sglang 0.5.1.post3__py3-none-any.whl → 0.5.2__py3-none-any.whl

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

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

@@ -0,0 +1,300 @@
+# Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/utils/cumsum.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.utils import check_shared_mem, input_guard
+
+BS_LIST = [32, 64] if check_shared_mem() else [16, 32]
+
+
+@triton.heuristics(
+    {
+        "HAS_SCALE": lambda args: args["scale"] is not None,
+        "IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
+    }
+)
+# @triton.autotune(
+#     configs=[triton.Config({}, num_warps=num_warps) for num_warps in [1, 2, 4, 8]],
+#     key=["B", "H", "BT", "IS_VARLEN", "REVERSE"],
+# )
+@triton.jit(do_not_specialize=["T"])
+def chunk_local_cumsum_scalar_kernel(
+    s,
+    o,
+    scale,
+    cu_seqlens,
+    chunk_indices,
+    T,
+    B: tl.constexpr,
+    H: tl.constexpr,
+    BT: tl.constexpr,
+    REVERSE: tl.constexpr,
+    HAS_SCALE: tl.constexpr,
+    IS_VARLEN: tl.constexpr,
+    HEAD_FIRST: 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
+
+    if HEAD_FIRST:
+        p_s = tl.make_block_ptr(
+            s + bos * H + i_h * T, (T,), (1,), (i_t * BT,), (BT,), (0,)
+        )
+        p_o = tl.make_block_ptr(
+            o + bos * H + i_h * T, (T,), (1,), (i_t * BT,), (BT,), (0,)
+        )
+    else:
+        p_s = tl.make_block_ptr(s + bos * H + i_h, (T,), (H,), (i_t * BT,), (BT,), (0,))
+        p_o = tl.make_block_ptr(o + bos * H + i_h, (T,), (H,), (i_t * BT,), (BT,), (0,))
+    # [BT]
+    b_s = tl.load(p_s, boundary_check=(0,)).to(tl.float32)
+    b_o = tl.cumsum(b_s, axis=0)
+    if REVERSE:
+        b_z = tl.sum(b_s, axis=0)
+        b_o = -b_o + b_z[None] + b_s
+    if HAS_SCALE:
+        b_o *= scale
+    tl.store(p_o, b_o.to(p_o.dtype.element_ty), boundary_check=(0,))
+
+
+@triton.heuristics(
+    {
+        "HAS_SCALE": lambda args: args["scale"] is not None,
+        "IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
+    }
+)
+@triton.autotune(
+    configs=[
+        triton.Config({"BS": BS}, num_warps=num_warps)
+        for BS in BS_LIST
+        for num_warps in [2, 4, 8]
+    ],
+    key=["B", "H", "S", "BT", "IS_VARLEN", "REVERSE"],
+)
+@triton.jit(do_not_specialize=["T"])
+def chunk_local_cumsum_vector_kernel(
+    s,
+    o,
+    scale,
+    cu_seqlens,
+    chunk_indices,
+    T,
+    B: tl.constexpr,
+    H: tl.constexpr,
+    S: tl.constexpr,
+    BT: tl.constexpr,
+    BS: tl.constexpr,
+    REVERSE: tl.constexpr,
+    HAS_SCALE: tl.constexpr,
+    IS_VARLEN: tl.constexpr,
+    HEAD_FIRST: tl.constexpr,
+):
+    i_s, 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_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_i = tl.arange(0, BT)
+    if REVERSE:
+        m_s = tl.where(o_i[:, None] <= o_i[None, :], 1.0, 0.0)
+    else:
+        m_s = tl.where(o_i[:, None] >= o_i[None, :], 1.0, 0.0)
+
+    if HEAD_FIRST:
+        p_s = tl.make_block_ptr(
+            s + (bos * H + i_h * T) * S,
+            (T, S),
+            (S, 1),
+            (i_t * BT, i_s * BS),
+            (BT, BS),
+            (1, 0),
+        )
+        p_o = tl.make_block_ptr(
+            o + (bos * H + i_h * T) * S,
+            (T, S),
+            (S, 1),
+            (i_t * BT, i_s * BS),
+            (BT, BS),
+            (1, 0),
+        )
+    else:
+        p_s = tl.make_block_ptr(
+            s + (bos * H + i_h) * S,
+            (T, S),
+            (H * S, 1),
+            (i_t * BT, i_s * BS),
+            (BT, BS),
+            (1, 0),
+        )
+        p_o = tl.make_block_ptr(
+            o + (bos * H + i_h) * S,
+            (T, S),
+            (H * S, 1),
+            (i_t * BT, i_s * BS),
+            (BT, BS),
+            (1, 0),
+        )
+    # [BT, BS]
+    b_s = tl.load(p_s, boundary_check=(0, 1)).to(tl.float32)
+    b_o = tl.dot(m_s, b_s, allow_tf32=False)
+    if HAS_SCALE:
+        b_o *= scale
+    tl.store(p_o, b_o.to(p_o.dtype.element_ty), boundary_check=(0, 1))
+
+
+def chunk_local_cumsum_scalar(
+    g: torch.Tensor,
+    chunk_size: int,
+    reverse: bool = False,
+    scale: float = None,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    head_first: bool = False,
+    output_dtype: Optional[torch.dtype] = torch.float,
+) -> torch.Tensor:
+    if head_first:
+        B, H, T = g.shape
+    else:
+        B, T, H = g.shape
+    assert chunk_size == 2 ** (
+        chunk_size.bit_length() - 1
+    ), "chunk_size must be a power of 2"
+    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)
+    g_org, g = g, torch.empty_like(g, dtype=output_dtype or g.dtype)
+    grid = (NT, B * H)
+    chunk_local_cumsum_scalar_kernel[grid](
+        s=g_org,
+        o=g,
+        scale=scale,
+        cu_seqlens=cu_seqlens,
+        chunk_indices=chunk_indices,
+        T=T,
+        B=B,
+        H=H,
+        BT=BT,
+        HEAD_FIRST=head_first,
+        REVERSE=reverse,
+        num_warps=8,
+        num_stages=3,
+    )
+    return g
+
+
+def chunk_local_cumsum_vector(
+    g: torch.Tensor,
+    chunk_size: int,
+    reverse: bool = False,
+    scale: float = None,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    head_first: bool = False,
+    output_dtype: Optional[torch.dtype] = torch.float,
+) -> torch.Tensor:
+    if head_first:
+        B, H, T, S = g.shape
+    else:
+        B, T, H, S = g.shape
+    BT = chunk_size
+    chunk_indices = (
+        prepare_chunk_indices(cu_seqlens, chunk_size)
+        if cu_seqlens is not None
+        else None
+    )
+    NT = triton.cdiv(T, BT) if cu_seqlens is None else len(chunk_indices)
+    assert chunk_size == 2 ** (
+        chunk_size.bit_length() - 1
+    ), "chunk_size must be a power of 2"
+
+    g_org, g = g, torch.empty_like(g, dtype=output_dtype or g.dtype)
+
+    def grid(meta):
+        return (triton.cdiv(meta["S"], meta["BS"]), NT, B * H)
+
+    # keep cumulative normalizer in fp32
+    # this kernel is equivalent to
+    # g = g.view(B, H, NT, BT, -1).cumsum(-2).view(B, H, T, -1)
+    chunk_local_cumsum_vector_kernel[grid](
+        s=g_org,
+        o=g,
+        scale=scale,
+        cu_seqlens=cu_seqlens,
+        chunk_indices=chunk_indices,
+        T=T,
+        B=B,
+        H=H,
+        S=S,
+        BT=BT,
+        HEAD_FIRST=head_first,
+        REVERSE=reverse,
+    )
+    return g
+
+
+@input_guard
+def chunk_local_cumsum(
+    g: torch.Tensor,
+    chunk_size: int,
+    reverse: bool = False,
+    scale: float = None,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    head_first: bool = False,
+    output_dtype: Optional[torch.dtype] = torch.float,
+    **kwargs,
+) -> torch.Tensor:
+    if cu_seqlens is not None:
+        assert (
+            g.shape[0] == 1
+        ), "Only batch size 1 is supported when cu_seqlens are provided"
+    if len(g.shape) == 3:
+        return chunk_local_cumsum_scalar(
+            g=g,
+            chunk_size=chunk_size,
+            reverse=reverse,
+            scale=scale,
+            cu_seqlens=cu_seqlens,
+            head_first=head_first,
+            output_dtype=output_dtype,
+        )
+    elif len(g.shape) == 4:
+        return chunk_local_cumsum_vector(
+            g=g,
+            chunk_size=chunk_size,
+            reverse=reverse,
+            scale=scale,
+            cu_seqlens=cu_seqlens,
+            head_first=head_first,
+            output_dtype=output_dtype,
+        )
+    else:
+        raise ValueError(
+            f"Unsupported input shape {g.shape}, "
+            f"which should be (B, T, H, D) if `head_first=False` "
+            f"or (B, H, T, D) otherwise"
+        )