liger-kernel 0.4.2__py3-none-any.whl → 0.5.1__py3-none-any.whl

liger_kernel/ops/cross_entropy.py

@@ -92,8 +92,8 @@ def liger_cross_entropy_kernel(
     # 3. [Online softmax] first pass: find max + sum
     m = float("-inf")  # m is the max value. use the notation from the paper
     d = 0.0  # d is the sum. use the notation from the paper
-    ori_X_y = tl.load(
-        X_ptr + y
+    ori_X_y = tl.load(X_ptr + y).cast(
+        tl.float32
     )  # we need to store the original value of X_y for the loss calculation
     if HAS_SOFTCAPPING:
         ori_X_y = softcap * tanh(ori_X_y / softcap)
@@ -106,8 +106,11 @@ def liger_cross_entropy_kernel(
     for i in range(0, n_cols, BLOCK_SIZE):
         X_offsets = i + tl.arange(0, BLOCK_SIZE)
         X_block = tl.load(
-            X_ptr + X_offsets, mask=X_offsets < n_cols, other=float("-inf")
-        )
+            X_ptr + X_offsets,
+            mask=X_offsets < n_cols,
+            other=float("-inf"),
+            # Ensure float32 precision for softmax calculation
+        ).cast(tl.float32)
         if HAS_SOFTCAPPING:
             X_block = softcap * tanh(X_block / softcap)
         block_max = tl.max(X_block)
@@ -141,8 +144,11 @@ def liger_cross_entropy_kernel(
     for i in range(0, n_cols, BLOCK_SIZE):
         X_offsets = i + tl.arange(0, BLOCK_SIZE)
         X_block = tl.load(
-            X_ptr + X_offsets, mask=X_offsets < n_cols, other=float("-inf")
-        )
+            X_ptr + X_offsets,
+            mask=X_offsets < n_cols,
+            other=float("-inf"),
+            # Ensure float32 precision for softmax calculation
+        ).cast(tl.float32)
         if HAS_SOFTCAPPING:
             intermediate = tanh(X_block / softcap)
             X_block = softcap * intermediate
@@ -279,11 +285,12 @@ def cross_entropy_forward(
         num_warps=32 if not is_hip() else 16,
     )
 
-    loss = torch.sum(loss_1d)
-    if return_z_loss == _TRUE.value:
-        z_loss = torch.sum(z_loss_1d)
+    if reduction == "none":
+        loss = loss_1d
+        z_loss = z_loss_1d if return_z_loss == _TRUE.value else None
     else:
-        z_loss = None
+        loss = torch.sum(loss_1d)
+        z_loss = torch.sum(z_loss_1d) if return_z_loss == _TRUE.value else None
 
     return loss, z_loss, _input
 

liger_kernel/ops/fused_linear_cross_entropy.py

@@ -26,7 +26,6 @@ def fused_linear_cross_entropy_forward(
     reduction="mean",
     softcap=None,
 ):
-    dtype = _input.dtype
     device = _input.device
 
     # inputs have shape: BT x H
@@ -74,9 +73,6 @@ def fused_linear_cross_entropy_forward(
         loss_1d_slice = loss_1d[start_idx:end_idx]  # chunk_size,
         n_non_ignore = (target_chunk != ignore_index).sum().item()
 
-        # when doing CE, use the upcasted precision
-        logits_chunk = logits_chunk.float()
-
         # ensure _input and target are contiguous
         logits_chunk = logits_chunk.contiguous()
         target_chunk = target_chunk.contiguous()
@@ -103,13 +99,6 @@ def fused_linear_cross_entropy_forward(
             num_warps=32 if not is_hip() else 16,
         )
 
-        # gradient of logits_chunk is computed in-place by the above triton kernel.
-        # Following HuggingFace model source code, we do the forward and backward
-        # w.r.t. logits in fp32 for numerical stability especially as the num classes (vocab size) is huge.
-        # (reference: https://github.com/huggingface/transformers/blob/v4.42.4/src/transformers/models/llama/modeling_llama.py#L1194)
-        # Propagating to lm_head's backward, we'll switch back to the original dtype.
-        logits_chunk = logits_chunk.to(dtype)
-
         # gradient of logits_chunk is computed in-place by the above triton kernel and is of shape: chunk_size x V
         # thus grad_input[start_idx: end_idx] should be of shape: chunk_size x H
         # additionally, since we are chunking the inputs, observe that the loss and gradients are calculated only

liger_kernel/ops/fused_linear_jsd.py

@@ -202,7 +202,7 @@ class LigerFusedLinearJSDFunction(torch.autograd.Function):
             teacher_input (torch.tensor): input of the last projection layer in teacher model, with shape (B*T, H), where B is batch size, T is sequence length, H is hidden dimension.
             teacher_weight (torch.tensor): the last projection layer in teacher model, with shape (V, H), where V is vocab size
             shift_labels (Optional[torch.LongTensor]): indicator of next predicted vocab with shape (BT) where each value is in [0, V-1].
-            jsd_beta (float): coefficient beta of generalized JSD in the open interval (0, 1). Default: `0.5`
+            jsd_beta (float): coefficient beta of generalized JSD in the interval [0, 1]. It implements forward/reverse KL when beta equals 0 and 1 respectively. Default: `0.5`
             ignore_index (int): the index to ignore. Default: -100
             temperature (float): temperature in softmax function to control the output probability distribution. Default: `1.0`
 

liger_kernel/ops/jsd.py

@@ -18,7 +18,7 @@ def _jsd_kernel(
     dX_ptr,
     dX_stride,
     label_ptr,
-    beta,
+    beta: tl.constexpr,
     n_non_ignore: int,
     ignore_index: tl.constexpr,
     n_cols,
@@ -50,17 +50,26 @@ def _jsd_kernel(
         X = tl.load(X_ptr + offsets, mask=mask, other=float("-inf")).to(tl.float32)
         Y = tl.load(Y_ptr + offsets, mask=mask, other=float("-inf")).to(tl.float32)
 
-        Q = tl.exp(X)
-        P = tl.exp(Y)
-        M = beta * P + (1 - beta) * Q
-        log_M = tl.log(M)
+        if beta == 0.0:  # forward KL
+            Y_prob = tl.exp(Y)
+            loss = Y_prob * (Y - X)
+            dX = -Y_prob
+        elif beta == 1.0:
+            X_prob = tl.exp(X)
+            loss = X_prob * (X - Y)
+            dX = loss + X_prob
+        else:
+            Q = tl.exp(X)
+            P = tl.exp(Y)
+            M = beta * P + (1 - beta) * Q
+            log_M = tl.log(M)
+
+            loss = beta * P * Y + (1 - beta) * Q * X - M * log_M
+            dX = (1 - beta) * Q * (X - log_M)
 
-        loss = beta * P * Y + (1 - beta) * Q * X - M * log_M
-        # reduction == "batchmean"
         loss = loss / n_non_ignore
+        dX = dX / n_non_ignore
         tl.store(loss_ptr + offsets, loss, mask=mask)
-
-        dX = (1 - beta) * Q * (X - log_M) / n_non_ignore
         tl.store(dX_ptr + offsets, dX, mask=mask)
 
 
@@ -142,7 +151,7 @@ class LigerJSDFunction(torch.autograd.Function):
             _input (torch.Tensor): predict values with shape (BT, V) in logspace
             target (torch.Tensor): ground truth values with shape (BT, V) in logspace
             shift_labels (Optional[torch.LongTensor]): indicator of next predicted vocab with shape (BT) where each value is in [0, V-1].
-            beta (float): coefficient beta of generalized JSD in the open interval (0, 1)
+            beta (float): coefficient beta of generalized JSD in the interval [0, 1]. It implements forward/reverse KL when beta equals 0 and 1 respectively. Default: `0.5`
             ignore_index (int): the index to ignore. Default: -100
 
         Returns:

liger_kernel/ops/layer_norm.py

@@ -180,8 +180,13 @@ def layer_norm_backward(dY, X, W, B, Mean, RSTD):
     dY = dY.view(-1, dim)
     n_rows, n_cols = dY.shape
 
+    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_subslice_count
+
     DX = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
-    sm_count = torch.cuda.get_device_properties(X.device).multi_processor_count
     _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)
 

liger_kernel/ops/qwen2vl_mrope.py

@@ -0,0 +1,238 @@
+import torch
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def _triton_qwen2vl_mrope(
+    q_ptr,
+    k_ptr,
+    cos,
+    sin,
+    sl,
+    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
+
+    cos_row_idx = pid % sl
+    t_cos = cos + cos_row_idx * hd
+    h_cos = t_cos + sl * hd
+    w_cos = h_cos + sl * hd
+    t_sin = sin + cos_row_idx * hd
+    h_sin = t_sin + sl * hd
+    w_sin = h_sin + 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,
+        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,
+        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, 1, seq_len, head_dim)
+        sin size: (3, 1, 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, 1, seq_len, head_dim)
+        sin size: (3, 1, 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

liger_kernel/ops/rms_norm.py

@@ -264,7 +264,12 @@ def rms_norm_backward(
     dY = dY.view(-1, dim)
     n_rows, n_cols = dY.shape
 
-    sm_count = torch.cuda.get_device_properties(X.device).multi_processor_count
+    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_subslice_count
+
     # fp32 for numerical stability especially.
     _dW = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)
 

liger_kernel/ops/utils.py

@@ -20,6 +20,8 @@ import triton
 import triton.language as tl
 from packaging.version import Version
 
+from liger_kernel.utils import infer_device
+
 
 def is_hip() -> bool:
     return torch.version.hip is not None
@@ -69,10 +71,11 @@ def compare_version(package: str, operator: Callable, target: str):
 
 
 def get_amp_custom_fwd_bwd() -> Callable:
+    device = infer_device()
     if compare_version("torch", operator.ge, "2.4.0"):
         return (
-            functools.partial(torch.amp.custom_fwd, device_type="cuda"),
-            functools.partial(torch.amp.custom_bwd, device_type="cuda"),
+            functools.partial(torch.amp.custom_fwd, device_type=device),
+            functools.partial(torch.amp.custom_bwd, device_type=device),
         )
     return torch.cuda.amp.custom_fwd, torch.cuda.amp.custom_bwd
 

liger_kernel/transformers/functional.py

@@ -10,21 +10,11 @@ from liger_kernel.ops.group_norm import LigerGroupNormFunction
 from liger_kernel.ops.jsd import LigerJSDFunction
 from liger_kernel.ops.kl_div import LigerKLDivLossFunction
 from liger_kernel.ops.layer_norm import LigerLayerNormFunction
+from liger_kernel.ops.qwen2vl_mrope import LigerQwen2VLMRopeFunction
 from liger_kernel.ops.rms_norm import LigerRMSNormFunction
 from liger_kernel.ops.rope import LigerRopeFunction
 from liger_kernel.ops.swiglu import LigerSiLUMulFunction
 
-liger_swiglu = LigerSiLUMulFunction.apply
-liger_fused_linear_cross_entropy = LigerFusedLinearCrossEntropyFunction.apply
-liger_geglu = LigerGELUMulFunction.apply
-liger_rms_norm = LigerRMSNormFunction.apply
-liger_rope = LigerRopeFunction.apply
-liger_layer_norm = LigerLayerNormFunction.apply
-liger_kl_div = LigerKLDivLossFunction.apply
-liger_jsd = LigerJSDFunction.apply
-liger_fused_linear_jsd = LigerFusedLinearJSDFunction.apply
-liger_group_norm = LigerGroupNormFunction.apply
-
 
 # conform to the function signature in https://pytorch.org/docs/stable/generated/torch.nn.functional.cross_entropy.html
 # `weight` and `size_average` are placeholders and not implemented yet
@@ -54,3 +44,130 @@ def liger_cross_entropy(
     if not return_z_loss:
         return loss
     return loss, z_loss
+
+
+def liger_fused_linear_cross_entropy(
+    input,
+    weight,
+    target,
+    bias=None,
+    ignore_index: int = -100,
+    lse_square_scale: float = 0.0,
+    label_smoothing: float = 0.0,
+    reduction: str = "mean",
+    softcap: Optional[float] = None,
+):
+    return LigerFusedLinearCrossEntropyFunction.apply(
+        input,
+        weight,
+        target,
+        bias,
+        ignore_index,
+        lse_square_scale,
+        label_smoothing,
+        reduction,
+        softcap,
+    )
+
+
+def liger_fused_linear_jsd(
+    student_input,
+    student_weight,
+    teacher_input,
+    teacher_weight,
+    shift_labels=None,
+    jsd_beta: float = 0.5,
+    ignore_index: int = -100,
+    temperature: float = 1.0,
+):
+    return LigerFusedLinearJSDFunction.apply(
+        student_input,
+        student_weight,
+        teacher_input,
+        teacher_weight,
+        shift_labels,
+        jsd_beta,
+        ignore_index,
+        temperature,
+    )
+
+
+def liger_geglu(a, b):
+    return LigerGELUMulFunction.apply(a, b)
+
+
+def liger_group_norm(
+    X,
+    affine_scaling_weight,
+    affine_shifting_bias,
+    num_channels,
+    num_groups,
+    eps,
+):
+    return LigerGroupNormFunction.apply(
+        X,
+        affine_scaling_weight,
+        affine_shifting_bias,
+        num_channels,
+        num_groups,
+        eps,
+    )
+
+
+def liger_jsd(
+    input,
+    target,
+    shift_labels=None,
+    beta: float = 0.5,
+    ignore_index: int = -100,
+):
+    return LigerJSDFunction.apply(
+        input,
+        target,
+        shift_labels,
+        beta,
+        ignore_index,
+    )
+
+
+# conform to the function signature in https://pytorch.org/docs/stable/generated/torch.nn.functional.kl_div.html#torch.nn.functional.kl_div
+# `size_average` and `mean` are being deprecated in torch API and are placeholders here
+def liger_kl_div(
+    input,
+    target,
+    size_average: bool = True,
+    reduce: bool = True,
+    reduction: str = "mean",
+    log_target: bool = False,
+    eps: float = 1e-10,
+):
+    # Note: the default reduction in torch is `mean`, but being `batchmean` in Liger
+    return LigerKLDivLossFunction.apply(
+        input,
+        target,
+        reduction,
+        log_target,
+        eps,
+    )
+
+
+def liger_layer_norm(X, W, B, eps):
+    return LigerLayerNormFunction.apply(X, W, B, eps)
+
+
+def liger_qwen2vl_mrope(q, k, cos, sin, mrope_section, unsqueeze_dim=1):
+    return LigerQwen2VLMRopeFunction.apply(q, k, cos, sin, mrope_section, unsqueeze_dim)
+
+
+def liger_rms_norm(
+    X, W, eps, offset: float = 0.0, casting_mode: str = "llama", in_place: bool = True
+):
+    return LigerRMSNormFunction.apply(X, W, eps, offset, casting_mode, in_place)
+
+
+def liger_rope(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    return LigerRopeFunction.apply(q, k, cos, sin, position_ids, unsqueeze_dim)
+
+
+def liger_swiglu(a, b):
+    return LigerSiLUMulFunction.apply(a, b)

liger_kernel/transformers/fused_linear_jsd.py

@@ -12,7 +12,7 @@ class LigerFusedLinearJSD(torch.nn.Module):
     the materialization of the large logits tensor.
 
     Args:
-        jsd_beta (float): coefficient beta of generalized JSD in the open interval (0, 1). Default: `0.5`
+        jsd_beta (float): coefficient beta of generalized JSD in the interval [0, 1]. It implements forward/reverse KL when beta equals 0 and 1 respectively. Default: `0.5`
         ignore_index (int): The index to ignore in the target. Default: `-100`
         temperature (float): temperature in softmax function to control the output probability distribution. Default: `1.0`
 
@@ -70,9 +70,6 @@ class LigerFusedLinearJSD(torch.nn.Module):
 
     def __init__(self, jsd_beta=0.5, ignore_index=-100, temperature=1.0):
         super().__init__()
-        assert (
-            jsd_beta > 0 and jsd_beta < 1
-        ), f"beta must be greater than 0 and less than 1. Got: {jsd_beta}"
         assert temperature != 0, "temperature cannot be 0."
         self.jsd_beta = jsd_beta
         self.temperature = temperature

liger_kernel/transformers/jsd.py

@@ -18,7 +18,7 @@ class LigerJSD(torch.nn.Module):
     :math:`P` denotes the teacher model and :math:`Q` denotes the student model.
 
     Args:
-        beta (float): coefficient beta of generalized JSD in the open interval (0, 1). Default: `0.5`
+        beta (float): coefficient beta of generalized JSD in the interval [0, 1]. It implements forward/reverse KL when beta equals 0 and 1 respectively. Default: `0.5`
         ignore_index (int): The index to ignore in the target. Default: `-100`
 
     Shape:
@@ -58,9 +58,6 @@ class LigerJSD(torch.nn.Module):
 
     def __init__(self, beta: float = 0.5, ignore_index: int = -100):
         super().__init__()
-        assert (
-            beta > 0 and beta < 1
-        ), f"beta must be greater than 0 and less than 1. Got: {beta}"
         self.beta = beta
         self.ignore_index = ignore_index
 

liger_kernel/transformers/monkey_patch.py

@@ -36,6 +36,7 @@ from liger_kernel.transformers.model.qwen2 import lce_forward as qwen2_lce_forwa
 from liger_kernel.transformers.model.qwen2 import (
     lce_forward_deprecated as qwen2_lce_forward_deprecated,
 )
+from liger_kernel.transformers.qwen2vl_mrope import liger_multimodal_rotary_pos_emb
 from liger_kernel.transformers.rms_norm import LigerRMSNorm
 from liger_kernel.transformers.rope import liger_rotary_pos_emb
 from liger_kernel.transformers.swiglu import (
@@ -610,9 +611,7 @@ def apply_liger_kernel_to_qwen2(
             logger.warning(TRANSFORMER_DEPRECATION_WARNING)
             modeling_qwen2.CrossEntropyLoss = LigerCrossEntropyLoss
 
-    # import pdb; pdb.set_trace()
     if fused_linear_cross_entropy:
-
         if transformer_version >= version.parse(SUPPORTED_TRANSFORMER_VERSION):
             modeling_qwen2.Qwen2ForCausalLM.forward = qwen2_lce_forward
         else:  # if version < 4.46.1
@@ -644,6 +643,7 @@ def apply_liger_kernel_to_qwen2(
 
 
 def apply_liger_kernel_to_qwen2_vl(
+    rope: bool = True,
     cross_entropy: bool = False,
     fused_linear_cross_entropy: bool = True,
     rms_norm: bool = True,
@@ -678,8 +678,10 @@ def apply_liger_kernel_to_qwen2_vl(
         lce_forward as qwen2_vl_lce_forward,
     )
 
-    # TODO: Support Qwen2-VL's multimodal RoPE implementation
-
+    if rope:
+        modeling_qwen2_vl.apply_multimodal_rotary_pos_emb = (
+            liger_multimodal_rotary_pos_emb
+        )
     if rms_norm:
         # https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py#L439
         modeling_qwen2_vl.Qwen2RMSNorm = LigerRMSNorm