vllm.model_executor.layers.mhc ¶

def _hc_head_fused_kernel(
    hs_flat: torch.Tensor,
    fn: torch.Tensor,
    hc_scale: torch.Tensor,
    hc_base: torch.Tensor,
    out: torch.Tensor,
    hidden_size: int,
    rms_eps: float,
    hc_eps: float,
    hc_mult: int,
) -> None:
    """Fill pre-allocated `out` (T, H) in-place with the hc_head result."""
    if hs_flat.shape[0] > 0:
        hc_head_fuse_tilelang(
            hs_flat,
            fn,
            hc_scale,
            hc_base,
            out,
            hidden_size,
            rms_eps,
            hc_eps,
            hc_mult,
        )

@tilelang.jit(
    pass_configs={
        tilelang.PassConfigKey.TL_DISABLE_WARP_SPECIALIZED: True,
        tilelang.PassConfigKey.TL_DISABLE_TMA_LOWER: True,
        tilelang.PassConfigKey.TL_PTXAS_REGISTER_USAGE_LEVEL: 10,
    },
)
def hc_head_fuse_tilelang(
    residual,
    fn,
    hc_scale,
    hc_base,
    out,
    hidden_size: int,
    rms_eps: float,
    hc_eps: float,
    hc_mult: int = 4,
    n_thr: int = 128,
    h_blk: int = 1024,
):
    """Two-pass fused kernel for hc_head.

    Pass 1: accumulate per-token squared sum and hc_mult dot-products
            (projections onto fn rows) using cross-thread reducers.
    Pass 2: apply sigmoid-gated weighted sum of residual channels to output.

    Avoids materialising mixes / rsqrt / pre tensors to global memory.
    """
    num_tokens = T.dynamic("num_tokens")
    hc_dim = hc_mult * hidden_size
    h_block = math.gcd(h_blk, hidden_size)
    n_h = hidden_size // h_block

    residual: T.Tensor[[num_tokens, hc_mult, hidden_size], T.bfloat16]  # type: ignore[no-redef,valid-type]
    fn: T.Tensor[[hc_mult, hc_dim], T.float32]  # type: ignore[no-redef,valid-type]
    hc_scale: T.Tensor[[1], T.float32]  # type: ignore[no-redef,valid-type]
    hc_base: T.Tensor[[hc_mult], T.float32]  # type: ignore[no-redef,valid-type]
    out: T.Tensor[[num_tokens, hidden_size], T.bfloat16]  # type: ignore[no-redef,valid-type]

    with T.Kernel(num_tokens, threads=n_thr) as i:
        T.pdl_sync()

        # ------------------------------------------------------------------
        # Pass 1 – for each residual channel m_c and h_block:
        #   • accumulate squared sum (for RMS norm denominator)
        #   • accumulate hc_mult dot-products with fn rows
        # ------------------------------------------------------------------
        sqrsum_r = T.alloc_reducer((1,), T.float32, replication="all")
        mixes_r = T.alloc_reducer((hc_mult,), T.float32, replication="all")
        T.fill(sqrsum_r, 0.0)
        T.fill(mixes_r, 0.0)

        for m_c in T.serial(hc_mult):
            for i_h in T.serial(n_h):
                x_local = T.alloc_fragment(h_block, T.float32)
                T.copy(residual[i, m_c, i_h * h_block], x_local)

                for k in T.Parallel(h_block):
                    sqrsum_r[0] += x_local[k] * x_local[k]

                for m_m in T.unroll(hc_mult):
                    fn_local = T.alloc_fragment(h_block, T.float32)
                    T.copy(fn[m_m, m_c * hidden_size + i_h * h_block], fn_local)
                    for k in T.Parallel(h_block):
                        mixes_r[m_m] += x_local[k] * fn_local[k]

        T.finalize_reducer(sqrsum_r)
        T.finalize_reducer(mixes_r)

        # ------------------------------------------------------------------
        # Compute pre_mix = sigmoid(mix * rsqrt * scale + base) + eps
        # ------------------------------------------------------------------
        pre_mix_shared = T.alloc_shared(hc_mult, T.float32)
        rsqrt_val = T.alloc_fragment(1, T.float32)
        rsqrt_val[0] = T.rsqrt(sqrsum_r[0] / hc_dim + rms_eps)
        for m in T.Parallel(hc_mult):
            pre_mix_shared[m] = (
                T.sigmoid(mixes_r[m] * rsqrt_val[0] * hc_scale[0] + hc_base[m]) + hc_eps
            )

        # ------------------------------------------------------------------
        # Pass 2 – apply_mix: pipelined weighted sum over residual channels
        # ------------------------------------------------------------------
        for i0_h in T.Pipelined(n_h, num_stages=2):
            xs = T.alloc_shared((hc_mult, h_block), T.bfloat16)
            xl = T.alloc_fragment((hc_mult, h_block), T.float32)
            T.copy(residual[i, 0, i0_h * h_block], xs, disable_tma=True)
            T.copy(xs, xl)

            ol = T.alloc_fragment(h_block, T.float32)
            T.clear(ol)
            for i_hc in T.serial(hc_mult):
                pre = pre_mix_shared[i_hc]
                for i1_h in T.Parallel(h_block):
                    ol[i1_h] += pre * xl[i_hc, i1_h]

            T.copy(ol, out[i, i0_h * h_block], disable_tma=True)

        T.pdl_trigger()

def mhc_pre(
    residual: torch.Tensor,
    fn: torch.Tensor,
    hc_scale: torch.Tensor,
    hc_base: torch.Tensor,
    rms_eps: float,
    hc_pre_eps: float,
    hc_sinkhorn_eps: float,
    hc_post_mult_value: float,
    sinkhorn_repeat: int,
    n_splits: int = 1,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    """
    Forward pass for mHC pre block.

    Args:
        residual: shape (..., hc_mult, hidden_size), dtype torch.bfloat16
        fn: shape (hc_mult3, hc_mult * hidden_size), dtype torch.float32
        hc_scale: shape (3,), dtype torch.float32
        hc_base: shape (hc_mult3,), dtype torch.float32
        rms_eps: RMS normalization epsilon
        hc_pre_eps: pre-mix epsilon
        hc_sinkhorn_eps: sinkhorn epsilon
        hc_post_mult_value: post-mix multiplier value
        sinkhorn_repeat: number of sinkhorn iterations
        n_splits: split-k factor;

    Returns:
        post_mix: shape (..., hc_mult), dtype torch.float32
        comb_mix: shape (..., hc_mult, hc_mult), dtype torch.float32
        layer_input: shape (..., hidden_size), dtype torch.bfloat16
    """

    # Validate shapes
    assert residual.dtype == torch.bfloat16
    assert fn.dtype == torch.float32
    assert hc_scale.dtype == torch.float32
    assert hc_base.dtype == torch.float32

    hc_mult = residual.shape[-2]
    hidden_size = residual.shape[-1]
    hc_mult2 = hc_mult * hc_mult
    hc_mult3 = hc_mult * 2 + hc_mult2

    hc_hidden_size = hc_mult * hidden_size
    assert fn.shape[0] == hc_mult3
    assert fn.shape[1] == hc_hidden_size
    assert hc_scale.shape == (3,)
    assert hc_base.shape == (hc_mult3,)

    outer_shape = residual.shape[:-2]

    residual_flat = residual.view(-1, hc_mult, hidden_size)
    num_tokens = residual_flat.shape[0]
    fn_flat = fn

    # these number are from deepgemm kernel impl
    block_k = 64
    block_m = 64
    n_splits = compute_num_split(block_k, hc_hidden_size, cdiv(num_tokens, block_m))

    post_mix = torch.empty(
        num_tokens,
        hc_mult,
        dtype=torch.float32,
        device=residual.device,
    )
    comb_mix = torch.empty(
        num_tokens,
        hc_mult2,
        dtype=torch.float32,
        device=residual.device,
    )
    layer_input = torch.empty(
        num_tokens,
        hidden_size,
        dtype=torch.bfloat16,
        device=residual.device,
    )

    gemm_out_mul = torch.empty(
        n_splits,
        num_tokens,
        hc_mult3,
        dtype=torch.float32,
        device=residual.device,
    )
    gemm_out_sqrsum = torch.empty(
        n_splits,
        num_tokens,
        dtype=torch.float32,
        device=residual.device,
    )

    from vllm.utils.deep_gemm import tf32_hc_prenorm_gemm

    tf32_hc_prenorm_gemm(
        residual_flat.view(num_tokens, hc_mult * hidden_size),
        fn_flat,
        gemm_out_mul,
        gemm_out_sqrsum,
        n_splits,
    )

    mhc_pre_big_fuse_tilelang(
        gemm_out_mul,
        gemm_out_sqrsum,
        hc_scale,
        hc_base,
        residual_flat,
        post_mix,
        comb_mix,
        layer_input,
        hidden_size,
        rms_eps,
        hc_pre_eps,
        hc_sinkhorn_eps,
        hc_post_mult_value,
        sinkhorn_repeat,
        n_splits,
        hc_mult,
    )

    post_mix = post_mix.view(*outer_shape, hc_mult, 1)
    comb_mix = comb_mix.view(*outer_shape, hc_mult, hc_mult)
    layer_input = layer_input.view(*outer_shape, hidden_size)

    return post_mix, comb_mix, layer_input

@tilelang.jit(
    pass_configs={
        tilelang.PassConfigKey.TL_DISABLE_WARP_SPECIALIZED: True,
        tilelang.PassConfigKey.TL_DISABLE_TMA_LOWER: True,
        tilelang.PassConfigKey.TL_PTXAS_REGISTER_USAGE_LEVEL: 10,
    },
)
def mhc_pre_big_fuse_tilelang(
    gemm_out_mul,
    gemm_out_sqrsum,
    hc_scale,
    hc_base,
    residual,
    post_mix,
    comb_mix,
    layer_input,
    hidden_size: int,
    rms_eps: float,
    hc_pre_eps: float,
    hc_sinkhorn_eps: float,
    hc_post_mult_value: float,
    sinkhorn_repeat: int,
    n_splits: int = 16,
    hc_mult: int = 4,
):
    """Deeply fused kernels, everything other than gemm & sqrsum in mHC pre block."""
    num_tokens = T.dynamic("num_tokens")
    hc_mult3 = hc_mult * (2 + hc_mult)
    hidden_block = math.gcd(512, hidden_size)

    gemm_out_mul: T.Tensor[[n_splits, num_tokens, hc_mult3], T.float32]  # type: ignore[no-redef, valid-type]
    gemm_out_sqrsum: T.Tensor[[n_splits, num_tokens], T.float32]  # type: ignore[no-redef, valid-type]
    hc_scale: T.Tensor[[3], T.float32]  # type: ignore[no-redef, valid-type]
    hc_base: T.Tensor[[hc_mult3], T.float32]  # type: ignore[no-redef, valid-type]
    residual: T.Tensor[[num_tokens, hc_mult, hidden_size], T.bfloat16]  # type: ignore[no-redef, valid-type]
    # outputs
    post_mix: T.Tensor[[num_tokens, hc_mult], T.float32]  # type: ignore[no-redef, valid-type]
    comb_mix: T.Tensor[[num_tokens, hc_mult * hc_mult], T.float32]  # type: ignore[no-redef, valid-type]
    layer_input: T.Tensor[[num_tokens, hidden_size], T.bfloat16]  # type: ignore[no-redef, valid-type]

    with T.Kernel(num_tokens, threads=96) as i:
        T.pdl_sync()
        ##################################################################
        # _pre_norm_fn_fwd_norm
        rms = T.alloc_fragment(1, T.float32)
        mixes = T.alloc_fragment(hc_mult3, T.float32)
        T.clear(mixes)
        rms[0] = 0
        for i_split in T.serial(n_splits):
            rms[0] += gemm_out_sqrsum[i_split, i]
        rms[0] = T.rsqrt(rms[0] / (hc_mult * hidden_size) + rms_eps)
        for j in T.Parallel(hc_mult3):
            mixes[j] = 0
            for i_split in T.serial(n_splits):
                mixes[j] += gemm_out_mul[i_split, i, j]
            mixes[j] *= rms[0]
        mixes_shared = T.alloc_shared(hc_mult3, T.float32)
        T.copy(mixes, mixes_shared)

        if T.get_thread_binding() < 32:
            ##################################################################
            # _pre_split_mixes_fwd (post & comb)
            cm = T.alloc_fragment((hc_mult, hc_mult), T.float32)
            for j in T.Parallel(hc_mult):
                post_mix[i, j] = (
                    T.sigmoid(
                        mixes_shared[j + hc_mult] * hc_scale[1] + hc_base[j + hc_mult]
                    )
                    * hc_post_mult_value
                )
            for j, k in T.Parallel(hc_mult, hc_mult):
                cm[j, k] = (
                    mixes_shared[j * hc_mult + k + hc_mult * 2] * hc_scale[2]
                    + hc_base[j * hc_mult + k + hc_mult * 2]
                )

            ##################################################################
            # _sinkhorn_fwd
            row_sum = T.alloc_fragment(hc_mult, T.float32)
            col_sum = T.alloc_fragment(hc_mult, T.float32)

            # comb = comb.softmax(-1) + eps
            row_max = T.alloc_fragment(hc_mult, T.float32)
            T.reduce_max(cm, row_max, dim=1)
            for j, k in T.Parallel(hc_mult, hc_mult):
                cm[j, k] = T.exp(cm[j, k] - row_max[j])
            T.reduce_sum(cm, row_sum, dim=1)
            for j, k in T.Parallel(hc_mult, hc_mult):
                cm[j, k] = cm[j, k] / row_sum[j] + hc_sinkhorn_eps

            # comb = comb / (comb.sum(-2) + eps)
            T.reduce_sum(cm, col_sum, dim=0)
            for j, k in T.Parallel(hc_mult, hc_mult):
                cm[j, k] = cm[j, k] / (col_sum[k] + hc_sinkhorn_eps)

            for _ in T.serial(sinkhorn_repeat - 1):
                # comb = comb / (comb.sum(-1) + eps)
                T.reduce_sum(cm, row_sum, dim=1)
                for j, k in T.Parallel(hc_mult, hc_mult):
                    cm[j, k] = cm[j, k] / (row_sum[j] + hc_sinkhorn_eps)

                # comb = comb / (comb.sum(-2) + eps)
                T.reduce_sum(cm, col_sum, dim=0)
                for j, k in T.Parallel(hc_mult, hc_mult):
                    cm[j, k] = cm[j, k] / (col_sum[k] + hc_sinkhorn_eps)

            # save comb_mix to global memory
            for j, k in T.Parallel(hc_mult, hc_mult):
                comb_mix[i, j * hc_mult + k] = cm[j, k]
        else:
            ##################################################################
            # _pre_split_mixes_fwd (pre)
            pre_mix_shared = T.alloc_shared(hc_mult, T.float32)
            for j in T.Parallel(hc_mult):
                pre_mix_shared[j] = (
                    T.sigmoid(
                        mixes_shared[j] * hc_scale[0] + hc_base[j],
                    )
                    + hc_pre_eps
                )
            ###################################################################
            # _pre_apply_mix_fwd
            for i0_h in T.Pipelined(hidden_size // hidden_block, num_stages=2):
                xs = T.alloc_shared((hc_mult, hidden_block), T.float32)
                xl = T.alloc_fragment((hc_mult, hidden_block), T.float32)
                T.copy(residual[i, 0, i0_h * hidden_block], xs)
                T.copy(xs, xl)

                ol = T.alloc_fragment(hidden_block, T.float32)
                T.clear(ol)

                for i_hc in T.serial(hc_mult):
                    pre = pre_mix_shared[i_hc]
                    for i1_h in T.Parallel(hidden_block):
                        ol[i1_h] += pre * xl[i_hc, i1_h]

                T.copy(ol, layer_input[i, i0_h * hidden_block])
        T.pdl_trigger()