vllm-cpu-amxbf16 0.11.2.post2__cp310-cp310-manylinux_2_17_x86_64.whl

vllm/model_executor/layers/fla/ops/cumsum.py

@@ -0,0 +1,280 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
+#
+# This file contains code copied from the flash-linear-attention project.
+# The original source code was licensed under the MIT license and included
+# the following copyright notice:
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+# ruff: noqa: E501
+import warnings
+
+import torch
+
+from vllm.triton_utils import tl, triton
+
+from .index import prepare_chunk_indices
+from .utils import check_shared_mem, input_guard
+
+BS_LIST = [32, 64] if check_shared_mem() else [16, 32]
+
+
+@triton.heuristics({"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,
+    cu_seqlens,
+    chunk_indices,
+    T,
+    B: tl.constexpr,
+    H: tl.constexpr,
+    BT: tl.constexpr,
+    REVERSE: 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
+    tl.store(p_o, b_o.to(p_o.dtype.element_ty), boundary_check=(0,))
+
+
+@triton.heuristics({"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,
+    cu_seqlens,
+    chunk_indices,
+    T,
+    B: tl.constexpr,
+    H: tl.constexpr,
+    S: tl.constexpr,
+    BT: tl.constexpr,
+    BS: tl.constexpr,
+    REVERSE: 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)
+    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,
+    cu_seqlens: torch.Tensor | None = None,
+    head_first: bool = False,
+    output_dtype: torch.dtype | None = 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](
+        g_org,
+        g,
+        cu_seqlens,
+        chunk_indices,
+        T=T,
+        B=B,
+        H=H,
+        BT=BT,
+        HEAD_FIRST=head_first,
+        REVERSE=reverse,
+    )
+    return g
+
+
+def chunk_local_cumsum_vector(
+    g: torch.Tensor,
+    chunk_size: int,
+    reverse: bool = False,
+    cu_seqlens: torch.Tensor | None = None,
+    head_first: bool = False,
+    output_dtype: torch.dtype | None = 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](
+        g_org,
+        g,
+        cu_seqlens,
+        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,
+    cu_seqlens: torch.Tensor | None = None,
+    head_first: bool = False,
+    output_dtype: torch.dtype | None = torch.float,
+    **kwargs,
+) -> torch.Tensor:
+    if not head_first and g.shape[1] < g.shape[2]:
+        warnings.warn(
+            f"Input tensor shape suggests potential format mismatch: seq_len ({g.shape[1]}) < num_heads ({g.shape[2]}). "
+            "This may indicate the inputs were passed in head-first format [B, H, T, ...] "
+            "when head_first=False was specified. "
+            "Please verify your input tensor format matches the expected shape [B, T, H, ...].",
+            stacklevel=2,
+        )
+    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, chunk_size, reverse, cu_seqlens, head_first, output_dtype
+        )
+    elif len(g.shape) == 4:
+        return chunk_local_cumsum_vector(
+            g, chunk_size, reverse, cu_seqlens, head_first, 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"
+        )

vllm/model_executor/layers/fla/ops/fused_recurrent.py

@@ -0,0 +1,390 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
+#
+# This file contains code copied from the flash-linear-attention project.
+# The original source code was licensed under the MIT license and included
+# the following copyright notice:
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+# ruff: noqa: E501
+
+import torch
+
+from vllm.triton_utils import tl, triton
+
+from .op import exp
+
+
+@triton.heuristics(
+    {
+        "USE_INITIAL_STATE": lambda args: args["h0"] is not None,
+        "IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
+        "IS_CONTINUOUS_BATCHING": lambda args: args["ssm_state_indices"] is not None,
+        "IS_SPEC_DECODING": lambda args: args["num_accepted_tokens"] is not None,
+    }
+)
+@triton.jit(do_not_specialize=["N", "T"])
+def fused_recurrent_gated_delta_rule_fwd_kernel(
+    q,
+    k,
+    v,
+    g,
+    beta,
+    o,
+    h0,
+    ht,
+    cu_seqlens,
+    ssm_state_indices,
+    num_accepted_tokens,
+    scale,
+    N: tl.int64,  # num of sequences
+    T: tl.int64,  # num of tokens
+    B: tl.constexpr,
+    H: tl.constexpr,
+    HV: tl.constexpr,
+    K: tl.constexpr,
+    V: tl.constexpr,
+    BK: tl.constexpr,
+    BV: tl.constexpr,
+    stride_init_state_token: tl.constexpr,
+    stride_final_state_token: tl.constexpr,
+    stride_indices_seq: tl.constexpr,
+    stride_indices_tok: tl.constexpr,
+    USE_INITIAL_STATE: tl.constexpr,  # whether to use initial state
+    INPLACE_FINAL_STATE: tl.constexpr,  # whether to store final state inplace
+    IS_BETA_HEADWISE: tl.constexpr,  # whether beta is headwise vector or scalar,
+    USE_QK_L2NORM_IN_KERNEL: tl.constexpr,
+    IS_VARLEN: tl.constexpr,
+    IS_CONTINUOUS_BATCHING: tl.constexpr,
+    IS_SPEC_DECODING: tl.constexpr,
+    IS_KDA: tl.constexpr,
+):
+    i_k, i_v, i_nh = tl.program_id(0), tl.program_id(1), tl.program_id(2)
+    i_n, i_hv = i_nh // HV, i_nh % HV
+    i_h = i_hv // (HV // H)
+    if IS_VARLEN:
+        bos, eos = (
+            tl.load(cu_seqlens + i_n).to(tl.int64),
+            tl.load(cu_seqlens + i_n + 1).to(tl.int64),
+        )
+        all = T
+        T = eos - bos
+    else:
+        bos, eos = i_n * T, i_n * T + T
+        all = B * T
+
+    if T == 0:
+        # no tokens to process for this sequence
+        return
+
+    o_k = i_k * BK + tl.arange(0, BK)
+    o_v = i_v * BV + tl.arange(0, BV)
+
+    p_q = q + (bos * H + i_h) * K + o_k
+    p_k = k + (bos * H + i_h) * K + o_k
+    p_v = v + (bos * HV + i_hv) * V + o_v
+    if IS_BETA_HEADWISE:
+        p_beta = beta + (bos * HV + i_hv) * V + o_v
+    else:
+        p_beta = beta + bos * HV + i_hv
+
+    if not IS_KDA:
+        p_g = g + bos * HV + i_hv
+    else:
+        p_gk = g + (bos * HV + i_hv) * K + o_k
+
+    p_o = o + ((i_k * all + bos) * HV + i_hv) * V + o_v
+
+    mask_k = o_k < K
+    mask_v = o_v < V
+    mask_h = mask_k[:, None] & mask_v[None, :]
+
+    b_h = tl.zeros([BK, BV], dtype=tl.float32)
+    if USE_INITIAL_STATE:
+        if IS_CONTINUOUS_BATCHING:
+            if IS_SPEC_DECODING:
+                i_t = tl.load(num_accepted_tokens + i_n).to(tl.int64) - 1
+            else:
+                i_t = 0
+            p_h0 = (
+                h0
+                + tl.load(ssm_state_indices + i_n * stride_indices_seq + i_t).to(
+                    tl.int64
+                )
+                * stride_init_state_token
+            )
+        else:
+            p_h0 = h0 + bos * HV * K * V
+        p_h0 = p_h0 + i_hv * K * V + o_k[:, None] * V + o_v[None, :]
+        b_h += tl.load(p_h0, mask=mask_h, other=0).to(tl.float32)
+
+    for i_t in range(0, T):
+        b_q = tl.load(p_q, mask=mask_k, other=0).to(tl.float32)
+        b_k = tl.load(p_k, mask=mask_k, other=0).to(tl.float32)
+        b_v = tl.load(p_v, mask=mask_v, other=0).to(tl.float32)
+
+        if USE_QK_L2NORM_IN_KERNEL:
+            b_q = b_q / tl.sqrt(tl.sum(b_q * b_q) + 1e-6)
+            b_k = b_k / tl.sqrt(tl.sum(b_k * b_k) + 1e-6)
+        b_q = b_q * scale
+        # [BK, BV]
+        if not IS_KDA:
+            b_g = tl.load(p_g).to(tl.float32)
+            b_h *= exp(b_g)
+        else:
+            b_gk = tl.load(p_gk).to(tl.float32)
+            b_h *= exp(b_gk[:, None])
+        # [BV]
+        b_v -= tl.sum(b_h * b_k[:, None], 0)
+        if IS_BETA_HEADWISE:
+            b_beta = tl.load(p_beta, mask=mask_v, other=0).to(tl.float32)
+        else:
+            b_beta = tl.load(p_beta).to(tl.float32)
+        b_v *= b_beta
+        # [BK, BV]
+        b_h += b_k[:, None] * b_v[None, :]
+        # [BV]
+        b_o = tl.sum(b_h * b_q[:, None], 0)
+        tl.store(p_o, b_o.to(p_o.dtype.element_ty), mask=mask_v)
+
+        # keep the states for multi-query tokens
+        if INPLACE_FINAL_STATE:
+            p_ht = (
+                ht
+                + tl.load(ssm_state_indices + i_n * stride_indices_seq + i_t).to(
+                    tl.int64
+                )
+                * stride_final_state_token
+            )
+        else:
+            p_ht = ht + (bos + i_t) * stride_final_state_token
+        p_ht = p_ht + i_hv * K * V + o_k[:, None] * V + o_v[None, :]
+        tl.store(p_ht, b_h.to(p_ht.dtype.element_ty), mask=mask_h)
+
+        p_q += H * K
+        p_k += H * K
+        p_o += HV * V
+        p_v += HV * V
+        if not IS_KDA:
+            p_g += HV
+        else:
+            p_gk += HV * K
+        p_beta += HV * (V if IS_BETA_HEADWISE else 1)
+
+
+def fused_recurrent_gated_delta_rule_fwd(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    g: torch.Tensor,
+    beta: torch.Tensor,
+    scale: float,
+    initial_state: torch.Tensor,
+    inplace_final_state: bool = True,
+    cu_seqlens: torch.LongTensor | None = None,
+    ssm_state_indices: torch.Tensor | None = None,
+    num_accepted_tokens: torch.Tensor | None = None,
+    use_qk_l2norm_in_kernel: bool = False,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    B, T, H, K, V = *k.shape, v.shape[-1]
+    HV = v.shape[2]
+    N = B if cu_seqlens is None else len(cu_seqlens) - 1
+    BK, BV = triton.next_power_of_2(K), min(triton.next_power_of_2(V), 8)
+    NK, NV = triton.cdiv(K, BK), triton.cdiv(V, BV)
+    assert NK == 1, "NK > 1 is not supported yet"
+    num_stages = 3
+    num_warps = 1
+
+    o = q.new_empty(NK, *v.shape)
+    if inplace_final_state:
+        final_state = initial_state
+    else:
+        final_state = q.new_empty(T, HV, K, V, dtype=initial_state.dtype)
+
+    stride_init_state_token = initial_state.stride(0)
+    stride_final_state_token = final_state.stride(0)
+
+    if ssm_state_indices is None:
+        stride_indices_seq, stride_indices_tok = 1, 1
+    elif ssm_state_indices.ndim == 1:
+        stride_indices_seq, stride_indices_tok = ssm_state_indices.stride(0), 1
+    else:
+        stride_indices_seq, stride_indices_tok = ssm_state_indices.stride()
+
+    grid = (NK, NV, N * HV)
+    fused_recurrent_gated_delta_rule_fwd_kernel[grid](
+        q=q,
+        k=k,
+        v=v,
+        g=g,
+        beta=beta,
+        o=o,
+        h0=initial_state,
+        ht=final_state,
+        cu_seqlens=cu_seqlens,
+        ssm_state_indices=ssm_state_indices,
+        num_accepted_tokens=num_accepted_tokens,
+        scale=scale,
+        N=N,
+        T=T,
+        B=B,
+        H=H,
+        HV=HV,
+        K=K,
+        V=V,
+        BK=BK,
+        BV=BV,
+        stride_init_state_token=stride_init_state_token,
+        stride_final_state_token=stride_final_state_token,
+        stride_indices_seq=stride_indices_seq,
+        stride_indices_tok=stride_indices_tok,
+        IS_BETA_HEADWISE=beta.ndim == v.ndim,
+        USE_QK_L2NORM_IN_KERNEL=use_qk_l2norm_in_kernel,
+        INPLACE_FINAL_STATE=inplace_final_state,
+        IS_KDA=False,
+        num_warps=num_warps,
+        num_stages=num_stages,
+    )
+    o = o.squeeze(0)
+    return o, final_state
+
+
+class FusedRecurrentFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        g: torch.Tensor,
+        beta: torch.Tensor,
+        scale: float,
+        initial_state: torch.Tensor,
+        inplace_final_state: bool = True,
+        cu_seqlens: torch.LongTensor | None = None,
+        ssm_state_indices: torch.Tensor | None = None,
+        num_accepted_tokens: torch.Tensor | None = None,
+        use_qk_l2norm_in_kernel: bool = False,
+    ):
+        o, final_state = fused_recurrent_gated_delta_rule_fwd(
+            q=q.contiguous(),
+            k=k.contiguous(),
+            v=v.contiguous(),
+            g=g.contiguous(),
+            beta=beta.contiguous(),
+            scale=scale,
+            initial_state=initial_state,
+            inplace_final_state=inplace_final_state,
+            cu_seqlens=cu_seqlens,
+            ssm_state_indices=ssm_state_indices,
+            num_accepted_tokens=num_accepted_tokens,
+            use_qk_l2norm_in_kernel=use_qk_l2norm_in_kernel,
+        )
+
+        return o, final_state
+
+
+def fused_recurrent_gated_delta_rule(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    g: torch.Tensor,
+    beta: torch.Tensor = None,
+    scale: float = None,
+    initial_state: torch.Tensor = None,
+    inplace_final_state: bool = True,
+    cu_seqlens: torch.LongTensor | None = None,
+    ssm_state_indices: torch.Tensor | None = None,
+    num_accepted_tokens: torch.Tensor | None = None,
+    use_qk_l2norm_in_kernel: bool = False,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    r"""
+    Args:
+        q (torch.Tensor):
+            queries of shape `[B, T, H, K]`.
+        k (torch.Tensor):
+            keys of shape `[B, T, H, K]`.
+        v (torch.Tensor):
+            values of shape `[B, T, HV, V]`.
+            GVA is applied if `HV > H`.
+        g (torch.Tensor):
+            g (decays) of shape `[B, T, HV]`.
+        beta (torch.Tensor):
+            betas of shape `[B, T, HV]`.
+        scale (Optional[int]):
+            Scale factor for the RetNet attention scores.
+            If not provided, it will default to `1 / sqrt(K)`. Default: `None`.
+        initial_state (Optional[torch.Tensor]):
+            Initial state of shape `[N, HV, K, V]` for `N` input sequences.
+            For equal-length input sequences, `N` equals the batch size `B`.
+            Default: `None`.
+        inplace_final_state: bool:
+            Whether to store the final state in-place to save memory.
+            Default: `True`.
+        cu_seqlens (torch.LongTensor):
+            Cumulative sequence lengths of shape `[N+1]` used for variable-length training,
+            consistent with the FlashAttention API.
+        ssm_state_indices (Optional[torch.Tensor]):
+            Indices to map the input sequences to the initial/final states.
+        num_accepted_tokens (Optional[torch.Tensor]):
+            Number of accepted tokens for each sequence during decoding.
+
+    Returns:
+        o (torch.Tensor):
+            Outputs of shape `[B, T, HV, V]`.
+        final_state (torch.Tensor):
+            Final state of shape `[N, HV, K, V]`.
+
+    Examples::
+        >>> import torch
+        >>> import torch.nn.functional as F
+        >>> from einops import rearrange
+        >>> from fla.ops.gated_delta_rule import fused_recurrent_gated_delta_rule
+        # inputs with equal lengths
+        >>> B, T, H, HV, K, V = 4, 2048, 4, 8, 512, 512
+        >>> q = torch.randn(B, T, H, K, device='cuda')
+        >>> k = F.normalize(torch.randn(B, T, H, K, device='cuda'), p=2, dim=-1)
+        >>> v = torch.randn(B, T, HV, V, device='cuda')
+        >>> g = F.logsigmoid(torch.rand(B, T, HV, device='cuda'))
+        >>> beta = torch.rand(B, T, HV, device='cuda').sigmoid()
+        >>> h0 = torch.randn(B, HV, K, V, device='cuda')
+        >>> o, ht = fused_gated_recurrent_delta_rule(
+            q, k, v, g, beta,
+            initial_state=h0,
+        )
+        # for variable-length inputs, the batch size `B` is expected to be 1 and `cu_seqlens` is required
+        >>> q, k, v, g, beta = map(lambda x: rearrange(x, 'b t ... -> 1 (b t) ...'), (q, k, v, g, beta))
+        # for a batch with 4 sequences, `cu_seqlens` with 5 start/end positions are expected
+        >>> cu_seqlens = q.new_tensor([0, 2048, 4096, 6144, 8192], dtype=torch.long)
+        >>> o_var, ht_var = fused_gated_recurrent_delta_rule(
+            q, k, v, g, beta,
+            initial_state=h0,
+            cu_seqlens=cu_seqlens
+        )
+    """
+    if cu_seqlens is not None and q.shape[0] != 1:
+        raise ValueError(
+            f"The batch size is expected to be 1 rather than {q.shape[0]} when using `cu_seqlens`."
+            f"Please flatten variable-length inputs before processing."
+        )
+    if scale is None:
+        scale = k.shape[-1] ** -0.5
+    else:
+        assert scale > 0, "scale must be positive"
+    if beta is None:
+        beta = torch.ones_like(q[..., 0])
+    o, final_state = FusedRecurrentFunction.apply(
+        q,
+        k,
+        v,
+        g,
+        beta,
+        scale,
+        initial_state,
+        inplace_final_state,
+        cu_seqlens,
+        ssm_state_indices,
+        num_accepted_tokens,
+        use_qk_l2norm_in_kernel,
+    )
+    return o, final_state

vllm/model_executor/layers/fla/ops/index.py

@@ -0,0 +1,41 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
+#
+# This file contains code copied from the flash-linear-attention project.
+# The original source code was licensed under the MIT license and included
+# the following copyright notice:
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+# ruff: noqa: E501
+import torch
+
+from vllm.triton_utils import triton
+
+from .utils import tensor_cache
+
+
+@tensor_cache
+def prepare_lens(cu_seqlens: torch.LongTensor) -> torch.LongTensor:
+    return cu_seqlens[1:] - cu_seqlens[:-1]
+
+
+@tensor_cache
+def prepare_chunk_indices(
+    cu_seqlens: torch.LongTensor, chunk_size: int
+) -> torch.LongTensor:
+    indices = torch.cat(
+        [
+            torch.arange(n)
+            for n in triton.cdiv(prepare_lens(cu_seqlens), chunk_size).tolist()
+        ]
+    )
+    return torch.stack([indices.eq(0).cumsum(0) - 1, indices], 1).to(cu_seqlens)
+
+
+@tensor_cache
+def prepare_chunk_offsets(
+    cu_seqlens: torch.LongTensor, chunk_size: int
+) -> torch.LongTensor:
+    return torch.cat(
+        [cu_seqlens.new_tensor([0]), triton.cdiv(prepare_lens(cu_seqlens), chunk_size)]
+    ).cumsum(-1)