liger-kernel-nightly 0.4.0.dev20241107052928__py3-none-any.whl → 0.6.3.dev20251121010306__py3-none-any.whl

liger_kernel/ops/poly_norm.py

@@ -0,0 +1,386 @@
+import operator
+
+import torch
+import triton
+import triton.language as tl
+
+from liger_kernel.ops.utils import calculate_settings
+from liger_kernel.ops.utils import compare_version
+from liger_kernel.ops.utils import ensure_contiguous
+
+if compare_version("triton", operator.ge, "3.0.0"):
+    try:
+        from triton.language.extra.libdevice import rsqrt
+    except ModuleNotFoundError:
+        from triton.language.extra.cuda.libdevice import rsqrt
+else:
+    from triton.language.math import rsqrt
+
+
+@triton.jit
+def _poly_norm_forward_kernel(
+    Y_ptr,
+    Y_row_stride,
+    X_ptr,
+    X_row_stride,
+    W_ptr,  # weight: [3] for [w0, w1, w2]
+    B_ptr,  # bias: scalar
+    RSTD_ptr,  # cache rstd for backward: shape (n_rows, 3)
+    RSTD_row_stride,
+    n_cols,
+    eps,
+    BLOCK_SIZE: tl.constexpr,
+):
+    """
+    PolyNorm formula:
+        y = w₀·norm(x³) + w₁·norm(x²) + w₂·norm(x) + b
+        where norm(u) = u / sqrt(mean(u²) + ε)
+
+    Reference:
+    1. https://github.com/BryceZhuo/PolyCom/
+    2. https://arxiv.org/pdf/2411.03884
+
+    Cache rstd values for backward pass
+    """
+    row_idx = tl.program_id(0).to(tl.int64)
+    col_offsets = tl.arange(0, BLOCK_SIZE)
+    mask = col_offsets < n_cols
+
+    # Load pointers
+    Y_ptr += row_idx * Y_row_stride
+    X_ptr += row_idx * X_row_stride
+    RSTD_ptr += row_idx * RSTD_row_stride
+
+    # Load input row
+    X_row = tl.load(X_ptr + col_offsets, mask=mask, other=0.0)
+
+    # Load weights and bias
+    w0 = tl.load(W_ptr + 0)
+    w1 = tl.load(W_ptr + 1)
+    w2 = tl.load(W_ptr + 2)
+    b = tl.load(B_ptr)
+
+    # Compute x³, x², x
+    X_pow3 = X_row * X_row * X_row
+    X_pow2 = X_row * X_row
+    X_pow1 = X_row
+
+    # Compute norm(x³): norm(u) = u * rsqrt(mean(u²) + eps)
+    mean_square_3 = tl.sum(X_pow3 * X_pow3, axis=0) / n_cols
+    rstd_3 = rsqrt(mean_square_3 + eps)
+    norm_x3 = X_pow3 * rstd_3
+
+    # Compute norm(x²)
+    mean_square_2 = tl.sum(X_pow2 * X_pow2, axis=0) / n_cols
+    rstd_2 = rsqrt(mean_square_2 + eps)
+    norm_x2 = X_pow2 * rstd_2
+
+    # Compute norm(x)
+    mean_square_1 = tl.sum(X_pow1 * X_pow1, axis=0) / n_cols
+    rstd_1 = rsqrt(mean_square_1 + eps)
+    norm_x1 = X_pow1 * rstd_1
+
+    # Cache rstd values for backward
+    tl.store(RSTD_ptr + 0, rstd_3)
+    tl.store(RSTD_ptr + 1, rstd_2)
+    tl.store(RSTD_ptr + 2, rstd_1)
+
+    # Compute output: y = w₀·norm(x³) + w₁·norm(x²) + w₂·norm(x) + b
+    Y_row = w0 * norm_x3 + w1 * norm_x2 + w2 * norm_x1 + b
+
+    # Store output
+    tl.store(Y_ptr + col_offsets, Y_row, mask=mask)
+
+
+@triton.jit
+def _poly_norm_backward_kernel(
+    dY_ptr,
+    dY_row_stride,
+    dX_ptr,
+    dX_row_stride,
+    X_ptr,
+    X_row_stride,
+    W_ptr,
+    RSTD_ptr,
+    RSTD_row_stride,
+    dW_ptr,  # shape: (n_programs, 3)
+    dW_row_stride,
+    dB_ptr,  # shape: (n_programs,)
+    n_rows,
+    n_cols,
+    rows_per_program: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+):
+    """
+    PolyNorm Backward Kernel Gradient:
+        ∂L/∂x_i = Σ_p w_p * [p*x_i^(p-1) * grad_i/D_p - (p/d)*x_i^(2p-1) * S_p/(D_p³)]
+
+    where:
+        - D_p = RMS(x^p) = 1/rstd_p
+        - S_p = sum(grad * x^p) over the row
+        - d = n_cols
+        - p ∈ {3, 2, 1}
+    """
+    row_block_id = tl.program_id(0).to(tl.int64)
+    row_start = row_block_id * rows_per_program
+    row_end = min((row_block_id + 1) * rows_per_program, n_rows)
+    col_offsets = tl.arange(0, BLOCK_SIZE)
+    mask = col_offsets < n_cols
+
+    # Initialize accumulators for weight and bias gradients (scalars)
+    dW0_acc = 0.0
+    dW1_acc = 0.0
+    dW2_acc = 0.0
+    dB_acc = 0.0
+
+    # Load weights
+    w0 = tl.load(W_ptr + 0).to(tl.float32)
+    w1 = tl.load(W_ptr + 1).to(tl.float32)
+    w2 = tl.load(W_ptr + 2).to(tl.float32)
+
+    dY_ptr += row_start * dY_row_stride
+    dX_ptr += row_start * dX_row_stride
+    X_ptr += row_start * X_row_stride
+    RSTD_ptr += row_start * RSTD_row_stride
+
+    for _ in range(row_start, row_end):
+        # Load input and gradient
+        dY_row = tl.load(dY_ptr + col_offsets, mask=mask, other=0.0).to(tl.float32)
+        X_row = tl.load(X_ptr + col_offsets, mask=mask, other=0.0).to(tl.float32)
+
+        # Load cached rstd values
+        rstd_3 = tl.load(RSTD_ptr + 0).to(tl.float32)
+        rstd_2 = tl.load(RSTD_ptr + 1).to(tl.float32)
+        rstd_1 = tl.load(RSTD_ptr + 2).to(tl.float32)
+
+        # Compute powers
+        X_pow3 = X_row * X_row * X_row
+        X_pow2 = X_row * X_row
+        X_pow1 = X_row
+
+        # Accumulate bias gradient: dB = sum(dY)
+        dB_acc += tl.sum(dY_row, axis=0)
+
+        # Compute gradient w.r.t. input using closed-form formula
+        # For p=3: ∂L/∂x from w0 * norm(x³)
+        S_3 = tl.sum(dY_row * X_pow3, axis=0)  # scalar
+        grad_x_3 = w0 * (
+            3.0 * X_pow2 * rstd_3 * dY_row
+            - (3.0 / n_cols) * X_row * X_row * X_row * X_row * X_row * (rstd_3 * rstd_3 * rstd_3) * S_3
+        )
+
+        # For p=2: ∂L/∂x from w1 * norm(x²)
+        S_2 = tl.sum(dY_row * X_pow2, axis=0)  # scalar
+        grad_x_2 = w1 * (
+            2.0 * X_row * rstd_2 * dY_row - (2.0 / n_cols) * X_row * X_row * X_row * (rstd_2 * rstd_2 * rstd_2) * S_2
+        )
+
+        # For p=1: ∂L/∂x from w2 * norm(x)
+        S_1 = tl.sum(dY_row * X_pow1, axis=0)  # scalar
+        grad_x_1 = w2 * (1.0 * rstd_1 * dY_row - (1.0 / n_cols) * X_row * (rstd_1 * rstd_1 * rstd_1) * S_1)
+
+        # Accumulate weight gradients using closed-form: dW_p = rstd_p * S_p
+        dW0_acc += rstd_3 * S_3
+        dW1_acc += rstd_2 * S_2
+        dW2_acc += rstd_1 * S_1
+
+        # Total gradient
+        dX_row = grad_x_3 + grad_x_2 + grad_x_1
+
+        # Store gradient
+        tl.store(dX_ptr + col_offsets, dX_row, mask=mask)
+
+        # Update pointers
+        dY_ptr += dY_row_stride
+        dX_ptr += dX_row_stride
+        X_ptr += X_row_stride
+        RSTD_ptr += RSTD_row_stride
+
+    # Store accumulated gradients (scalars)
+    tl.store(dW_ptr + row_block_id * dW_row_stride + 0, dW0_acc)
+    tl.store(dW_ptr + row_block_id * dW_row_stride + 1, dW1_acc)
+    tl.store(dW_ptr + row_block_id * dW_row_stride + 2, dW2_acc)
+    tl.store(dB_ptr + row_block_id, dB_acc)
+
+
+def poly_norm_forward(X, W, B, eps=1e-6):
+    """
+    PolyNorm Forward Pass
+
+    Args:
+        X: input tensor of shape (*, H) where H is hidden dimension
+        W: weight tensor of shape (3,) for [w0, w1, w2]
+        B: bias scalar tensor
+        eps: epsilon for numerical stability
+
+    Returns:
+        Y: output tensor of same shape as X
+        X: reshaped input (for backward)
+        RSTD: cached rstd values (for backward)
+        BLOCK_SIZE: block size used
+        num_warps: number of warps used
+    """
+    shape = X.shape
+    dim = shape[-1]
+    X = X.view(-1, dim)
+    n_rows, n_cols = X.shape
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+
+    # RSTD is to cache rstd for each row
+    Y = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
+    RSTD = torch.empty((n_rows, 3), dtype=torch.float32, device=X.device)
+
+    # Check constraints
+    assert W.shape[0] == 3, "Weight tensor must have shape (3,)"
+    assert B.numel() == 1, "Bias must be a scalar"
+
+    # XPU-specific optimization
+    kernel_args = {}
+    if X.device.type == "xpu":
+        kernel_args["grf_mode"] = "large"
+
+    # Launch kernel
+    _poly_norm_forward_kernel[(n_rows,)](
+        Y,
+        Y.stride(0),
+        X,
+        X.stride(0),
+        W,
+        B,
+        RSTD,
+        RSTD.stride(0),
+        n_cols,
+        eps,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+        **kernel_args,
+    )
+
+    return Y.view(*shape), X, RSTD, BLOCK_SIZE, num_warps
+
+
+def poly_norm_backward(dY, X, W, RSTD, BLOCK_SIZE, num_warps, in_place):
+    """
+    PolyNorm Backward Pass
+
+    Args:
+        dY: gradient of output
+        X: input tensor (already reshaped to 2D)
+        W: weight tensor
+        RSTD: cached rstd values from forward
+        BLOCK_SIZE: block size from forward
+        num_warps: number of warps from forward
+        in_place: whether to in-place modify dY to store dX (saves memory)
+
+    Returns:
+        dX: gradient w.r.t. input
+        dW: gradient w.r.t. weight
+        dB: gradient w.r.t. bias
+    """
+    shape = dY.shape
+    dim = shape[-1]
+    dY = dY.view(-1, dim)
+    n_rows, n_cols = dY.shape
+
+    # Get number of SMs for parallelization
+    import math
+
+    sm_count = 1
+    if X.device.type == "cuda":
+        sm_count = torch.cuda.get_device_properties(X.device).multi_processor_count
+    elif X.device.type == "xpu":
+        sm_count = torch.xpu.get_device_properties(X.device).gpu_eu_count
+
+    # Allocate or reuse gradients
+    if in_place is True:
+        dX = dY
+    else:
+        dX = torch.zeros_like(dY)
+
+    _dW = torch.empty((sm_count, 3), dtype=torch.float32, device=W.device)
+    _dB = torch.empty((sm_count,), dtype=torch.float32, device=W.device)
+
+    rows_per_program = math.ceil(n_rows / sm_count)
+    grid = (sm_count,)
+
+    # XPU-specific optimization
+    kernel_args = {}
+    if X.device.type == "xpu":
+        kernel_args["grf_mode"] = "large"
+
+    # Launch backward kernel
+    _poly_norm_backward_kernel[grid](
+        dY,
+        dY.stride(0),
+        dX,
+        dX.stride(0),
+        X,
+        X.stride(0),
+        W,
+        RSTD,
+        RSTD.stride(0),
+        _dW,
+        _dW.stride(0),
+        _dB,
+        n_rows,
+        n_cols,
+        rows_per_program,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+        **kernel_args,
+    )
+
+    # Reduce gradients across SMs
+    dX = dX.view(*shape)
+    dW = _dW.sum(dim=0).to(W.dtype)
+    dB = _dB.sum().to(W.dtype)
+
+    return dX, dW, dB
+
+
+class LigerPolyNormFunction(torch.autograd.Function):
+    """
+    PolyNorm Function with forward and backward pass
+
+    PolyNorm formula:
+        y = w₀·norm(x³) + w₁·norm(x²) + w₂·norm(x) + b
+        where norm(u) = u / sqrt(mean(u²) + ε)
+
+    Backward uses closed-form gradient:
+        ∂L/∂x_i = Σ_p w_p * [p*x_i^(p-1) * grad_i/D_p - (p/d)*x_i^(2p-1) * S_p/(D_p³)]
+    """
+
+    @staticmethod
+    @ensure_contiguous
+    def forward(ctx, X, W, B, eps=1e-6, in_place=True):
+        """
+        Args:
+            X: input tensor of shape (B, T, H) or (BxT, H)
+            W: weight tensor of shape (3,) for [w0, w1, w2]
+            B: bias scalar
+            eps: epsilon for numerical stability
+            in_place: whether to in-place modify grad_output in backward (saves memory)
+
+        Returns:
+            Y: output tensor of same shape as X
+        """
+        Y, X, RSTD, BLOCK_SIZE, num_warps = poly_norm_forward(X, W, B, eps)
+        ctx.BLOCK_SIZE = BLOCK_SIZE
+        ctx.num_warps = num_warps
+        ctx.in_place = in_place
+        ctx.save_for_backward(X, W, RSTD)
+        return Y
+
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, grad_output):
+        """
+        Args:
+            grad_output: gradient of output
+
+        Returns:
+            dX, dW, dB: gradients w.r.t. X, W, B
+        """
+        X, W, RSTD = ctx.saved_tensors
+        dX, dW, dB = poly_norm_backward(grad_output, X, W, RSTD, ctx.BLOCK_SIZE, ctx.num_warps, ctx.in_place)
+        return dX, dW, dB, None, None

liger_kernel/ops/qwen2vl_mrope.py

@@ -0,0 +1,222 @@
+import torch
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def _triton_qwen2vl_mrope(
+    q_ptr,
+    k_ptr,
+    cos,
+    sin,
+    sl,
+    bs: tl.constexpr,
+    n_qh: tl.constexpr,
+    n_kh: tl.constexpr,
+    hd: tl.constexpr,
+    pad_n_qh: tl.constexpr,
+    pad_n_kh: tl.constexpr,
+    pad_hd: tl.constexpr,
+    mrope_section_t: tl.constexpr,
+    mrope_section_h: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    BACKWARD_PASS: tl.constexpr = False,
+):
+    pid = tl.program_id(0)
+
+    # locate start address
+    q_ptr = q_ptr + pid * (n_qh * hd)
+    k_ptr = k_ptr + pid * (n_kh * hd)
+
+    # ####################################################################
+    # get the cos(mθ_{i...d/2}) and sin(mθ_{i...d/2}) for token position
+    # m of this program instance
+    # ####################################################################
+
+    # 1. program instances are laid out in a 1D vector of size bsz * seq_len, which
+    # effectively represents a 2D grid of size [bsz, seq_len] with seq_len dimension
+    # being the fastest changing dimension. Thus we can simply do pid // sl to get the batch index
+    # and pid % sl to get the sequence index.
+    # 2. We only need the left half of cos and sin matrix because the right half is just
+    # a clone of the left half.
+    t_end = mrope_section_t
+    h_end = t_end + mrope_section_h
+
+    t_cos = cos + pid * hd
+    h_cos = t_cos + bs * sl * hd
+    w_cos = h_cos + bs * sl * hd
+    t_sin = sin + pid * hd
+    h_sin = t_sin + bs * sl * hd
+    w_sin = h_sin + bs * sl * hd
+
+    cos_offsets = tl.arange(0, pad_hd // 2)
+    t_mask = cos_offsets < t_end
+    h_mask = (t_end <= cos_offsets) & (cos_offsets < h_end)
+    w_mask = (h_end <= cos_offsets) & (cos_offsets < hd // 2)
+    t_cos_row = tl.load(t_cos + cos_offsets, mask=t_mask, other=0)
+    h_cos_row = tl.load(h_cos + cos_offsets, mask=h_mask, other=0)
+    w_cos_row = tl.load(w_cos + cos_offsets, mask=w_mask, other=0)
+    t_sin_row = tl.load(t_sin + cos_offsets, mask=t_mask, other=0)
+    h_sin_row = tl.load(h_sin + cos_offsets, mask=h_mask, other=0)
+    w_sin_row = tl.load(w_sin + cos_offsets, mask=w_mask, other=0)
+    cos_row = t_cos_row + h_cos_row + w_cos_row
+    sin_row = t_sin_row + h_sin_row + w_sin_row
+
+    # ####################################################################
+    # Load the left and right half of q and k for the current
+    # program instance (i.e. for the current token) separately
+    # ####################################################################
+    # left half of the head
+    first_half_q_offsets = tl.arange(0, pad_n_qh)[:, None] * hd + tl.arange(0, pad_hd // 2)[None, :]
+    first_half_k_offsets = tl.arange(0, pad_n_kh)[:, None] * hd + tl.arange(0, pad_hd // 2)[None, :]
+    first_q_mask = (tl.arange(0, pad_n_qh)[:, None] < n_qh) & (tl.arange(0, pad_hd // 2)[None, :] < hd // 2)
+    first_k_mask = (tl.arange(0, pad_n_kh)[:, None] < n_kh) & (tl.arange(0, pad_hd // 2)[None, :] < hd // 2)
+    q_tile_1 = tl.load(q_ptr + first_half_q_offsets, mask=first_q_mask, other=0).to(sin_row.dtype)
+    k_tile_1 = tl.load(k_ptr + first_half_k_offsets, mask=first_k_mask, other=0).to(sin_row.dtype)
+
+    # right half of the head
+    second_half_q_offsets = first_half_q_offsets + (hd // 2)
+    second_half_k_offsets = first_half_k_offsets + (hd // 2)
+    second_q_mask = first_q_mask
+    second_k_mask = first_k_mask
+    q_tile_2 = tl.load(q_ptr + second_half_q_offsets, mask=second_q_mask, other=0).to(sin_row.dtype)
+    k_tile_2 = tl.load(k_ptr + second_half_k_offsets, mask=second_k_mask, other=0).to(sin_row.dtype)
+
+    if not BACKWARD_PASS:
+        # y = [x1, x2] * [cos, cos] + [-x2, x1] * [sin, sin]
+        new_q_tile_1 = q_tile_1 * cos_row - q_tile_2 * sin_row
+        tl.store(q_ptr + first_half_q_offsets, new_q_tile_1, mask=first_q_mask)
+        new_q_tile_2 = q_tile_2 * cos_row + q_tile_1 * sin_row
+        tl.store(q_ptr + second_half_q_offsets, new_q_tile_2, mask=second_q_mask)
+
+        new_k_tile_1 = k_tile_1 * cos_row - k_tile_2 * sin_row
+        tl.store(k_ptr + first_half_k_offsets, new_k_tile_1, mask=first_k_mask)
+        new_k_tile_2 = k_tile_2 * cos_row + k_tile_1 * sin_row
+        tl.store(k_ptr + second_half_k_offsets, new_k_tile_2, mask=second_k_mask)
+    else:
+        # with some math, we can get:
+        # dy = [dx1, dx2] * [cos, cos] + [-dx2, dx1] * [-sin, -sin]
+        new_q_tile_1 = q_tile_1 * cos_row + q_tile_2 * sin_row
+        tl.store(q_ptr + first_half_q_offsets, new_q_tile_1, mask=first_q_mask)
+        new_q_tile_2 = q_tile_2 * cos_row - q_tile_1 * sin_row
+        tl.store(q_ptr + second_half_q_offsets, new_q_tile_2, mask=second_q_mask)
+
+        new_k_tile_1 = k_tile_1 * cos_row + k_tile_2 * sin_row
+        tl.store(k_ptr + first_half_k_offsets, new_k_tile_1, mask=first_k_mask)
+        new_k_tile_2 = k_tile_2 * cos_row - k_tile_1 * sin_row
+        tl.store(k_ptr + second_half_k_offsets, new_k_tile_2, mask=second_k_mask)
+
+
+def qwen2vl_mrope_forward(q, k, cos, sin, mrope_section):
+    # transpose it back to the physical shape because Triton looks at the physical storage
+    # note: q and k are incontiguous before the transformation and will become contiguous after transpose
+    q = q.transpose(1, 2)
+    k = k.transpose(1, 2)
+
+    batch_size, seq_len, n_q_head, head_dim = q.shape
+    n_kv_head = k.shape[2]
+    pad_hd = triton.next_power_of_2(head_dim)
+    pad_n_q_head = triton.next_power_of_2(n_q_head)
+    pad_n_kv_head = triton.next_power_of_2(n_kv_head)
+    BLOCK_SIZE = max(pad_n_q_head, pad_n_kv_head)
+
+    n_row = batch_size * seq_len
+
+    # ensure tensors passed into the kernel are contiguous. It will be no-op if they are already contiguous
+    q = q.contiguous()
+    k = k.contiguous()
+    cos = cos.contiguous()
+    sin = sin.contiguous()
+
+    _triton_qwen2vl_mrope[(n_row,)](
+        q,
+        k,
+        cos,
+        sin,
+        seq_len,
+        batch_size,
+        n_q_head,
+        n_kv_head,
+        head_dim,
+        pad_n_q_head,
+        pad_n_kv_head,
+        pad_hd,
+        mrope_section[0],
+        mrope_section[1],
+        BLOCK_SIZE=BLOCK_SIZE,
+        BACKWARD_PASS=False,
+    )
+    return q.transpose(1, 2), k.transpose(1, 2), cos, sin
+
+
+def qwen2vl_mrope_backward(dq, dk, cos, sin, mrope_section):
+    dq = dq.transpose(1, 2)
+    dk = dk.transpose(1, 2)
+
+    batch_size, seq_len, n_q_head, head_dim = dq.shape
+    n_kv_head = dk.shape[2]
+    pad_hd = triton.next_power_of_2(head_dim)
+    pad_n_q_head = triton.next_power_of_2(n_q_head)
+    pad_n_kv_head = triton.next_power_of_2(n_kv_head)
+    BLOCK_SIZE = max(pad_n_q_head, pad_n_kv_head)
+
+    n_row = batch_size * seq_len
+
+    # ensure dq and dk are contiguous
+    dq = dq.contiguous()
+    dk = dk.contiguous()
+
+    # backward is similar to forward except swapping few ops
+    _triton_qwen2vl_mrope[(n_row,)](
+        dq,
+        dk,
+        cos,
+        sin,
+        seq_len,
+        batch_size,
+        n_q_head,
+        n_kv_head,
+        head_dim,
+        pad_n_q_head,
+        pad_n_kv_head,
+        pad_hd,
+        mrope_section[0],
+        mrope_section[1],
+        BLOCK_SIZE=BLOCK_SIZE,
+        BACKWARD_PASS=True,
+    )
+    return dq.transpose(1, 2), dk.transpose(1, 2)
+
+
+class LigerQwen2VLMRopeFunction(torch.autograd.Function):
+    """
+    Triton implementation of the Qwen2VL Multimodal Rotary Positional Embedding (M-RoPE) operation.
+
+    Please find the corresponding HuggingFace implementation here:
+    https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py
+    """
+
+    @staticmethod
+    def forward(ctx, q, k, cos, sin, mrope_section, unsqueeze_dim=1):
+        """
+        q size: (bsz, n_q_head, seq_len, head_dim)
+        k size: (bsz, n_kv_head, seq_len, head_dim)
+        cos size: (3, bsz, seq_len, head_dim)
+        sin size: (3, bsz, seq_len, head_dim)
+        """
+        q, k, cos, sin = qwen2vl_mrope_forward(q, k, cos, sin, mrope_section)
+        ctx.save_for_backward(cos, sin)
+        ctx.mrope_section = mrope_section
+        return q, k
+
+    def backward(ctx, dq, dk):
+        """
+        dq size: (bsz, n_q_head, seq_len, head_dim)
+        dk size: (bsz, n_kv_head, seq_len, head_dim)
+        cos size: (3, bsz, seq_len, head_dim)
+        sin size: (3, bsz, seq_len, head_dim)
+        """
+        cos, sin = ctx.saved_tensors
+        mrope_section = ctx.mrope_section
+        dq, dk = qwen2vl_mrope_backward(dq, dk, cos, sin, mrope_section)
+        return dq, dk, None, None, None, None