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

liger_kernel/ops/rms_norm.py

@@ -21,8 +21,10 @@ 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.ops.utils import torch_to_triton_dtype
+from liger_kernel.utils import get_npu_multi_processor_count
+from liger_kernel.utils import is_npu_available
 
-if compare_version("triton", operator.ge, "3.0.0"):
+if compare_version("triton", operator.ge, "3.0.0") and not is_npu_available():
     try:
         # typical import path with dispatch available
         from triton.language.extra.libdevice import rsqrt
@@ -52,6 +54,7 @@ def _rms_norm_forward_kernel(
     eps,
     offset,
     casting_mode: tl.constexpr,  # constexpr so the `if` blocks can be optimized out
+    elementwise_affine: tl.constexpr,
     BLOCK_SIZE: tl.constexpr,
 ):
     """
@@ -63,17 +66,18 @@ def _rms_norm_forward_kernel(
     3. https://arxiv.org/pdf/1910.07467
     """
 
-    row_idx = tl.program_id(0)
+    row_idx = tl.program_id(0).to(tl.int64)
     col_offsets = tl.arange(0, BLOCK_SIZE)
     mask = col_offsets < n_cols
 
-    Y_ptr += row_idx * Y_row_stride
-    X_ptr += row_idx * X_row_stride
-    RSTD_ptr += row_idx * RSTD_row_stride
+    y_base = Y_ptr + row_idx * Y_row_stride
+    x_base = X_ptr + row_idx * X_row_stride
+    rstd_base = RSTD_ptr + row_idx * RSTD_row_stride
 
-    X_row = tl.load(X_ptr + col_offsets, mask=mask, other=0)
+    X_row = tl.load(x_base + col_offsets, mask=mask, other=0)
     X_row_dtype = X_row.dtype
-    W_row = tl.load(W_ptr + col_offsets, mask=mask, other=0)
+    if elementwise_affine:
+        W_row = tl.load(W_ptr + col_offsets, mask=mask, other=0)
 
     # On Llama, only rstd is computed on fp32
     if casting_mode == _CASTING_MODE_LLAMA:
@@ -81,7 +85,8 @@ def _rms_norm_forward_kernel(
 
     # Gemma computes everything on fp32, and then casts back the output to the original dtype
     if casting_mode == _CASTING_MODE_GEMMA:
-        W_row = W_row.to(tl.float32)
+        if elementwise_affine:
+            W_row = W_row.to(tl.float32)
         X_row = X_row.to(tl.float32)
 
     if casting_mode == _CASTING_MODE_NONE:
@@ -94,7 +99,7 @@ def _rms_norm_forward_kernel(
     # We can save time by caching rms with minimal memory overhead
     # because rms is much smaller compared to X_row, as rms is for each row.
     # However, on the computation side, it can save 4 operations (*, sum, /, sqrt).
-    tl.store(RSTD_ptr, rstd)
+    tl.store(rstd_base, rstd)
 
     X_row = X_row * rstd
 
@@ -102,12 +107,15 @@ def _rms_norm_forward_kernel(
     if casting_mode == _CASTING_MODE_LLAMA:
         X_row = X_row.to(X_row_dtype)
 
-    Y_row = X_row * (offset + W_row)
+    if elementwise_affine:
+        Y_row = X_row * (offset + W_row)
+    else:
+        Y_row = X_row
 
     if casting_mode == _CASTING_MODE_GEMMA:
         Y_row = Y_row.to(X_row_dtype)
 
-    tl.store(Y_ptr + col_offsets, Y_row, mask=mask)
+    tl.store(y_base + col_offsets, Y_row, mask=mask)
 
 
 @triton.jit
@@ -128,8 +136,9 @@ def _rms_norm_backward_kernel(
     n_rows,
     n_cols,
     offset,
-    rows_per_program: tl.constexpr,
+    rows_per_program,
     casting_mode: tl.constexpr,
+    elementwise_affine: tl.constexpr,
     BLOCK_SIZE: tl.constexpr,
 ):
     """
@@ -137,61 +146,256 @@ def _rms_norm_backward_kernel(
     dw = sum(dy * (x / RMS)). summation over BxT dimension
     """
 
-    row_block_id = tl.program_id(0)
+    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
 
-    dW_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+    if elementwise_affine:
+        dW_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
 
-    dY_ptr += row_start * dY_row_stride
-    dX_ptr += row_start * dX_row_stride
+    if elementwise_affine:
+        W_row = tl.load(W_ptr + col_offsets, mask=mask, other=0.0)
+        W_row = W_row + offset
 
-    X_ptr += row_start * X_row_stride
-    RSTD_ptr += row_start
+    for row_idx in range(row_start, row_end):
+        dy_base = dY_ptr + row_idx * dY_row_stride
+        dx_base = dX_ptr + row_idx * dX_row_stride
 
-    W_row = tl.load(W_ptr + col_offsets, mask=mask, other=0.0)
-    W_row = W_row + offset
+        x_base = X_ptr + row_idx * X_row_stride
+        rstd_base = RSTD_ptr + row_idx * RSTD_row_stride
 
-    for _ in range(row_start, row_end):
-        dY_row = tl.load(dY_ptr + col_offsets, mask=mask, other=0.0)
-        X_row = tl.load(X_ptr + col_offsets, mask=mask, other=0.0)
+        dY_row = tl.load(dy_base + col_offsets, mask=mask, other=0.0)
+        X_row = tl.load(x_base + col_offsets, mask=mask, other=0.0)
 
         # Get cached rms
-        rstd_row = tl.load(RSTD_ptr)
+        rstd_row = tl.load(rstd_base)
 
         X_row = X_row.to(tl.float32)
 
         # Different bacward graphs for different casting modes
         if casting_mode == _CASTING_MODE_LLAMA:
-            m = (dY_row * W_row).to(tl.float32)
+            if elementwise_affine:
+                m = (dY_row * W_row).to(tl.float32)
+            else:
+                m = dY_row.to(tl.float32)
 
         elif casting_mode == _CASTING_MODE_GEMMA:
             dY_row = dY_row.to(tl.float32)
-            m = dY_row * W_row
+            if elementwise_affine:
+                m = dY_row * W_row
+            else:
+                m = dY_row
         else:
-            m = dY_row * W_row
+            if elementwise_affine:
+                m = dY_row * W_row
+            else:
+                m = dY_row
 
         dX_row = rstd_row * m
 
         dX_row += (rstd_row) * (-(1 / n_cols) * rstd_row * rstd_row * tl.sum(m * X_row, axis=0) * X_row)
 
-        # calculate the gradient of W
+        if elementwise_affine:
+            # calculate the gradient of W
+            if casting_mode == _CASTING_MODE_LLAMA:
+                dW_row += dY_row * (X_row * rstd_row).to(X_dtype)
+            else:
+                # here X_row is already in fp32 (see previous if block)
+                dW_row += dY_row * (X_row * rstd_row)
+
+        tl.store(dx_base + col_offsets, dX_row.to(X_dtype), mask=mask)
+
+    if elementwise_affine:
+        tl.store(dW_ptr + row_block_id * dW_row_stride + col_offsets, dW_row, mask=mask)
+
+
+@triton.jit
+def _block_rms_norm_forward_kernel(
+    Y_ptr,
+    Y_row_stride,
+    X_ptr,
+    X_row_stride,
+    W_ptr,
+    W_row_stride,
+    RSTD_ptr,
+    RSTD_row_stride,
+    n_rows,
+    n_cols,
+    eps,
+    offset,
+    casting_mode: tl.constexpr,  # constexpr so the `if` blocks can be optimized out
+    elementwise_affine: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    BLOCK_ROW: tl.constexpr,
+):
+    """
+    y_i = (x_i / (RMS)) * (offset + wi), RMS = sqrt(sum(x_i^2) / N)
+
+    Reference:
+    1. https://triton-lang.org/main/getting-started/tutorials/05-layer-norm.html
+    2. https://github.com/unslothai/unsloth/blob/fd753fed99ed5f10ef8a9b7139588d9de9ddecfb/unsloth/kernels/rms_layernorm.py#L22
+    3. https://arxiv.org/pdf/1910.07467
+    """
+
+    row_idx = tl.program_id(0) * BLOCK_ROW + tl.arange(0, BLOCK_ROW)
+    col_offsets = tl.arange(0, BLOCK_SIZE)
+    row_mask = row_idx < n_rows
+    col_mask = col_offsets < n_cols
+
+    X_row = tl.load(
+        X_ptr + row_idx[:, None] * X_row_stride + col_offsets[None, :],
+        mask=row_mask[:, None] & col_mask[None, :],
+        other=0,
+    )
+    X_row_dtype = X_row.dtype
+    if elementwise_affine:
+        W_row = tl.load(W_ptr + col_offsets, mask=col_mask, other=0)
+
+    # On Llama, only rstd is computed on fp32
+    if casting_mode == _CASTING_MODE_LLAMA:
+        X_row = X_row.to(tl.float32)
+
+    # Gemma computes everything on fp32, and then casts back the output to the original dtype
+    if casting_mode == _CASTING_MODE_GEMMA:
+        if elementwise_affine:
+            W_row = W_row.to(tl.float32)
+        X_row = X_row.to(tl.float32)
+
+    if casting_mode == _CASTING_MODE_NONE:
+        eps = eps.to(X_row_dtype)
+        offset = offset.to(X_row_dtype)
+
+    mean_square = tl.sum(X_row * X_row, axis=1) / n_cols
+    rstd = rsqrt(mean_square + eps)
+
+    # We can save time by caching rms with minimal memory overhead
+    # because rms is much smaller compared to X_row, as rms is for each row.
+    # However, on the computation side, it can save 4 operations (*, sum, /, sqrt).
+    tl.store(RSTD_ptr + row_idx * RSTD_row_stride, rstd, row_mask)
+
+    X_row = X_row * rstd[:, None]
+
+    # On Llama, the multiplication with the weight is done on the original dtype
+    if casting_mode == _CASTING_MODE_LLAMA:
+        X_row = X_row.to(X_row_dtype)
+
+    if elementwise_affine:
+        Y_row = X_row * (offset + W_row)[None, :]
+    else:
+        Y_row = X_row
+
+    if casting_mode == _CASTING_MODE_GEMMA:
+        Y_row = Y_row.to(X_row_dtype)
+
+    tl.store(
+        Y_ptr + row_idx[:, None] * Y_row_stride + col_offsets[None, :],
+        Y_row,
+        mask=row_mask[:, None] & col_mask[None, :],
+    )
+
+
+@triton.jit
+def _block_rms_norm_backward_kernel(
+    dY_ptr,
+    dY_row_stride,
+    dX_ptr,
+    dX_row_stride,
+    X_ptr,
+    X_row_stride,
+    X_dtype: tl.constexpr,
+    W_ptr,
+    W_row_stride,
+    RSTD_ptr,
+    RSTD_row_stride,
+    dW_ptr,
+    dW_row_stride,
+    n_rows,
+    n_cols,
+    offset,
+    casting_mode: tl.constexpr,
+    elementwise_affine: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    BLOCK_ROW: tl.constexpr,
+):
+    """
+    dx = (1 / RMS) * [dy * (w + offset - (1 / N) * (1 / RMS^2) * ((dy * (w + offset)) dot x) * x]. * means element-wise multiplication, whileas dot means dot product
+    dw = sum(dy * (x / RMS)). summation over BxT dimension
+    """
+
+    pid = tl.program_id(0).cast(tl.int64)
+    NUM_SMS = tl.num_programs(0)
+
+    col_offsets = tl.arange(0, BLOCK_SIZE)
+    col_mask = col_offsets < n_cols
+
+    if elementwise_affine:
+        dW_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+
+        W_row = tl.load(W_ptr + col_offsets, mask=col_mask, other=0.0)
+        W_row = W_row + offset
+
+    for start in range(pid * BLOCK_ROW, n_rows, NUM_SMS * BLOCK_ROW):
+        row_idx = start + tl.arange(0, BLOCK_ROW)
+        row_mask = row_idx < n_rows
+        dY_row = tl.load(
+            dY_ptr + row_idx[:, None] * dY_row_stride + col_offsets[None, :],
+            mask=row_mask[:, None] & col_mask[None, :],
+            other=0.0,
+        )
+        X_row = tl.load(
+            X_ptr + row_idx[:, None] * X_row_stride + col_offsets[None, :],
+            mask=row_mask[:, None] & col_mask[None, :],
+            other=0.0,
+        )
+
+        # Get cached rms
+        rstd_row = tl.load(RSTD_ptr + row_idx * RSTD_row_stride, row_mask)
+
+        X_row = X_row.to(tl.float32)
+
+        # Different bacward graphs for different casting modes
         if casting_mode == _CASTING_MODE_LLAMA:
-            dW_row += dY_row * (X_row * rstd_row).to(X_dtype)
+            if elementwise_affine:
+                m = (dY_row * W_row[None, :]).to(tl.float32)
+            else:
+                m = dY_row.to(tl.float32)
+
+        elif casting_mode == _CASTING_MODE_GEMMA:
+            dY_row = dY_row.to(tl.float32)
+            if elementwise_affine:
+                m = dY_row * W_row[None, :]
+            else:
+                m = dY_row
         else:
-            # here X_row is already in fp32 (see previous if block)
-            dW_row += dY_row * (X_row * rstd_row)
+            if elementwise_affine:
+                m = dY_row * W_row[None, :]
+            else:
+                m = dY_row
 
-        tl.store(dX_ptr + col_offsets, dX_row.to(X_dtype), mask=mask)
+        dX_row = rstd_row[:, None] * m
 
-        dY_ptr += dY_row_stride
-        dX_ptr += dX_row_stride
-        X_ptr += X_row_stride
-        RSTD_ptr += RSTD_row_stride
+        dX_row += (rstd_row[:, None]) * (
+            -(1 / n_cols) * (rstd_row * rstd_row * tl.sum(m * X_row, axis=1))[:, None] * X_row
+        )
+
+        if elementwise_affine:
+            if casting_mode == _CASTING_MODE_LLAMA:
+                # TODO(tcc): use tl.sum(..., dtype=tl.float32) once we upgrade to triton>=3.3.0
+                dW_row += tl.sum((dY_row * (X_row * rstd_row[:, None]).to(X_dtype)).to(tl.float32), 0)
+            else:
+                # here X_row is already in fp32 (see previous if block)
+                dW_row += tl.sum(dY_row * (X_row * rstd_row[:, None]), 0)
+
+        tl.store(
+            dX_ptr + row_idx[:, None] * dX_row_stride + col_offsets[None, :],
+            dX_row,
+            mask=row_mask[:, None] & col_mask[None, :],
+        )
 
-    tl.store(dW_ptr + row_block_id * dW_row_stride + col_offsets, dW_row, mask=mask)
+    if elementwise_affine:
+        tl.store(dW_ptr + pid * dW_row_stride + col_offsets, dW_row, mask=col_mask)
 
 
 _str_to_casting_mode = {
@@ -201,7 +405,7 @@ _str_to_casting_mode = {
 }
 
 
-def rms_norm_forward(X, W, eps, offset, casting_mode):
+def rms_norm_forward(X, W, eps, offset, casting_mode, row_mode):
     if not isinstance(casting_mode, int):
         assert casting_mode in _str_to_casting_mode, f"Invalid casting mode: {casting_mode}"
         casting_mode = _str_to_casting_mode[casting_mode]
@@ -220,29 +424,64 @@ def rms_norm_forward(X, W, eps, offset, casting_mode):
     rstd_dtype = torch.float32 if casting_mode in (_CASTING_MODE_LLAMA.value, _CASTING_MODE_GEMMA.value) else X.dtype
     RSTD = torch.empty(n_rows, dtype=rstd_dtype, device=X.device)
 
-    # Check constraints.
-    assert X.shape[1] == W.shape[0], "Incompatible hidden size dimension between tensor1.shape[1] and tensor2.shape[0]"
-
-    _rms_norm_forward_kernel[(n_rows,)](
-        Y,
-        Y.stride(0),
-        X,
-        X.stride(0),
-        W,
-        W.stride(0),
-        RSTD,
-        RSTD.stride(0),
-        n_cols,
-        eps,
-        offset,
-        casting_mode,
-        BLOCK_SIZE=BLOCK_SIZE,
-        num_warps=num_warps,
-    )
+    if W is not None:
+        # Check constraints.
+        assert X.shape[1] == W.shape[0], (
+            "Incompatible hidden size dimension between tensor1.shape[1] and tensor2.shape[0]"
+        )
+        elementwise_affine = True
+    else:
+        elementwise_affine = False
+
+    # XPU-specific optimization
+    kernel_args = {}
+    if X.device.type == "xpu":
+        kernel_args["grf_mode"] = "large"
+    if BLOCK_SIZE > 256 or n_rows < 4096 * 8 or row_mode:
+        _rms_norm_forward_kernel[(n_rows,)](
+            Y,
+            Y.stride(0),
+            X,
+            X.stride(0),
+            W,
+            W.stride(0) if elementwise_affine else 0,
+            RSTD,
+            RSTD.stride(0),
+            n_cols,
+            eps,
+            offset,
+            casting_mode,
+            elementwise_affine=elementwise_affine,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=num_warps,
+            **kernel_args,  # XPU-specific optimization
+        )
+    else:
+        BLOCK_ROW = 16
+        kernel_args["BLOCK_ROW"] = BLOCK_ROW
+        _block_rms_norm_forward_kernel[(triton.cdiv(n_rows, BLOCK_ROW),)](
+            Y,
+            Y.stride(0),
+            X,
+            X.stride(0),
+            W,
+            W.stride(0) if elementwise_affine else 0,
+            RSTD,
+            RSTD.stride(0),
+            n_rows,
+            n_cols,
+            eps,
+            offset,
+            casting_mode,
+            elementwise_affine=elementwise_affine,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=num_warps,
+            **kernel_args,  # XPU-specific optimization
+        )
     return Y.view(*shape), X, RSTD, BLOCK_SIZE, num_warps, casting_mode
 
 
-def rms_norm_backward(dY, X, W, RSTD, offset, casting_mode, BLOCK_SIZE, num_warps, in_place):
+def rms_norm_backward(dY, X, W, RSTD, offset, casting_mode, BLOCK_SIZE, num_warps, in_place, row_mode):
     shape = dY.shape
     dim = shape[-1]
     dY = dY.view(-1, dim)
@@ -252,10 +491,17 @@ def rms_norm_backward(dY, X, W, RSTD, offset, casting_mode, BLOCK_SIZE, num_warp
     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_subslice_count
-
-    # fp32 for numerical stability especially.
-    _dW = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)
+        sm_count = torch.xpu.get_device_properties(X.device).gpu_eu_count
+    elif X.device.type == "npu":
+        sm_count = get_npu_multi_processor_count()
+
+    if W is not None:
+        # fp32 for numerical stability especially.
+        _dW = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)
+        elementwise_affine = True
+    else:
+        _dW = None
+        elementwise_affine = False
 
     if n_cols > BLOCK_SIZE:
         raise RuntimeError("This layer norm doesn't support feature dim >= 64KB.")
@@ -267,30 +513,68 @@ def rms_norm_backward(dY, X, W, RSTD, offset, casting_mode, BLOCK_SIZE, num_warp
     else:
         dX = torch.zeros_like(dY)
 
-    _rms_norm_backward_kernel[grid](
-        dY,
-        dY.stride(0),
-        dX,
-        dX.stride(0),
-        X,
-        X.stride(0),
-        torch_to_triton_dtype[X.dtype],
-        W,
-        W.stride(0),
-        RSTD,
-        RSTD.stride(0),
-        _dW,
-        _dW.stride(0),
-        n_rows,
-        n_cols,
-        offset,
-        rows_per_program,
-        casting_mode,
-        BLOCK_SIZE=BLOCK_SIZE,
-        num_warps=num_warps,
-    )
+    # XPU-specific optimization
+    kernel_args = {}
+    if X.device.type == "xpu":
+        kernel_args["grf_mode"] = "large"
+
+    if BLOCK_SIZE > 256 or n_rows < 4096 * 8 or row_mode:
+        _rms_norm_backward_kernel[grid](
+            dY,
+            dY.stride(0),
+            dX,
+            dX.stride(0),
+            X,
+            X.stride(0),
+            torch_to_triton_dtype[X.dtype],
+            W,
+            W.stride(0) if elementwise_affine else 0,
+            RSTD,
+            RSTD.stride(0),
+            _dW,
+            _dW.stride(0) if elementwise_affine else 0,
+            n_rows,
+            n_cols,
+            offset,
+            rows_per_program,
+            casting_mode,
+            elementwise_affine=elementwise_affine,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=num_warps,
+            **kernel_args,  # XPU-specific optimization
+        )
+    else:
+        BLOCK_ROW = 16
+        kernel_args["BLOCK_ROW"] = BLOCK_ROW
+        _block_rms_norm_backward_kernel[grid](
+            dY,
+            dY.stride(0),
+            dX,
+            dX.stride(0),
+            X,
+            X.stride(0),
+            torch_to_triton_dtype[X.dtype],
+            W,
+            W.stride(0) if elementwise_affine else 0,
+            RSTD,
+            RSTD.stride(0),
+            _dW,
+            _dW.stride(0) if elementwise_affine else 0,
+            n_rows,
+            n_cols,
+            offset,
+            casting_mode,
+            elementwise_affine=elementwise_affine,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=num_warps,
+            **kernel_args,  # XPU-specific optimization
+        )
     dX = dX.view(*shape)
-    dW = _dW.sum(dim=0).to(W.dtype)
+
+    if elementwise_affine:
+        dW = _dW.sum(dim=0).to(W.dtype)
+    else:
+        dW = None
 
     return dX, dW
 
@@ -319,18 +603,30 @@ class LigerRMSNormFunction(torch.autograd.Function):
 
     @staticmethod
     @ensure_contiguous
-    def forward(ctx, X, W, eps, offset=0.0, casting_mode="llama", in_place=True):
+    def forward(ctx, X, W, eps, offset=0.0, casting_mode="llama", in_place=True, row_mode=None):
         """
         X: (B, T, H) or (BxT, H)
         W: (H,)
         """
-        Y, X, RSTD, BLOCK_SIZE, num_warps, casting_mode = rms_norm_forward(X, W, eps, offset, casting_mode)
+        if isinstance(X, torch.distributed.tensor.DTensor):
+            # Input tensor is output of a tensor parallel module and
+            # needs to be gathered to a local tensor to compute
+            # RMSE layer norm on each TP worker.
+            # TODO: support CP.
+            X = X.full_tensor()
+
+        Y, X, RSTD, BLOCK_SIZE, num_warps, casting_mode = rms_norm_forward(X, W, eps, offset, casting_mode, row_mode)
         ctx.offset = offset
         ctx.casting_mode = casting_mode
         ctx.in_place = in_place
+        ctx.row_mode = row_mode
         ctx.BLOCK_SIZE = BLOCK_SIZE
         ctx.num_warps = num_warps
-        ctx.save_for_backward(X, W, RSTD)
+        ctx.elementwise_affine = W is not None
+        if W is not None:
+            ctx.save_for_backward(X, W, RSTD)
+        else:
+            ctx.save_for_backward(X, RSTD)
         return Y
 
     @staticmethod
@@ -339,16 +635,19 @@ class LigerRMSNormFunction(torch.autograd.Function):
         """
         Y: (B, T, H) or (BxT, H)
         """
-        X, W, RSTD = ctx.saved_tensors
+        if ctx.elementwise_affine:
+            X, W, RSTD = ctx.saved_tensors
+        else:
+            X, RSTD = ctx.saved_tensors
+            W = None
+
+        if isinstance(dY, torch.distributed.tensor.DTensor):
+            # Gradients are output of a tensor parallel module and
+            # needs to be gathered to a local tensor for computing RMSE layer.
+            # TODO: support CP.
+            dY = dY.full_tensor()
+
         dX, dW = rms_norm_backward(
-            dY,
-            X,
-            W,
-            RSTD,
-            ctx.offset,
-            ctx.casting_mode,
-            ctx.BLOCK_SIZE,
-            ctx.num_warps,
-            ctx.in_place,
+            dY, X, W, RSTD, ctx.offset, ctx.casting_mode, ctx.BLOCK_SIZE, ctx.num_warps, ctx.in_place, ctx.row_mode
         )
-        return dX, dW, None, None, None, None
+        return dX, dW, None, None, None, None, None

liger_kernel/ops/rope.py

@@ -32,7 +32,7 @@ def _triton_rope(
 
     # cos size: (1, seq_len, head_dim) or (bsz, seq_len, head_dim)
     # stride: (seq_len * head_dim, head_dim, 1)
-    pid = tl.program_id(0)
+    pid = tl.program_id(0).to(tl.int64)
 
     # locate start address
     q_ptr = q_ptr + pid * q_row_stride