty), boundary_check=(0, 1))
+
+
+def chunk_fwd_o(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    h: torch.Tensor,
+    g: Optional[torch.Tensor] = None,  # cumsum of log decay
+    scale: Optional[float] = None,
+    cu_seqlens: Optional[torch.LongTensor] = None,
+    chunk_size: int = 64,
+) -> torch.Tensor:
+    B, T, Hg, K, V = *q.shape, v.shape[-1]
+    H = v.shape[-2]
+    BT = min(chunk_size, max(16, triton.next_power_of_2(T)))
+    chunk_indices = (
+        prepare_chunk_indices(cu_seqlens, BT) if cu_seqlens is not None else None
+    )
+    NT = triton.cdiv(T, BT) if cu_seqlens is None else len(chunk_indices)
+    if scale is None:
+        scale = k.shape[-1] ** -0.5
+
+    o = torch.empty_like(v)
+
+    def grid(meta):
+        return (triton.cdiv(V, meta["BV"]), NT, B * H)
+
+    chunk_fwd_kernel_o[grid](
+        q,
+        k,
+        v,
+        h,
+        g,
+        o,
+        cu_seqlens,
+        chunk_indices,
+        scale,
+        T=T,
+        H=H,
+        Hg=Hg,
+        K=K,
+        V=V,
+        BT=BT,
+        BK=128,
+        BV=64,
+        num_warps=4,
+        num_stages=2,
+    )
+    return o

sglang/srt/layers/attention/fla/chunk_scaled_dot_kkt.py

@@ -0,0 +1,151 @@
+# Adapted from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/common/chunk_scaled_dot_kkt.py
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from typing import Optional
+
+import torch
+import triton
+import triton.language as tl
+
+from sglang.srt.layers.attention.fla.index import prepare_chunk_indices
+from sglang.srt.layers.attention.fla.op import safe_exp
+
+
+@triton.heuristics(
+    {
+        "IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
+        "USE_G": lambda args: args["g_cumsum"] is not None,
+    }
+)
+# @triton.autotune(
+#     configs=[
+#         triton.Config({"BK": BK}, num_warps=num_warps, num_stages=num_stages)
+#         for BK in [32, 64, 128]
+#         for num_warps in [2, 4, 8]
+#         for num_stages in [2, 3, 4]
+#     ],
+#     key=["H", "K", "BT", "IS_VARLEN"],
+# )
+@triton.jit(do_not_specialize=["T"])
+def chunk_scaled_dot_kkt_fwd_kernel(
+    k,
+    beta,
+    g_cumsum,
+    A,
+    cu_seqlens,
+    chunk_indices,
+    T,
+    H: tl.constexpr,
+    Hg: tl.constexpr,
+    K: tl.constexpr,
+    BT: tl.constexpr,
+    BK: tl.constexpr,
+    IS_VARLEN: tl.constexpr,
+    USE_G: tl.constexpr,
+):
+    i_t, i_bh = tl.program_id(0), tl.program_id(1)
+    i_b, i_h = i_bh // H, i_bh % H
+    if IS_VARLEN:
+        i_n, i_t = tl.load(chunk_indices + i_t * 2).to(tl.int32), tl.load(
+            chunk_indices + i_t * 2 + 1
+        ).to(tl.int32)
+        bos, eos = tl.load(cu_seqlens + i_n).to(tl.int32), tl.load(
+            cu_seqlens + i_n + 1
+        ).to(tl.int32)
+        T = eos - bos
+    else:
+        bos, eos = i_b * T, i_b * T + T
+    o_t = tl.arange(0, BT)
+
+    p_beta = tl.make_block_ptr(
+        beta + bos * H + i_h, (T,), (H,), (i_t * BT,), (BT,), (0,)
+    )
+    b_beta = tl.load(p_beta, boundary_check=(0,))
+
+    b_A = tl.zeros([BT, BT], dtype=tl.float32)
+    for i_k in range(tl.cdiv(K, BK)):
+        p_k = tl.make_block_ptr(
+            k + (bos * Hg + i_h // (H // Hg)) * K,
+            (T, K),
+            (Hg * K, 1),
+            (i_t * BT, i_k * BK),
+            (BT, BK),
+            (1, 0),
+        )
+        b_k = tl.load(p_k, boundary_check=(0, 1))
+        b_kb = b_k * b_beta[:, None]
+        b_A += tl.dot(b_kb.to(b_k.dtype), tl.trans(b_k))
+
+    if USE_G:
+        p_g = tl.make_block_ptr(
+            g_cumsum + bos * H + i_h, (T,), (H,), (i_t * BT,), (BT,), (0,)
+        )
+        b_g = tl.load(p_g, boundary_check=(0,))
+        b_g_diff = b_g[:, None] - b_g[None, :]
+        b_A = b_A * safe_exp(b_g_diff)
+
+    b_A = tl.where(o_t[:, None] > o_t[None, :], b_A, 0)
+    p_A = tl.make_block_ptr(
+        A + (bos * H + i_h) * BT, (T, BT), (BT * H, 1), (i_t * BT, 0), (BT, BT), (1, 0)
+    )
+    tl.store(p_A, b_A.to(p_A.dtype.element_ty), boundary_check=(0, 1))
+
+
+def chunk_scaled_dot_kkt_fwd(
+    k: torch.Tensor,
+    beta: torch.Tensor,
+    g_cumsum: Optional[torch.Tensor] = None,
+    cu_seqlens: Optional[torch.LongTensor] = None,
+    chunk_size: int = 64,
+    output_dtype: torch.dtype = torch.float32,
+) -> torch.Tensor:
+    r"""
+    Compute beta * K * K^T.
+
+    Args:
+        k (torch.Tensor):
+            The key tensor of shape `[B, T, H, K]`.
+        beta (torch.Tensor):
+            The beta tensor of shape `[B, T, H]`.
+        g_cumsum (torch.Tensor):
+            The cumulative sum of the gate tensor of shape `[B, T, H]`.
+            Default: None
+        cu_seqlens (torch.LongTensor):
+            The cumulative sequence lengths of the input tensor.
+            Default: None
+        chunk_size (int):
+            The chunk size. Default: 64.
+        output_dtype (torch.dtype):
+            The dtype of the output tensor. Default: `torch.float32`
+
+    Returns:
+        beta * K * K^T of shape `[B, T, H, BT]` where `BT` is the chunk size.
+    """
+
+    B, T, Hg, K = k.shape
+
+    H = beta.shape[-1]
+    BT = chunk_size
+    chunk_indices = (
+        prepare_chunk_indices(cu_seqlens, BT) if cu_seqlens is not None else None
+    )
+    NT = triton.cdiv(T, BT) if cu_seqlens is None else len(chunk_indices)
+    A = torch.empty(B, T, H, BT, device=k.device, dtype=output_dtype)
+    chunk_scaled_dot_kkt_fwd_kernel[(NT, B * H)](
+        k=k,
+        beta=beta,
+        g_cumsum=g_cumsum,
+        A=A,
+        cu_seqlens=cu_seqlens,
+        chunk_indices=chunk_indices,
+        T=T,
+        H=H,
+        Hg=Hg,
+        K=K,
+        BT=BT,
+        BK=64,
+        num_warps=8,
+        num_stages=3,
+    )
+    return A