liger-kernel-nightly 0.5.10.dev20250611191801__py3-none-any.whl → 0.6.4.dev20260112233432__py3-none-any.whl

liger_kernel/ops/geglu.py

@@ -7,8 +7,9 @@ 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
+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 tanh
@@ -40,7 +41,7 @@ def _geglu_tanh_forward_kernel(a, b, c, stride, n_cols: tl.constexpr, BLOCK_SIZE
     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 * b_row
+    c_row = geglu_a.cast(b_row.dtype) * b_row
     tl.store(c + col_offsets, c_row, mask=mask)
 
 
@@ -66,8 +67,9 @@ def _geglu_tanh_backward_kernel(dc, a, b, stride, n_cols: tl.constexpr, BLOCK_SI
     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 * geglu_a
+    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)))
@@ -78,7 +80,7 @@ def _geglu_tanh_backward_kernel(dc, a, b, stride, n_cols: tl.constexpr, BLOCK_SI
     da_row = dc_row * b_row * (term1 + term2)
 
     tl.store(a + col_offsets, da_row, mask=mask)
-    tl.store(b + col_offsets, db_row, mask=mask)
+    tl.store(b + col_offsets, db_row.to(dc_row.dtype), mask=mask)
 
 
 def geglu_forward(a, b):

liger_kernel/ops/group_norm.py

@@ -6,8 +6,9 @@ import triton.language as tl
 
 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"):
+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
@@ -77,15 +78,14 @@ def _group_norm_forward_kernel(
     for channel_idx in tl.range(group_idx * channels_per_group, (group_idx + 1) * channels_per_group):
         W = tl.load(W_ptr + channel_idx)
         B = tl.load(B_ptr + channel_idx)
-        for i in range(0, hidden_size_per_channel, BLOCK_SIZE):
+        # Calculate channel offset within the group
+        channel_offset = (channel_idx - group_idx * channels_per_group) * hidden_size_per_channel
+        for i in tl.range(0, hidden_size_per_channel, BLOCK_SIZE):
             hidden_size_offsets = i + block_range
             mask = hidden_size_offsets < hidden_size_per_channel
-            X = tl.load(X_ptr + hidden_size_offsets, mask=mask, other=m)
+            X = tl.load(X_ptr + channel_offset + hidden_size_offsets, mask=mask, other=m)
             Y = (X - m) * rstd * W + B
-            tl.store(Y_ptr + hidden_size_offsets, Y, mask=mask)
-
-        X_ptr += hidden_size_per_channel
-        Y_ptr += hidden_size_per_channel
+            tl.store(Y_ptr + channel_offset + hidden_size_offsets, Y, mask=mask)
 
     tl.store(Mean_ptr + batch_idx * Mean_row_stride + group_idx * Mean_col_stride, m)
     tl.store(RSTD_ptr + batch_idx * RSTD_row_stride + group_idx * RSTD_col_stride, rstd)

liger_kernel/ops/grpo_loss.py

@@ -128,7 +128,9 @@ def _grpo_loss_fwd_kernel(
     per_token_loss1 = coef_1 * advantage
     per_token_loss2 = coef_2 * advantage
     per_token_loss = -tl.minimum(per_token_loss1, per_token_loss2)
-    is_clipped = per_token_loss1 < per_token_loss2
+    is_low_clipped = (coef_1 < 1 - EPS_LOW) & (advantage < 0)
+    is_high_clipped = (coef_1 > 1 + EPS_HIGH) & (advantage > 0)
+    is_clipped = is_low_clipped | is_high_clipped
 
     if BETA != 0.0:
         REF_LOGP += off_b * L + off_l

liger_kernel/ops/kl_div.py

@@ -21,7 +21,12 @@ def get_num_warps(BLOCK_SIZE):
     return num_warps
 
 
-MAX_FUSED_SIZE = 65536 // 4  # 65536 // 4 or 8 works the best
+if infer_device() == "xpu":
+    MAX_FUSED_SIZE = 8192
+elif infer_device() == "npu":
+    MAX_FUSED_SIZE = 8192
+else:
+    MAX_FUSED_SIZE = 65536 // 4  # 65536 // 4 or 8 works the best
 
 REDUCTION_LITERAL = Literal["none", "sum", "mean", "batchmean"]
 
@@ -116,11 +121,7 @@ def _kldiv_kernel_backward(
 
 def kldiv_forward_triton(y_pred, y_true, log_target, reduction, eps):  # [BT, V]
     BT, V = y_pred.shape
-    BLOCK_SIZE = (
-        min(8192, triton.next_power_of_2(V))
-        if infer_device() == "xpu"
-        else min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
-    )
+    BLOCK_SIZE = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
     num_warps = 32 if infer_device() == "xpu" else get_num_warps(BLOCK_SIZE)
 
     grid = (BT,)
@@ -159,11 +160,7 @@ def kldiv_forward_triton(y_pred, y_true, log_target, reduction, eps):  # [BT, V]
 
 def kldiv_backward_triton(target, grad_output, new_grads, log_target):
     BT, V = target.shape
-    BLOCK_SIZE = (
-        min(8192, triton.next_power_of_2(V))
-        if infer_device() == "xpu"
-        else min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
-    )
+    BLOCK_SIZE = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
     num_warps = 32 if infer_device() == "xpu" else get_num_warps(BLOCK_SIZE)
 
     grid = (BT,)

liger_kernel/ops/layer_norm.py

@@ -8,8 +8,9 @@ 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
+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
@@ -43,111 +44,157 @@ def _layer_norm_forward_kernel(
     https://arxiv.org/abs/1607.06450
     https://github.com/karpathy/llm.c/blob/master/doc/layernorm/layernorm.md
     """
-    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
-    Mean_ptr += row_idx * Mean_row_stride
-    RSTD_ptr += row_idx * RSTD_row_stride
-
-    X_row = tl.load(X_ptr + col_offsets, mask=mask, other=0)
-    W_row = tl.load(W_ptr + col_offsets, mask=mask, other=0)
-    B_row = tl.load(B_ptr + col_offsets, mask=mask, other=0)
-
-    mean = tl.sum(X_row, axis=0) / n_cols
-    Xmm = tl.where(mask, X_row - mean, 0)
-    var = tl.sum(Xmm * Xmm, axis=0) / n_cols
+    # Pre-load weights and bias in fp32 to avoid repeated conversions
+    W_row = tl.load(W_ptr + col_offsets, mask=mask, other=0.0)
+    B_row = tl.load(B_ptr + col_offsets, mask=mask, other=0.0)
+    W_f32 = W_row.to(tl.float32)
+    B_f32 = B_row.to(tl.float32)
+
+    # Calculate pointers for this row
+    row_X_ptr = X_ptr + row_idx * X_row_stride
+    row_Y_ptr = Y_ptr + row_idx * Y_row_stride
+    row_Mean_ptr = Mean_ptr + row_idx * Mean_row_stride
+    row_RSTD_ptr = RSTD_ptr + row_idx * RSTD_row_stride
+
+    # Load input data and convert to fp32 for numerical stability
+    X_row = tl.load(row_X_ptr + col_offsets, mask=mask, other=0.0)
+    X_f32 = X_row.to(tl.float32)
+
+    # Compute statistics in fp32 for numerical stability
+    mean = tl.sum(X_f32, axis=0) / n_cols
+    X_centered = X_f32 - mean
+    # Apply mask to variance calculation to exclude contributions from masked elements
+    X_centered_masked = tl.where(mask, X_centered, 0.0)
+    var = tl.sum(X_centered_masked * X_centered_masked, axis=0) / n_cols
     rstd = rsqrt(var + eps)
 
-    tl.store(Mean_ptr, mean)
-    tl.store(RSTD_ptr, rstd)
+    # Store statistics (convert back to original dtype only once)
+    tl.store(row_Mean_ptr, mean.to(X_row.dtype))
+    tl.store(row_RSTD_ptr, rstd.to(X_row.dtype))
 
-    Y_row = Xmm * rstd * W_row + B_row
+    # Fused normalization and affine transformation
+    # Y = (X - mean) * rstd * W + B = X_centered * rstd * W + B
+    Y_f32 = X_centered * rstd * W_f32 + B_f32
 
-    tl.store(Y_ptr + col_offsets, Y_row, mask=mask)
+    # Store output (single conversion back to original dtype)
+    tl.store(row_Y_ptr + col_offsets, Y_f32.to(X_row.dtype), mask=mask)
 
 
 @triton.jit
 def _layer_norm_backward_kernel(
     X_ptr,  # pointer to input, shape (n_rows, n_cols)
+    stride_x,  # stride of each row in input
     W_ptr,  # pointer to weights, shape (n_cols,)
     Mean_ptr,  # pointer to mean, shape (n_rows,)
+    stride_mean,  # stride of each row in mean
     RSTD_ptr,  # pointer to rstd, shape (n_rows,)
+    stride_rstd,  # stride of each row in rstd
     DX_ptr,  # pointer to input grad, shape (n_rows, n_cols)
-    DW_ptr,  # pointer to weights grad, shape (n_cols,)
-    DB_ptr,  # pointer to bias grad, shape (n_cols,)
-    DY_ptr,  # pointer to output grad, shape (n_rows, n_cols)
-    stride_x,  # stride of each row in input
     stride_dx,  # stride of each row in input grad
+    DW_ptr,  # pointer to weights grad, shape (n_cols,)
     stride_dw,  # stride of each row in weights grad
+    DB_ptr,  # pointer to bias grad, shape (n_cols,)
     stride_db,  # stride of each row in bias grad
+    DY_ptr,  # pointer to output grad, shape (n_rows, n_cols)
     stride_dy,  # stride of each row in output grad
     n_rows,
     n_cols,
     rows_per_program: tl.constexpr,
     BLOCK_SIZE: tl.constexpr,
-    dtype: tl.constexpr,
 ):
     """
     References:
     https://arxiv.org/abs/1607.06450
     https://github.com/karpathy/llm.c/blob/master/doc/layernorm/layernorm.md
-    https://triton-lang.org/main/getting-started/tutorials/05-layer-norm.html
-    https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/ops/triton/layer_norm.py
     """
-    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)
     cols = tl.arange(0, BLOCK_SIZE)
     mask = cols < n_cols
 
-    dw_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+    dW_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
     db_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
 
-    X_ptr += row_start * stride_x
-    Mean_ptr += row_start
-    RSTD_ptr += row_start
-    DX_ptr += row_start * stride_dx
-    DY_ptr += row_start * stride_dy
+    # Pre-load weights once (same optimization as forward pass)
+    w = tl.load(W_ptr + cols, mask=mask, other=0.0)
+    w_f32 = w.to(tl.float32)
+
+    # Calculate pointers for this specific row
+    row_X_ptr = X_ptr + row_start * stride_x
+    row_DX_ptr = DX_ptr + row_start * stride_dx
+    row_DY_ptr = DY_ptr + row_start * stride_dy
+    row_Mean_ptr = Mean_ptr + row_start
+    row_RSTD_ptr = RSTD_ptr + row_start
 
     for _ in range(row_start, row_end):
-        x = tl.load(X_ptr + cols, mask=mask, other=0.0)
-        w = tl.load(W_ptr + cols, mask=mask, other=0.0)
-        dy = tl.load(DY_ptr + cols, mask=mask, other=0.0)
-        mean = tl.load(Mean_ptr)
-        rstd = tl.load(RSTD_ptr)
-
-        x_hat = (x - mean) * rstd
-        wdy = w * dy
+        # Load data for this row
+        x = tl.load(row_X_ptr + cols, mask=mask, other=0.0)
+        dy = tl.load(row_DY_ptr + cols, mask=mask, other=0.0)
+        mean = tl.load(row_Mean_ptr)
+        rstd = tl.load(row_RSTD_ptr)
+
+        # Convert to fp32 for numerical stability
+        x_f32 = x.to(tl.float32)
+        dy_f32 = dy.to(tl.float32)
+        mean_f32 = mean.to(tl.float32)
+        rstd_f32 = rstd.to(tl.float32)
+
+        # Compute backward pass for this row
+        x_hat = (x_f32 - mean_f32) * rstd_f32
+        wdy = w_f32 * dy_f32
         c1 = tl.sum(x_hat * wdy, axis=0) / n_cols
         c2 = tl.sum(wdy, axis=0) / n_cols
-        dx = (wdy - (x_hat * c1 + c2)) * rstd
-        tl.store(DX_ptr + cols, dx.to(dtype), mask=mask)
+        dx = (wdy - (x_hat * c1 + c2)) * rstd_f32
+
+        # Store input gradient
+        tl.store(row_DX_ptr + cols, dx, mask=mask)
 
-        dw_row += dy * x_hat
-        db_row += dy
+        # Accumulate weight and bias gradients for this thread block's assigned rows
+        dw = dy_f32 * x_hat
+        db = dy_f32
+        dW_row += dw
+        db_row += db
 
-        X_ptr += stride_x
-        Mean_ptr += 1
-        RSTD_ptr += 1
-        DX_ptr += stride_dx
-        DY_ptr += stride_dy
+        row_X_ptr += stride_x
+        row_DX_ptr += stride_dx
+        row_DY_ptr += stride_dy
+        row_Mean_ptr += stride_mean
+        row_RSTD_ptr += stride_rstd
 
-    tl.store(DW_ptr + row_block_id * stride_dw + cols, dw_row.to(dtype), mask=mask)
-    tl.store(DB_ptr + row_block_id * stride_db + cols, db_row.to(dtype), mask=mask)
+    tl.store(DW_ptr + row_block_id * stride_dw + cols, dW_row, mask=mask)
+    tl.store(DB_ptr + row_block_id * stride_db + cols, db_row, mask=mask)
 
 
 def layer_norm_forward(X, W, B, eps):
+    """
+    Args:
+        X: Input tensor of shape (..., hidden_size)
+        W: Weight tensor of shape (hidden_size,)
+        B: Bias tensor of shape (hidden_size,)
+        eps: Small constant for numerical stability
+
+    Returns:
+        Tuple of (output, input, mean, rstd, block_size, num_warps)
+    """
     shape = X.shape
     dim = shape[-1]
     X = X.view(-1, dim)
     n_rows, n_cols = X.shape
+
+    # Calculate optimal block size and warp configuration
     BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+
+    # Allocate output tensors
     Y = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
     Mean = torch.empty(n_rows, dtype=X.dtype, device=X.device)
     RSTD = torch.empty(n_rows, dtype=X.dtype, device=X.device)
+
+    # Validate input dimensions
     if X.shape[1] != W.shape[0]:
         raise ValueError(
             f"Incompatible dimensions: input feature size (X.shape[1]={X.shape[1]}) "
@@ -159,7 +206,9 @@ def layer_norm_forward(X, W, B, eps):
     if X.device.type == "xpu":
         kernel_args["grf_mode"] = "large"
 
-    _layer_norm_forward_kernel[(n_rows,)](
+    # Launch kernel with one thread block per row for optimal performance
+    grid = (n_rows,)
+    _layer_norm_forward_kernel[grid](
         Y,
         Y.stride(0),
         X,
@@ -176,12 +225,25 @@ def layer_norm_forward(X, W, B, eps):
         eps,
         BLOCK_SIZE=BLOCK_SIZE,
         num_warps=num_warps,
-        **kernel_args,  # XPU-specific optimization
+        **kernel_args,
     )
+
     return Y.view(*shape), X, Mean, RSTD, BLOCK_SIZE, num_warps
 
 
 def layer_norm_backward(dY, X, W, B, Mean, RSTD):
+    """
+    Args:
+        dY: Gradient of output
+        X: Input tensor
+        W: Weight tensor
+        B: Bias tensor
+        Mean: Pre-computed mean
+        RSTD: Pre-computed reciprocal standard deviation
+
+    Returns:
+        Tuple of (input_grad, weight_grad, bias_grad)
+    """
     shape = dY.shape
     dim = shape[-1]
     dY = dY.view(-1, dim)
@@ -193,59 +255,52 @@ def layer_norm_backward(dY, X, W, B, Mean, RSTD):
     elif X.device.type == "xpu":
         sm_count = torch.xpu.get_device_properties(X.device).gpu_eu_count
 
-    DX = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
-    _DW = torch.empty((sm_count, n_cols), dtype=W.dtype, device=W.device)
-    _DB = torch.empty((sm_count, n_cols), dtype=W.dtype, device=W.device)
+    # fp32 for numerical stability especially.
+    _DW = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)
+    _DB = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)
 
+    # Calculate optimal block size and warp configuration
     BLOCK_SIZE, num_warps = calculate_settings(n_cols)
     if n_cols > BLOCK_SIZE:
-        raise RuntimeError(
-            f"Feature dimension {n_cols} exceeds maximum supported size of {BLOCK_SIZE}. Consider using a smaller feature dimension."
-        )
-
+        raise RuntimeError(f"Feature dimension {n_cols} exceeds maximum supported size of {BLOCK_SIZE}.")
     rows_per_program = math.ceil(n_rows / sm_count)
     grid = (sm_count,)
-    triton_dtype = (
-        tl.float32
-        if X.dtype == torch.float32
-        else tl.bfloat16
-        if X.dtype == torch.bfloat16
-        else tl.float16
-        if X.dtype == torch.float16
-        else tl.float32  # fallback to float32 for other types
-    )
 
+    # Allocate gradient tensors
+    DX = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
+
+    kernel_args = {"num_warps": num_warps}
     # XPU-specific optimization
-    kernel_args = {}
     if X.device.type == "xpu":
         kernel_args.update({"grf_mode": "large", "num_warps": 32, "num_stages": 4})
 
+    # Launch kernel with one thread block per row for optimal performance
     _layer_norm_backward_kernel[grid](
         X,
+        X.stride(0),
         W,
         Mean,
+        Mean.stride(0),
         RSTD,
+        RSTD.stride(0),
         DX,
-        _DW,
-        _DB,
-        dY,
-        X.stride(0),
         DX.stride(0),
+        _DW,
         _DW.stride(0),
+        _DB,
         _DB.stride(0),
+        dY,
         dY.stride(0),
         n_rows,
         n_cols,
-        rows_per_program,
+        rows_per_program=rows_per_program,
         BLOCK_SIZE=BLOCK_SIZE,
-        dtype=triton_dtype,
-        **kernel_args,  # XPU-specific optimization
+        **kernel_args,
     )
 
-    DW = _DW.sum(dim=0).to(W.dtype)
-    DB = _DB.sum(dim=0).to(W.dtype)
-
     DX = DX.view(*shape)
+    DW = _DW.sum(dim=0).to(W.dtype)
+    DB = _DB.sum(dim=0).to(B.dtype)
     return DX, DW, DB
 
 

liger_kernel/ops/llama4_rope.py

@@ -0,0 +1,225 @@
+import torch
+import triton
+import triton.language as tl
+
+
+def _prepare_freqs(freqs_cis: torch.Tensor, seq_len: int, head_dim_half: int):
+    # Split or unpack complex frequencies into real and imag parts
+    if freqs_cis.is_complex():
+        freqs_real = freqs_cis.real
+        freqs_imag = freqs_cis.imag
+    else:
+        # Already split: last dim should be 2*head_dim_half
+        if freqs_cis.shape[-1] == 2 * head_dim_half:
+            freqs_real = freqs_cis[..., :head_dim_half]
+            freqs_imag = freqs_cis[..., head_dim_half:]
+        else:
+            raise ValueError(
+                f"Unexpected freqs_cis shape for non-complex input: {freqs_cis.shape}, expected last dim = {2 * head_dim_half}"
+            )
+
+    # Canonicalize to shape (seq_len, head_dim_half):
+    # 1) Ensure the last dimension is head_dim_half
+    if freqs_real.shape[-1] != head_dim_half:
+        raise ValueError(f"Unexpected last dim for freqs: {freqs_real.shape[-1]} (expected {head_dim_half})")
+    # 2) Flatten all leading dims to a single row dimension
+    freqs_real = freqs_real.reshape(-1, head_dim_half)
+    freqs_imag = freqs_imag.reshape(-1, head_dim_half)
+    # 3) If we have fewer rows than seq_len, allow broadcasting when single row
+    if freqs_real.shape[0] < seq_len:
+        if freqs_real.shape[0] == 1:
+            freqs_real = freqs_real.expand(seq_len, -1)
+            freqs_imag = freqs_imag.expand(seq_len, -1)
+        else:
+            raise ValueError(f"Insufficient rows in freqs: {freqs_real.shape[0]} < seq_len={seq_len}")
+    # 4) If we have more rows than seq_len (e.g., batch present), take the first seq_len rows
+    elif freqs_real.shape[0] > seq_len:
+        freqs_real = freqs_real[:seq_len]
+        freqs_imag = freqs_imag[:seq_len]
+
+    return freqs_real, freqs_imag
+
+
+def _maybe_to_dtype(t: torch.Tensor, dtype: torch.dtype) -> torch.Tensor:
+    return t if t.dtype == dtype else t.to(dtype)
+
+
+def _maybe_contiguous(t: torch.Tensor) -> torch.Tensor:
+    return t if t.is_contiguous() else t.contiguous()
+
+
+def _cast_and_contiguous(q, k, freqs_real, freqs_imag):
+    # Choose compute dtype: use fp32 only when inputs are fp32; otherwise keep input dtype for performance
+    compute_dtype = torch.float32 if q.dtype == torch.float32 else q.dtype
+
+    # Make sure q/k share the same dtype before casting to compute dtype
+    if k.dtype != q.dtype:
+        k = k.to(q.dtype)
+
+    q = _maybe_contiguous(_maybe_to_dtype(q, compute_dtype))
+    k = _maybe_contiguous(_maybe_to_dtype(k, compute_dtype))
+    freqs_real = _maybe_contiguous(_maybe_to_dtype(freqs_real, compute_dtype))
+    freqs_imag = _maybe_contiguous(_maybe_to_dtype(freqs_imag, compute_dtype))
+    return q, k, freqs_real, freqs_imag
+
+
+@triton.jit
+def _llama4_rope_kernel(
+    q_ptr,
+    k_ptr,
+    freqs_real_ptr,
+    freqs_imag_ptr,
+    q_row_stride,
+    k_row_stride,
+    q_head_stride,
+    k_head_stride,
+    freqs_row_stride,
+    seq_len,
+    batch_size,
+    imag_sign,
+    head_dim_half: tl.constexpr,
+    n_q_heads: tl.constexpr,
+    n_k_heads: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+):
+    """
+    H100-optimized RoPE kernel with improved parallelization across heads and dimensions.
+    Grid: (batch*seq, head)
+    """
+    # 2D grid
+    pid_bs = tl.program_id(0)  # over batch*seq
+    pid_h = tl.program_id(1)  # over heads
+
+    batch_idx = pid_bs // seq_len
+    seq_idx = pid_bs % seq_len
+
+    # Bounds check
+    if batch_idx >= batch_size or seq_idx >= seq_len:
+        return
+
+    # Base pointers for this (batch, seq) position
+    base_offset = batch_idx * seq_len + seq_idx
+    q_base = q_ptr + base_offset * q_row_stride
+    k_base = k_ptr + base_offset * k_row_stride
+
+    # Tiling over dim/2
+    for d_start in tl.static_range(0, head_dim_half, BLOCK_SIZE):
+        d_indices = d_start + tl.arange(0, BLOCK_SIZE)
+        mask_d = d_indices < head_dim_half
+
+        # Load frequencies once per tile (freqs layout: [seq_len, head_dim_half])
+        freq_idx = d_indices
+        freqs_real = tl.load(freqs_real_ptr + seq_idx * freqs_row_stride + freq_idx, mask=mask_d, other=0.0)
+        freqs_imag = tl.load(freqs_imag_ptr + seq_idx * freqs_row_stride + freq_idx, mask=mask_d, other=0.0)
+        freqs_imag = freqs_imag * imag_sign
+
+        # Process one query head per program in pid_h
+        if pid_h < n_q_heads:
+            q_head_ptr = q_base + pid_h * q_head_stride
+            q_real = tl.load(q_head_ptr + d_indices * 2, mask=mask_d, other=0.0)
+            q_imag = tl.load(q_head_ptr + d_indices * 2 + 1, mask=mask_d, other=0.0)
+
+            # Complex multiply with FMAs: (a+ib)*(c+i d) = (a*c - b*d) + i(a*d + b*c)
+            new_q_real = tl.math.fma(q_real, freqs_real, -(q_imag * freqs_imag))
+            new_q_imag = tl.math.fma(q_real, freqs_imag, q_imag * freqs_real)
+
+            tl.store(q_head_ptr + d_indices * 2, new_q_real, mask=mask_d)
+            tl.store(q_head_ptr + d_indices * 2 + 1, new_q_imag, mask=mask_d)
+
+        # Process one key head per program in pid_h
+        if pid_h < n_k_heads:
+            k_head_ptr = k_base + pid_h * k_head_stride
+            k_real = tl.load(k_head_ptr + d_indices * 2, mask=mask_d, other=0.0)
+            k_imag = tl.load(k_head_ptr + d_indices * 2 + 1, mask=mask_d, other=0.0)
+
+            new_k_real = tl.math.fma(k_real, freqs_real, -(k_imag * freqs_imag))
+            new_k_imag = tl.math.fma(k_real, freqs_imag, k_imag * freqs_real)
+
+            tl.store(k_head_ptr + d_indices * 2, new_k_real, mask=mask_d)
+            tl.store(k_head_ptr + d_indices * 2 + 1, new_k_imag, mask=mask_d)
+
+
+def _select_kernel_meta(head_dim_half: int):
+    # Heuristic tuning for block size and num_warps
+    if head_dim_half >= 256:
+        return 128, 8
+    if head_dim_half >= 96:
+        return 128, 4
+    if head_dim_half >= 48:
+        return 64, 4
+    if head_dim_half >= 24:
+        return 32, 2
+    return 16, 2
+
+
+def llama4_rope_forward(q, k, freqs_cis, BLOCK_SIZE: int = None, imag_sign: float = 1.0):
+    # Save original dtype for casting back
+    original_dtype = q.dtype
+
+    batch_size, seq_len, n_q_heads, head_dim = q.shape
+    _, _, n_k_heads, _ = k.shape
+    head_dim_half = head_dim // 2
+
+    # Prepare frequencies
+    freqs_real, freqs_imag = _prepare_freqs(freqs_cis, seq_len, head_dim_half)
+
+    # Cast to appropriate dtype and make contiguous only when needed
+    q, k, freqs_real, freqs_imag = _cast_and_contiguous(q, k, freqs_real, freqs_imag)
+
+    # H100-optimized meta-params
+    if BLOCK_SIZE is None:
+        BLOCK_SIZE, num_warps = _select_kernel_meta(head_dim_half)
+    else:
+        # Provide a default num_warps if caller pins BLOCK_SIZE
+        _, num_warps = _select_kernel_meta(head_dim_half)
+
+    # 2D grid: one program per (batch, seq, head)
+    n_heads_max = max(n_q_heads, n_k_heads)
+    grid = (batch_size * seq_len, n_heads_max)
+
+    # Launch kernel
+    _llama4_rope_kernel[grid](
+        q,
+        k,
+        freqs_real,
+        freqs_imag,
+        q.stride(1),
+        k.stride(1),
+        q.stride(2),
+        k.stride(2),
+        freqs_real.stride(0),
+        seq_len,
+        batch_size,
+        imag_sign,
+        head_dim_half,
+        n_q_heads,
+        n_k_heads,
+        BLOCK_SIZE,
+        num_warps=num_warps,
+        num_stages=2,
+    )
+
+    # Cast back to original dtype only if it differs from compute dtype
+    if q.dtype != original_dtype:
+        q = q.to(original_dtype)
+    if k.dtype != original_dtype:
+        k = k.to(original_dtype)
+
+    return q, k
+
+
+class LigerLlama4RopeFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, q, k, freqs_cis, BLOCK_SIZE: int = None):
+        q_out, k_out = llama4_rope_forward(q, k, freqs_cis, BLOCK_SIZE, imag_sign=1.0)
+        ctx.save_for_backward(freqs_cis.detach() if isinstance(freqs_cis, torch.Tensor) else freqs_cis)
+        ctx.BLOCK_SIZE = BLOCK_SIZE
+        return q_out, k_out
+
+    @staticmethod
+    def backward(ctx, dq, dk):
+        (freqs_cis,) = ctx.saved_tensors
+        BLOCK_SIZE = getattr(ctx, "BLOCK_SIZE", None)
+        # Use imag_sign=-1.0 for conjugate without materializing a new tensor
+        dq_out, dk_out = llama4_rope_forward(dq, dk, freqs_cis, BLOCK_SIZE, imag_sign=-1.0)
+        return dq_out, dk_out, None