liger-kernel-nightly 0.5.5.dev20250402185702__py3-none-any.whl → 0.6.4.dev20260112233432__py3-none-any.whl

liger_kernel/ops/backends/_ascend/ops/geglu.py

@@ -0,0 +1,266 @@
+"""
+UB-aware GEGLU implementation for Ascend NPU.
+
+This implementation automatically adjusts block sizes to fit within UB constraints,
+preventing UB overflow errors when running on Ascend NPU.
+
+It reuses the original kernels when possible, and only uses tiling when necessary.
+"""
+
+import operator
+
+import torch
+import triton
+import triton.language as tl
+
+from liger_kernel.ops.backends._ascend.ub_manager import compute_default_tiling_strategy
+from liger_kernel.ops.utils import calculate_settings
+from liger_kernel.ops.utils import compare_version
+from liger_kernel.ops.utils import ensure_contiguous
+from liger_kernel.utils import is_npu_available
+
+if compare_version("triton", operator.ge, "3.0.0") and not is_npu_available():
+    try:
+        from triton.language.extra.libdevice import tanh
+    except ModuleNotFoundError:
+        from triton.language.extra.cuda.libdevice import tanh
+else:
+    from triton.language.math import tanh
+
+
+@triton.jit
+def _geglu_tanh_forward_kernel_npu(a, b, c, stride, n_cols: tl.constexpr, BLOCK_SIZE: tl.constexpr):
+    """
+    UB-aware GEGLU forward kernel for NPU.
+
+    Uses tiling loop to handle cases where BLOCK_SIZE < n_cols (due to UB constraints).
+    When BLOCK_SIZE >= n_cols, the loop executes only once, maintaining original behavior.
+    """
+    program_id = tl.program_id(0).to(tl.int64)
+
+    # locate start index
+    a += program_id * stride
+    b += program_id * stride
+    c += program_id * stride
+
+    # Process in tiles when BLOCK_SIZE < n_cols
+    for i in range(0, n_cols, BLOCK_SIZE):
+        col_offsets = i + tl.arange(0, BLOCK_SIZE)
+        mask = col_offsets < n_cols
+
+        a_row = tl.load(a + col_offsets, mask=mask, other=0).to(tl.float32)
+        b_row = tl.load(b + col_offsets, mask=mask, other=0)
+
+        # tanh approximation form of GELU is computed with:
+        # 0.5 * a * (1 + tanh(sqrt(2 / pi) * (a + 0.044715 * a^3)))
+        sqrt_2_over_pi = 0.7978845608028654  # sqrt(2 / pi)
+        a_cubed = a_row * a_row * a_row
+        tanh_arg = sqrt_2_over_pi * (a_row + 0.044715 * a_cubed)
+        tanh_result = tanh(tanh_arg)
+        geglu_a = 0.5 * a_row * (1 + tanh_result)
+        c_row = geglu_a.cast(b_row.dtype) * b_row
+
+        tl.store(c + col_offsets, c_row, mask=mask)
+
+
+@triton.jit
+def _geglu_tanh_backward_kernel_npu(dc, a, b, stride, n_cols: tl.constexpr, BLOCK_SIZE: tl.constexpr):
+    """
+    UB-aware GEGLU backward kernel for NPU.
+
+    Uses tiling loop to handle cases where BLOCK_SIZE < n_cols (due to UB constraints).
+    When BLOCK_SIZE >= n_cols, the loop executes only once, maintaining original behavior.
+    """
+    program_id = tl.program_id(0).to(tl.int64)
+
+    # locate start index
+    dc += program_id * stride
+    a += program_id * stride
+    b += program_id * stride
+
+    # Process in tiles when BLOCK_SIZE < n_cols
+    for i in range(0, n_cols, BLOCK_SIZE):
+        col_offsets = i + tl.arange(0, BLOCK_SIZE)
+        mask = col_offsets < n_cols
+
+        dc_row = tl.load(dc + col_offsets, mask=mask, other=0)
+        a_row = tl.load(a + col_offsets, mask=mask, other=0).to(tl.float32)
+        b_row = tl.load(b + col_offsets, mask=mask, other=0)
+
+        # recomputation to save memory
+        sqrt_2_over_pi = 0.7978845608028654  # sqrt(2 / pi)
+        a_cubed = a_row * a_row * a_row
+        tanh_arg = sqrt_2_over_pi * (a_row + 0.044715 * a_cubed)
+        tanh_result = tanh(tanh_arg)
+        geglu_a = 0.5 * a_row * (1 + tanh_result)
+        geglu_a = geglu_a.to(dc_row.dtype).to(tl.float32)
+
+        db_row = dc_row.cast(tl.float32) * geglu_a
+
+        # Gradient w.r.t. a can be computed with:
+        # b * (0.5 * (1 + tanh(z)) + 0.5 * a * (1 - tanh(z)^2) * (sqrt(2/pi) * (1 + 3 * 0.044715 * a^2)))
+        # where z = sqrt(2/pi) * (a + 0.044715 * a^3)
+        term1 = 0.5 * (1 + tanh_result)
+        tanh_sq = tanh_result * tanh_result
+        term2 = 0.5 * a_row * (1 - tanh_sq) * (sqrt_2_over_pi * (1 + 3 * 0.044715 * a_row * a_row))
+        da_row = dc_row * b_row * (term1 + term2)
+
+        tl.store(a + col_offsets, da_row, mask=mask)
+        tl.store(b + col_offsets, db_row.to(dc_row.dtype), mask=mask)
+
+
+def geglu_forward(a, b):
+    """
+    UB-aware GEGLU forward pass for NPU.
+
+    Automatically adjusts block size to fit within UB constraints.
+    """
+    ori_shape = a.shape
+
+    n_cols = ori_shape[-1]
+    a = a.view(-1, n_cols)
+    b = b.view(-1, n_cols)
+    c = torch.empty_like(a)
+    n_rows = a.shape[0]
+
+    # Calculate desired block size
+    desired_block_size, num_warps = calculate_settings(n_cols)
+
+    # Compute tiling strategy based on UB capacity
+    dtype_size = a.element_size()
+    # GEGLU forward tiling strategy:
+    # - Calculates maximum safe block size based on UB capacity
+    # - Memory analysis (only buffers that occupy UB, excluding temporary variables):
+    #   * Inputs: a_row (4 bytes, float32), b_row (dtype_size bytes)
+    #   * Output: c_row (dtype_size bytes)
+    #   * Temporary variables (a_cubed, tanh_arg, tanh_result, geglu_a) are optimized to registers
+    #     and don't occupy UB since they are only used once
+    #   * For float16: a_row(4) + b_row(2) + c_row(2) = 8 bytes/element, ratio = 8/2 = 4.0
+    #   * For float32: a_row(4) + b_row(4) + c_row(4) = 12 bytes/element, ratio = 12/4 = 3.0
+    # - Uses memory_multiplier=4.0 (float16) or 3.0 (float32) * BLOCK_SIZE * dtype_size * 8 bits
+    # - shapes: ((n_cols,),)
+    # - tiling_dims: (0,) means first dimension can be tiled
+    # - Returns: ((block_size,),)
+    shapes = ((n_cols,),)
+    if dtype_size == 2:
+        memory_multiplier = 4.0
+    else:
+        memory_multiplier = 3.0
+    tile_shapes = compute_default_tiling_strategy(
+        safety_margin=0.80,
+        dtype_size=dtype_size,
+        memory_multiplier=memory_multiplier,
+        shapes=shapes,
+        tiling_dims=(0,),
+    )
+
+    if tile_shapes is not None and len(tile_shapes) > 0 and len(tile_shapes[0]) > 0:
+        # Strategy returns ((block_size,),)
+        adjusted_block_size = tile_shapes[0][0]
+    else:
+        # Fallback to desired block size if no best practice found (no tiling needed)
+        adjusted_block_size = desired_block_size
+    # Always use the unified NPU kernel
+    # When adjusted_block_size >= n_cols, the loop executes only once (no tiling)
+    # When adjusted_block_size < n_cols, the loop handles tiling automatically
+    _geglu_tanh_forward_kernel_npu[(n_rows,)](
+        a,
+        b,
+        c,
+        c.stride(-2),
+        n_cols=n_cols,
+        BLOCK_SIZE=adjusted_block_size,
+        num_warps=num_warps,
+    )
+    return a, b, c.view(*ori_shape)
+
+
+def geglu_backward(a, b, dc):
+    """
+    UB-aware GEGLU backward pass for NPU.
+
+    Automatically adjusts block size to fit within UB constraints.
+    """
+    ori_shape = dc.shape
+    n_cols = ori_shape[-1]
+    dc = dc.view(-1, n_cols)
+    n_rows = dc.shape[0]
+
+    # Calculate desired block size
+    desired_block_size, num_warps = calculate_settings(n_cols)
+
+    # Compute tiling strategy based on UB capacity
+    dtype_size = dc.element_size()
+    # GEGLU backward tiling strategy:
+    # - Calculates maximum safe block size based on UB capacity
+    # - Memory analysis: Peak memory usage occurs when executing line 103 (term1 calculation)
+    #   At this point, the following buffers simultaneously occupy UB:
+    #   1. dc_row = tl.load(dc + col_offsets, ...)  # dtype_size bytes
+    #   2. a_row = tl.load(a + col_offsets, ...).to(tl.float32)  # 4 bytes (float32)
+    #   3. b_row = tl.load(b + col_offsets, ...)  # dtype_size bytes
+    #   4. tanh_result = tanh(tanh_arg)  # 4 bytes (float32), used in lines 95, 103, 104
+    #   5. geglu_a = 0.5 * a_row * (1 + tanh_result)  # 4 bytes (float32), used in lines 96, 98
+    #   6. db_row = dc_row.cast(tl.float32) * geglu_a  # 4 bytes (float32, computed at line 98, stored at line 109)
+    #   Note: term1 (line 103) is a temporary variable optimized to registers and doesn't occupy UB
+    #   Temporary variables (a_cubed, tanh_arg, term1, tanh_sq, term2) are optimized to registers
+    #   and don't occupy UB since they are only used once
+    #   * For float16: dc_row(2) + a_row(4) + b_row(2) + tanh_result(4) + geglu_a(4) + db_row(4)
+    #     = 20 bytes/element, ratio = 20/2 = 10.0
+    #   * For float32: dc_row(4) + a_row(4) + b_row(4) + tanh_result(4) + geglu_a(4) + db_row(4)
+    #     = 24 bytes/element, ratio = 24/4 = 6.0
+    # - Uses memory_multiplier=10.0 (float16) or 6.0 (float32) * BLOCK_SIZE * dtype_size * 8 bits
+    # - shapes: ((n_cols,),)
+    # - tiling_dims: (0,) means first dimension can be tiled
+    # - Returns: ((block_size,),)
+    shapes = ((n_cols,),)
+    if dtype_size == 2:
+        memory_multiplier = 10.0
+    else:
+        memory_multiplier = 6.0
+    tile_shapes = compute_default_tiling_strategy(
+        safety_margin=0.80,
+        dtype_size=dtype_size,
+        memory_multiplier=memory_multiplier,
+        shapes=shapes,
+        tiling_dims=(0,),
+    )
+
+    if tile_shapes is not None and len(tile_shapes) > 0 and len(tile_shapes[0]) > 0:
+        # Strategy returns ((block_size,),)
+        adjusted_block_size = tile_shapes[0][0]
+    else:
+        # Fallback to desired block size if no best practice found (no tiling needed)
+        adjusted_block_size = desired_block_size
+
+    # Always use the unified NPU kernel
+    # When adjusted_block_size >= n_cols, the loop executes only once (no tiling)
+    # When adjusted_block_size < n_cols, the loop handles tiling automatically
+    _geglu_tanh_backward_kernel_npu[(n_rows,)](
+        dc,
+        a,
+        b,
+        dc.stride(-2),
+        n_cols=n_cols,
+        BLOCK_SIZE=adjusted_block_size,
+        num_warps=num_warps,
+    )
+
+    return a.view(*ori_shape), b.view(*ori_shape)
+
+
+class LigerGELUMulFunction(torch.autograd.Function):
+    """UB-aware GEGLU function for Ascend NPU."""
+
+    @staticmethod
+    @ensure_contiguous
+    def forward(ctx, a, b):
+        a, b, c = geglu_forward(a, b)
+        ctx.save_for_backward(a, b)
+        return c
+
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, dc):
+        a, b = ctx.saved_tensors
+        a, b = geglu_backward(a, b, dc)
+        return a, b

liger_kernel/ops/backends/_ascend/ops/qwen2vl_mrope.py

@@ -0,0 +1,285 @@
+import torch
+import triton
+import triton.language as tl
+
+from liger_kernel.ops.backends._ascend.ub_manager import compute_default_tiling_strategy
+
+
+@triton.jit
+def _triton_qwen2vl_mrope_npu(
+    q_ptr,
+    q_row_stride,
+    k_ptr,
+    k_row_stride,
+    cos,
+    sin,
+    sl,
+    bs: tl.constexpr,
+    n_qh: tl.constexpr,
+    n_kh: tl.constexpr,
+    hd: tl.constexpr,
+    mrope_section_t: tl.constexpr,
+    mrope_section_h: tl.constexpr,
+    BLOCK_Q: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+    BACKWARD_PASS: tl.constexpr = False,
+):
+    pid = tl.program_id(0).to(tl.int64)
+
+    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
+
+    q_base = q_ptr + pid * q_row_stride
+    k_base = k_ptr + pid * k_row_stride
+
+    d_idx = tl.arange(0, hd // 2)
+    d_mask = d_idx < (hd // 2)
+
+    pos_mask_t = d_idx < t_end
+    pos_mask_h = (d_idx >= t_end) & (d_idx < h_end)
+
+    text_cos_vals = tl.load(t_cos + d_idx, mask=d_mask, other=0)
+    text_sin_vals = tl.load(t_sin + d_idx, mask=d_mask, other=0)
+    height_cos_vals = tl.load(h_cos + d_idx, mask=d_mask, other=0)
+    height_sin_vals = tl.load(h_sin + d_idx, mask=d_mask, other=0)
+    width_cos_vals = tl.load(w_cos + d_idx, mask=d_mask, other=0)
+    width_sin_vals = tl.load(w_sin + d_idx, mask=d_mask, other=0)
+
+    cos_vals = tl.where(pos_mask_t, text_cos_vals, tl.where(pos_mask_h, height_cos_vals, width_cos_vals))
+    sin_vals = tl.where(pos_mask_t, text_sin_vals, tl.where(pos_mask_h, height_sin_vals, width_sin_vals))
+
+    for qh_block in range(0, n_qh, BLOCK_Q):
+        qh_idx = tl.arange(0, BLOCK_Q) + qh_block
+        qh_mask = qh_idx < n_qh
+
+        block_mask = qh_mask[:, None] & d_mask[None, :]
+        offsets = qh_idx[:, None] * hd + d_idx[None, :]
+
+        q_left = tl.load(q_base + offsets, mask=block_mask, other=0)
+        q_right = tl.load(q_base + offsets + (hd // 2), mask=block_mask, other=0)
+
+        if not BACKWARD_PASS:
+            new_left = q_left * cos_vals - q_right * sin_vals
+            new_right = q_right * cos_vals + q_left * sin_vals
+        else:
+            new_left = q_left * cos_vals + q_right * sin_vals
+            new_right = q_right * cos_vals - q_left * sin_vals
+
+        tl.store(q_base + offsets, new_left, mask=block_mask)
+        tl.store(q_base + offsets + (hd // 2), new_right, mask=block_mask)
+
+    for kh_block in range(0, n_kh, BLOCK_K):
+        kh_idx = tl.arange(0, BLOCK_K) + kh_block
+        kh_mask = kh_idx < n_kh
+
+        block_mask = kh_mask[:, None] & d_mask[None, :]
+        offsets = kh_idx[:, None] * hd + d_idx[None, :]
+
+        k_left = tl.load(k_base + offsets, mask=block_mask, other=0)
+        k_right = tl.load(k_base + offsets + (hd // 2), mask=block_mask, other=0)
+
+        if not BACKWARD_PASS:
+            new_left = k_left * cos_vals - k_right * sin_vals
+            new_right = k_right * cos_vals + k_left * sin_vals
+        else:
+            new_left = k_left * cos_vals + k_right * sin_vals
+            new_right = k_right * cos_vals - k_left * sin_vals
+
+        tl.store(k_base + offsets, new_left, mask=block_mask)
+        tl.store(k_base + offsets + (hd // 2), new_right, mask=block_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)
+
+    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()
+
+    # Compute tiling strategy based on UB capacity
+    dtype_size = q.element_size()
+    # MROPE forward tiling strategy:
+    # - cos_vals and sin_vals (include text, height and width) are loaded once outside loops (shared): (pad_hd // 2) * 4 = 2 * pad_hd elements each
+    # - In q heads loop (peak memory):
+    #   * q_left: BLOCK_Q * (pad_hd // 2) elements
+    #   * q_right: BLOCK_Q * (pad_hd // 2) elements
+    #   * new_left: BLOCK_Q * (pad_hd // 2) elements (intermediate result)
+    #   * new_right: BLOCK_Q * (pad_hd // 2) elements (intermediate result)
+    #   * Total: 4 * BLOCK_Q * (pad_hd // 2) = 2 * BLOCK_Q * pad_hd elements
+    # - In k heads loop (peak memory):
+    #   * k_left: BLOCK_K * (pad_hd // 2) elements
+    #   * k_right: BLOCK_K * (pad_hd // 2) elements
+    #   * new_left: BLOCK_K * (pad_hd // 2) elements (intermediate result)
+    #   * new_right: BLOCK_K * (pad_hd // 2) elements (intermediate result)
+    #   * Total: 4 * BLOCK_K * (pad_hd // 2) = 2 * BLOCK_K * pad_hd elements
+    # - Since q and k are processed separately, peak memory is max(BLOCK_Q, BLOCK_K) case
+    # - Plus shared cos/sin: 2 * (pad_hd // 2) = pad_hd elements
+    # - Conservative estimate: (2 * BLOCK_SIZE * pad_hd + pad_hd) * dtype_size * 8 bits
+    # - Simplified: (2 * BLOCK_SIZE + 2) * pad_hd * dtype_size * 8 bits
+    # - For safety, use: memory_multiplier=3.0 * BLOCK_SIZE * pad_hd * dtype_size * 8 bits
+    # - shapes: ((pad_n_q_head, pad_hd), (pad_n_kv_head, pad_hd))
+    # - tiling_dims: (0, 0) means first dimension of each shape can be tiled
+    # - Returns: ((block_size_q, pad_hd), (block_size_kv, pad_hd))
+    shapes = ((pad_n_q_head, pad_hd), (pad_n_kv_head, pad_hd))
+    tile_shapes = compute_default_tiling_strategy(
+        safety_margin=0.90,
+        dtype_size=dtype_size,
+        memory_multiplier=3.0,
+        shapes=shapes,
+        tiling_dims=(0, 0),
+    )
+
+    if tile_shapes is not None and len(tile_shapes) == len(shapes):
+        # Strategy returns ((block_size_q, pad_hd), (block_size_kv, pad_hd))
+        q_tile_shape, k_tile_shape = tile_shapes
+        BLOCK_Q, _ = q_tile_shape
+        BLOCK_K, _ = k_tile_shape
+    else:
+        # Fallback to conservative defaults
+        BLOCK_Q = triton.next_power_of_2(pad_n_q_head)
+        BLOCK_K = triton.next_power_of_2(pad_n_kv_head)
+    _triton_qwen2vl_mrope_npu[(n_row,)](
+        q,
+        q.stride(1),
+        k,
+        k.stride(1),
+        cos,
+        sin,
+        seq_len,
+        batch_size,
+        n_q_head,
+        n_kv_head,
+        head_dim,
+        mrope_section[0],
+        mrope_section[1],
+        BLOCK_Q,
+        BLOCK_K,
+        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)
+
+    n_row = batch_size * seq_len
+
+    # ensure dq and dk are contiguous
+    dq = dq.contiguous()
+    dk = dk.contiguous()
+
+    # Compute tiling strategy based on UB capacity
+    dtype_size = dq.element_size()
+    # MROPE backward tiling strategy:
+    # - cos_vals and sin_vals (include text, height and width) are loaded once outside loops (shared): (pad_hd // 2) * 4 = 2 * pad_hd elements each
+    # - In q heads loop (peak memory):
+    #   * q_left: BLOCK_Q * (pad_hd // 2) elements
+    #   * q_right: BLOCK_Q * (pad_hd // 2) elements
+    #   * new_left: BLOCK_Q * (pad_hd // 2) elements (intermediate result)
+    #   * new_right: BLOCK_Q * (pad_hd // 2) elements (intermediate result)
+    #   * Total: 4 * BLOCK_Q * (pad_hd // 2) = 2 * BLOCK_Q * pad_hd elements
+    # - In k heads loop (peak memory):
+    #   * k_left: BLOCK_K * (pad_hd // 2) elements
+    #   * k_right: BLOCK_K * (pad_hd // 2) elements
+    #   * new_left: BLOCK_K * (pad_hd // 2) elements (intermediate result)
+    #   * new_right: BLOCK_K * (pad_hd // 2) elements (intermediate result)
+    #   * Total: 4 * BLOCK_K * (pad_hd // 2) = 2 * BLOCK_K * pad_hd elements
+    # - Since q and k are processed separately, peak memory is max(BLOCK_Q, BLOCK_K) case
+    # - Plus shared cos/sin: 2 * (pad_hd // 2) = pad_hd elements
+    # - Conservative estimate: (2 * BLOCK_SIZE * pad_hd + pad_hd) * dtype_size * 8 bits
+    # - Simplified: (2 * BLOCK_SIZE + 2) * pad_hd * dtype_size * 8 bits
+    # - For safety, use: memory_multiplier=3.0 * BLOCK_SIZE * pad_hd * dtype_size * 8 bits
+    # - shapes: ((pad_n_q_head, pad_hd), (pad_n_kv_head, pad_hd))
+    # - tiling_dims: (0, 0) means first dimension of each shape can be tiled
+    # - Returns: ((block_size_q, pad_hd), (block_size_kv, pad_hd))
+    shapes = ((pad_n_q_head, pad_hd), (pad_n_kv_head, pad_hd))
+    tile_shapes = compute_default_tiling_strategy(
+        safety_margin=0.90,
+        dtype_size=dtype_size,
+        memory_multiplier=3.0,
+        shapes=shapes,
+        tiling_dims=(0, 0),
+    )
+
+    if tile_shapes is not None and len(tile_shapes) == len(shapes):
+        # Strategy returns ((block_size_q, pad_hd), (block_size_kv, pad_hd))
+        q_tile_shape, k_tile_shape = tile_shapes
+        BLOCK_Q, _ = q_tile_shape
+        BLOCK_K, _ = k_tile_shape
+    else:
+        # Fallback to conservative defaults
+        BLOCK_Q = triton.next_power_of_2(pad_n_q_head)
+        BLOCK_K = triton.next_power_of_2(pad_n_kv_head)
+    _triton_qwen2vl_mrope_npu[(n_row,)](
+        dq,
+        dq.stride(1),
+        dk,
+        dk.stride(1),
+        cos,
+        sin,
+        seq_len,
+        batch_size,
+        n_q_head,
+        n_kv_head,
+        head_dim,
+        mrope_section[0],
+        mrope_section[1],
+        BLOCK_Q,
+        BLOCK_K,
+        BACKWARD_PASS=True,
+    )
+    return dq.transpose(1, 2), dk.transpose(1, 2)
+
+
+class LigerQwen2VLMRopeFunction(torch.autograd.Function):
+    @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