vllm-cpp-sys 0.0.1

# triton_kernels/chunk_delta_h.py
#
# GDN chunked state-recurrence (delta_h), the SANCTIONED CUDA-only Triton AOT
# fast-path for the PROVEN codegen-bound kernel (see .agents/discipline.md
# "SANCTIONED EXCEPTION", .agents/mission.md, porting-inventory.md §9). Compiled
# to a cubin at BUILD time (cmake/TritonAOT.cmake, gated VLLM_CPP_TRITON=OFF by
# default) and embedded in libvllm; the RUNTIME is Triton/Python-free (cubin via
# the CUDA driver API). The portable hand-C++ CUDA kernel
# (GdnChunkDeltaHRegRingKernel in src/vt/cuda/cuda_gdn.cu) + the CPU reference are
# PRESERVED as the fallback and remain the default when VLLM_CPP_TRITON=OFF.
#
# Ported VERBATIM FROM (vLLM 0.24.0 oracle venv):
#   vllm/model_executor/layers/fla/ops/chunk_delta_h.py:42-315
#     @triton.jit chunk_gated_delta_rule_fwd_kernel_h_blockdim64
#   (upstream flash-linear-attention; MIT). The kernel BODY is byte-for-byte the
#   FLA source. The ONLY adaptations for AOT compilation are:
#     (1) the @triton.heuristics / @triton.autotune decorators are removed — the
#         flags (USE_*) and dims (H,Hg,K,V,BT,BV) are PINNED per-shape via the
#         triton.tools.compile SIGNATURE (see CMakeLists.txt), and the winning
#         (BV,num_warps,num_stages) is selected there, exactly what the autotuner
#         would pick for the gate shape;
#     (2) one trailing runtime scalar `NH` (= N*H, the number of (sequence,head)
#         pairs) is appended: the FLA launch grid is (cdiv(V,BV), N*H) but N (the
#         sequence count) is NOT a kernel argument, so the grid-y extent is carried
#         in `NH` for the AOT launcher's baked grid expression. `NH` is unused by
#         the kernel body (dead arg → eliminated → the compute codegen is
#         identical to FLA's).
#
# Pinned flags for the vllm.cpp GDN chunked-prefill call site (matches the hand
# path in cuda_gdn.cu): USE_G=1, USE_GK=0, USE_INITIAL_STATE=1,
# STORE_FINAL_STATE=1, SAVE_NEW_VALUE=1, IS_VARLEN=1, USE_EXP2=0. Buffer layout is
# a 1:1 drop-in (verified stride-for-stride against the FLA pointer arithmetic):
#   k=[T,Hg,K] bf16, v(=u)=[T,H,V] bf16, w=[T,H,K] bf16, v_new=[T,H,V] bf16,
#   g(=gcum within-chunk cumsum)=[T,H] f32, h(=hstate snapshot)=[NT,H,V,K] bf16,
#   h0/ht(=state)=[N,H,V,K] f32, cu_seqlens=[N+1] i32, chunk_offsets=[N] i32.
import triton
import triton.language as tl

# fla/ops/op.py: exp/exp2 (FLA_USE_FAST_OPS=0 default -> tl.exp / tl.exp2).
exp = tl.exp
exp2 = tl.exp2


@triton.jit(do_not_specialize=["T", "NH"])
def chunk_gated_delta_rule_fwd_kernel_h_blockdim64(
    k,
    v,
    w,
    v_new,
    g,
    gk,
    h,
    h0,
    ht,
    cu_seqlens,
    chunk_offsets,
    T,
    NH,  # AOT grid-carrier (= N*H): grid-y extent; unused by the body. See header.
    H: tl.constexpr,
    Hg: tl.constexpr,
    K: tl.constexpr,
    V: tl.constexpr,
    BT: tl.constexpr,
    BV: tl.constexpr,
    USE_G: tl.constexpr,
    USE_GK: tl.constexpr,
    USE_INITIAL_STATE: tl.constexpr,
    STORE_FINAL_STATE: tl.constexpr,
    SAVE_NEW_VALUE: tl.constexpr,
    IS_VARLEN: tl.constexpr,
    USE_EXP2: 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

    # [BV, BK]
    b_h1 = tl.zeros([BV, 64], dtype=tl.float32)
    if K > 64:
        b_h2 = tl.zeros([BV, 64], dtype=tl.float32)
    if K > 128:
        b_h3 = tl.zeros([BV, 64], dtype=tl.float32)
    if K > 192:
        b_h4 = tl.zeros([BV, 64], dtype=tl.float32)

    # calculate offset
    h += ((boh * H + i_h) * V * K).to(tl.int64)
    v += ((bos * H + i_h) * V).to(tl.int64)
    k += ((bos * Hg + i_h // (H // Hg)) * K).to(tl.int64)
    w += ((bos * H + i_h) * K).to(tl.int64)
    if SAVE_NEW_VALUE:
        v_new += ((bos * H + i_h) * V).to(tl.int64)
    stride_v = H * V
    stride_h = H * V * K
    stride_k = Hg * K
    stride_w = H * K
    if USE_INITIAL_STATE:
        h0 = h0 + i_nh * V * K
    if STORE_FINAL_STATE:
        ht = ht + i_nh * V * K

    # load initial state
    if USE_INITIAL_STATE:
        p_h0_1 = tl.make_block_ptr(h0, (V, K), (K, 1), (i_v * BV, 0), (BV, 64), (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, (V, K), (K, 1), (i_v * BV, 64), (BV, 64), (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, (V, K), (K, 1), (i_v * BV, 128), (BV, 64), (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, (V, K), (K, 1), (i_v * BV, 192), (BV, 64), (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.to(tl.int64) * stride_h,
            (V, K),
            (K, 1),
            (i_v * BV, 0),
            (BV, 64),
            (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.to(tl.int64) * stride_h,
                (V, K),
                (K, 1),
                (i_v * BV, 64),
                (BV, 64),
                (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.to(tl.int64) * stride_h,
                (V, K),
                (K, 1),
                (i_v * BV, 128),
                (BV, 64),
                (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.to(tl.int64) * stride_h,
                (V, K),
                (K, 1),
                (i_v * BV, 192),
                (BV, 64),
                (1, 0),
            )
            tl.store(p_h4, b_h4.to(p_h4.dtype.element_ty), boundary_check=(0, 1))

        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 = tl.dot(b_w, tl.trans(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 += tl.dot(b_w, tl.trans(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 += tl.dot(b_w, tl.trans(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 += tl.dot(b_w, tl.trans(b_h4).to(b_w.dtype))
        p_v = tl.make_block_ptr(
            v, (T, V), (stride_v, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
        )
        b_v = tl.load(p_v, boundary_check=(0, 1)) - b_v

        if SAVE_NEW_VALUE:
            p_v = 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, b_v.to(p_v.dtype.element_ty), boundary_check=(0, 1))

        last_idx = min((i_t.to(tl.int64) + 1) * BT, T) - 1
        if USE_G:
            m_t = (i_t.to(tl.int64) * BT + tl.arange(0, BT)) < T
            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,))
            if USE_EXP2:
                b_v = b_v * tl.where(m_t, exp2(b_g_last - b_g), 0)[:, None]
                b_g_last = exp2(b_g_last)
            else:
                b_v = b_v * tl.where(m_t, exp(b_g_last - b_g), 0)[:, None]
                b_g_last = exp(b_g_last)
            b_h1 *= b_g_last
            if K > 64:
                b_h2 *= b_g_last
            if K > 128:
                b_h3 *= b_g_last
            if K > 192:
                b_h4 *= b_g_last

        if USE_GK:
            o_k1 = tl.arange(0, 64)
            b_gk_last1 = tl.load(
                gk + (bos + last_idx) * H * K + i_h * K + o_k1,
                mask=(o_k1 < K),
                other=0.0,
            )
            if USE_EXP2:
                b_h1 *= exp2(b_gk_last1)[None, :]
            else:
                b_h1 *= exp(b_gk_last1)[None, :]
            if K > 64:
                o_k2 = 64 + o_k1
                b_gk_last2 = tl.load(
                    gk + (bos + last_idx) * H * K + i_h * K + o_k2,
                    mask=(o_k2 < K),
                    other=0.0,
                )
                if USE_EXP2:
                    b_h2 *= exp2(b_gk_last2)[None, :]
                else:
                    b_h2 *= exp(b_gk_last2)[None, :]
            if K > 128:
                o_k3 = 128 + o_k1
                b_gk_last3 = tl.load(
                    gk + (bos + last_idx) * H * K + i_h * K + o_k3,
                    mask=(o_k3 < K),
                    other=0.0,
                )
                if USE_EXP2:
                    b_h3 *= exp2(b_gk_last3)[None, :]
                else:
                    b_h3 *= exp(b_gk_last3)[None, :]
            if K > 192:
                o_k4 = 192 + o_k1
                b_gk_last4 = tl.load(
                    gk + (bos + last_idx) * H * K + i_h * K + o_k4,
                    mask=(o_k4 < K),
                    other=0.0,
                )
                if USE_EXP2:
                    b_h4 *= exp2(b_gk_last4)[None, :]
                else:
                    b_h4 *= exp(b_gk_last4)[None, :]
        b_v = b_v.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.trans(tl.dot(b_k, b_v))
        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.trans(tl.dot(b_k, b_v))
        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.trans(tl.dot(b_k, b_v))
        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.trans(tl.dot(b_k, b_v))
    # epilogue
    if STORE_FINAL_STATE:
        p_ht = tl.make_block_ptr(ht, (V, K), (K, 1), (i_v * BV, 0), (BV, 64), (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, (V, K), (K, 1), (i_v * BV, 64), (BV, 64), (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, (V, K), (K, 1), (i_v * BV, 128), (BV, 64), (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, (V, K), (K, 1), (i_v * BV, 192), (BV, 64), (1, 0)
            )
            tl.store(p_ht, b_h4.to(p_ht.dtype.element_ty), boundary_check=(0, 1))