Kernels:

kernels-community
/

liger-kernels

Trusted publisher

Kernel card Files Files and versions

xet

Community

kernels-bot commited on 2 days ago

Commit

95d01a6

verified ·

1 Parent(s): b70db7c

Uploaded using `kernel-builder`.

Browse files

Files changed (21) hide show

build/torch-rocm/__init__.py +10 -0
build/torch-rocm/_ops.py +38 -0
build/torch-rocm/cross_entropy.py +558 -0
build/torch-rocm/dyt.py +164 -0
build/torch-rocm/fused_linear_cross_entropy.py +400 -0
build/torch-rocm/geglu.py +143 -0
build/torch-rocm/group_norm.py +311 -0
build/torch-rocm/jsd.py +201 -0
build/torch-rocm/kl_div.py +259 -0
build/torch-rocm/layer_norm.py +320 -0
build/torch-rocm/layers.py +342 -0
build/torch-rocm/liger_kernels/__init__.py +26 -0
build/torch-rocm/metadata.json +45 -0
build/torch-rocm/metadata.json.sigstore +1 -0
build/torch-rocm/qwen2vl_mrope.py +222 -0
build/torch-rocm/rms_norm.py +654 -0
build/torch-rocm/rope.py +239 -0
build/torch-rocm/swiglu.py +176 -0
build/torch-rocm/tiled_mlp.py +159 -0
build/torch-rocm/tvd.py +218 -0
build/torch-rocm/utils.py +176 -0

build/torch-rocm/__init__.py ADDED Viewed

	@@ -0,0 +1,10 @@

+from . import layers
+from .layers import LigerForCausalLMLoss, liger_rotary_pos_emb
+__all__ = [
+    "layers",
+    # kept for BC
+    "LigerForCausalLMLoss",
+    "liger_rotary_pos_emb",
+]

build/torch-rocm/_ops.py ADDED Viewed

	@@ -0,0 +1,38 @@

+import torch
+def get_backend() -> str:
+    """Detect the backend by inspecting torch."""
+    import torch
+    if hasattr(torch, "neuron"):
+        # Needs to be sorted before specific Torch builds, since Neuron
+        # extension can be loaded into e.g. CUDA Torch builds.
+        return "neuron"
+    elif torch.version.cuda is not None:
+        return "cuda"
+    elif torch.version.hip is not None:
+        return "rocm"
+    elif torch.backends.mps.is_available():
+        return "metal"
+    elif hasattr(torch.version, "xpu") and torch.version.xpu is not None:
+        return "xpu"
+    else:
+        return "cpu"
+def _find_ops_name() -> str:
+    kernel_name = "liger_kernels"
+    unique_id = "4d9f798"
+    backend = get_backend()
+    return f"_{kernel_name}_{backend}_{unique_id}"
+_OPS_NAME = _find_ops_name()
+ops = getattr(torch.ops, _OPS_NAME)
+def add_op_namespace_prefix(op_name: str) -> str:
+    """
+    Prefix op by namespace.
+    """
+    return f"{_OPS_NAME}::{op_name}"

build/torch-rocm/cross_entropy.py ADDED Viewed

	@@ -0,0 +1,558 @@

+import operator
+from typing import Optional
+import torch
+import triton
+import triton.language as tl
+from .utils import compare_version
+from .utils import element_mul_kernel
+from .utils import is_hip
+from .utils import infer_device
+from .utils import is_npu_available
+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
+    except ModuleNotFoundError:
+        # for working with NGC containers
+        from triton.language.extra.cuda.libdevice import tanh
+else:
+    from triton.language.math import tanh
+@triton.jit
+def liger_cross_entropy_kernel(
+    X_ptr,
+    X_stride,
+    Y_ptr,
+    Y_stride,
+    weight_ptr,
+    loss_ptr,
+    z_loss_ptr,
+    loss_stride,
+    token_accuracy_ptr,
+    token_accuracy_stride,
+    predicted_tokens_ptr,
+    predicted_tokens_stride,
+    n_cols,
+    n_non_ignore,
+    sum_non_ignore_weight,
+    weight_sum,
+    ignore_index,
+    lse_square_scale: tl.constexpr,
+    label_smoothing: tl.constexpr,
+    reduction: tl.constexpr,  # set it as constexpr since reduction is always known at compile time
+    softcap,
+    RETURN_Z_LOSS: tl.constexpr,
+    RETURN_TOKEN_ACCURACY: tl.constexpr,
+    RETURN_PREDICTED_TOKENS: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    HAS_WEIGHT: tl.constexpr,
+    HAS_SOFTCAPPING: tl.constexpr,
+    HAS_GRADIENTS: tl.constexpr,
+):
+    """
+    This kernel computes both cross entropy loss and the gradient of the input.
+    We only consider hard label + mean reduction for now. Please refer to https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html for the math.
+    Parameters:
+    X_ptr: Pointer to input tensor.
+    X_stride (int): The stride of the input tensor.
+    Y_ptr: Pointer to target tensor.
+    Y_stride (int): The stride of the target tensor.
+    weight_ptr: Pointer to weight tensor.
+    loss_ptr: Pointer to tensor to store the loss.
+    z_loss_ptr: Pointer to tensor to store the z loss. No operation if RETURN_Z_LOSS is 0.
+    loss_stride (int): The stride of the loss tensor.
+    token_accuracy_ptr: Pointer to tensor to store the per-token accuracy. No operation if RETURN_TOKEN_ACCURACY is 0.
+    token_accuracy_stride (int): The stride of the token accuracy tensor.
+    n_cols (int): The number of columns in the input tensor.
+    n_non_ignore (float): The number of non-ignored elements in the batch.
+    sum_non_ignore_weight (float): The sum of non-ignored target's weights in the batch.
+    weight_sum (float): The sum of weight tensor.
+    ignore_index (int): The index to ignore in the target.
+    label_smoothing (float): The amount of smoothing when computing the loss, where 0.0 means no smoothing.
+    lse_square_scale (float): The scaler of (logsumexp(_input)) ^ 2 adding to the loss for the stability of training.
+    reduction (str): The string for the reduction to apply
+    softcap (float): The upper threshold for scaling logits to the range (-softcap, +softcap).
+    RETURN_Z_LOSS (int): The boolean value to decide whether to store z loss to z_loss_ptr or not. It must be 0 or 1.
+    RETURN_TOKEN_ACCURACY (int): The boolean value to decide whether to store per-token accuracy to token_accuracy_ptr or not. It must be 0 or 1.
+    BLOCK_SIZE (int): The block size for Triton operations.
+    HAS_WEIGHT (bool): The boolean value to determine whether assigning weight to each of the classes.
+    HAS_SOFTCAPPING (bool): The boolean value to determine whether applying soft-capping or not.
+    HAS_GRADIENTS (bool): The boolean value to determine whether calculating gradients in forward pass.
+    """
+    # https://github.com/triton-lang/triton/issues/1058
+    # If B*T*V is too large, program_id * stride will overflow out of int32, so we convert to int64
+    program_id = tl.program_id(0).to(tl.int64)
+    # 1. Load Y_ptr first because if the target is ignore_index, we can return right away
+    Y_ptr += program_id * Y_stride
+    y = tl.load(Y_ptr)
+    # 2. locate the start index
+    X_ptr += program_id * X_stride
+    if y == ignore_index:
+        # set all X_ptr as 0
+        for i in range(0, n_cols, BLOCK_SIZE):
+            X_offsets = i + tl.arange(0, BLOCK_SIZE)
+            tl.store(X_ptr + X_offsets, 0.0, mask=X_offsets < n_cols)
+        # For ignored tokens, set token accuracy to 0
+        if RETURN_TOKEN_ACCURACY:
+            token_accuracy_ptr += program_id * token_accuracy_stride
+            tl.store(token_accuracy_ptr, 0.0)
+        if RETURN_PREDICTED_TOKENS:
+            predicted_tokens_ptr += program_id * predicted_tokens_stride
+            tl.store(predicted_tokens_ptr, -1)
+        return
+    loss_ptr += program_id * loss_stride
+    if RETURN_Z_LOSS:
+        z_loss_ptr += program_id * loss_stride
+    if RETURN_TOKEN_ACCURACY:
+        token_accuracy_ptr += program_id * token_accuracy_stride
+    if RETURN_PREDICTED_TOKENS:
+        predicted_tokens_ptr += program_id * predicted_tokens_stride
+    if HAS_WEIGHT:
+        weight_y = tl.load(weight_ptr + y).cast(tl.float32)
+    # Online softmax: 2 loads + 1 store (compared with 3 loads + 1 store for the safe softmax)
+    # Refer to Algorithm 3 in the paper: https://arxiv.org/pdf/1805.02867
+    # 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
+    argmax_idx = 0  # Track the index of the maximum value for token accuracy / predicted tokens computation
+    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)
+    # Label smoothing is a general case of normal cross entropy
+    # See the full derivation at https://github.com/linkedin/Liger-Kernel/pull/198#issue-2503665310
+    scaled_x_sum = 0.0
+    eps = label_smoothing / n_cols
+    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"),
+            # 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)
+        # Track argmax for accuracy / predicted tokens computation
+        if RETURN_TOKEN_ACCURACY or RETURN_PREDICTED_TOKENS:
+            # Find the index of the maximum value in this block
+            is_max_mask = X_block == block_max
+            # Mask out invalid indices with a value larger than n_cols
+            masked_offsets = tl.where(is_max_mask, X_offsets, n_cols)
+            # Get the first (smallest) index where max occurs
+            current_block_argmax_idx = tl.min(masked_offsets)
+            is_new_max = block_max > m
+            argmax_idx = tl.where(is_new_max, current_block_argmax_idx, argmax_idx)
+        if label_smoothing > 0:
+            # scale X beforehand to avoid overflow
+            if HAS_WEIGHT:
+                weight_block = tl.load(weight_ptr + X_offsets, mask=X_offsets < n_cols)
+                scaled_x_sum += tl.sum(tl.where(X_offsets < n_cols, -eps * X_block * weight_block, 0.0))
+            else:
+                scaled_x_sum += tl.sum(tl.where(X_offsets < n_cols, -eps * X_block, 0.0))
+        m_new = tl.maximum(m, block_max)
+        d = d * tl.exp(m - m_new) + tl.sum(tl.exp(X_block - m_new))
+        m = m_new
+    # log (sum(e^(X_i))) = log (sum(e ^ (max(X) * e ^ (X_i - max(X)))))
+    #                    = log (e^(max(X)) * sum(e ^ (X_i - max(X))))
+    #                    = max(X) + log (sum(e ^ (X_i - max(X)))) = m + log d
+    lse = m + tl.log(d)
+    # 4. [Online Softmax] Second pass: compute gradients
+    # For 'mean' reduction, gradients are normalized by number of non-ignored elements (N)
+    # dx_y = (softmax(x_y) - 1) / N
+    # dx_i = softmax(x_i) / N, i != y
+    # For label smoothing:
+    # dx_i = (softmax(x_i) - label_smoothing / V) / N, V = n_cols, i != y
+    # dx_y = (softmax(x_y) - label_smoothing / V - (1 - label_smoothing)) / N
+    #      = dx_i - (1 - label_smoothing) / N
+    # With Z loss:
+    # dx_i = ((1 + 2 * lse_square_scale * lse) * softmax(x_i) - label_smoothing / V) / N, i != y
+    # dx_y = dx_i - (1 - label_smoothing) / N
+    # For 'sum' reduction, no normalization is applied:
+    # dx_y = softmax(x_y) - 1
+    # dx_i = softmax(x_i), for i ≠ y
+    if HAS_GRADIENTS:
+        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"),
+                # Ensure float32 precision for softmax calculation
+            ).cast(tl.float32)
+            if HAS_SOFTCAPPING:
+                intermediate = tanh(X_block / softcap)
+                X_block = softcap * intermediate
+            if not HAS_WEIGHT:
+                # softmax(x_i)
+                X_block = tl.exp(X_block - m) / d
+                # derivative of z-loss: 2 * lse_square_scale * lse * softmax(x_i)
+                X_block += 2 * lse_square_scale * lse * X_block
+                # smoothing term
+                X_block += -eps
+                # special handle dx_y
+                X_block = tl.where(X_offsets != y, X_block, X_block - (1 - label_smoothing))
+                # reduction scale
+                if reduction == "mean":
+                    X_block = X_block / n_non_ignore
+            else:
+                weight_block = tl.load(weight_ptr + X_offsets, mask=X_offsets < n_cols)
+                softmax_X = tl.exp(X_block - m) / d
+                # derivative of original_loss
+                dloss_ori = (1 - label_smoothing) * softmax_X
+                # specially handle dx_y
+                dloss_ori = tl.where(X_offsets != y, dloss_ori, dloss_ori - (1 - label_smoothing))
+                dloss_ori = dloss_ori * weight_y
+                # derivative of smooth_loss
+                dloss_smooth = eps * (-weight_block + softmax_X * weight_sum)
+                # derivative of z-loss
+                dz_loss = 2 * lse_square_scale * lse * softmax_X
+                # reduction scale
+                if reduction == "mean":
+                    dloss_ori = dloss_ori / sum_non_ignore_weight
+                    dloss_smooth = dloss_smooth / sum_non_ignore_weight
+                    # TODO: Implement weighted z_loss. Currently, z_loss is not scaled by weight.
+                    dz_loss = dz_loss / n_non_ignore
+                # derivative of total_loss
+                X_block = dloss_ori + dloss_smooth + dz_loss
+            # chain rule softcapping
+            # d(softcap * tanh(x / softcap)) = (1 - tanh^2(x / softcap))
+            if HAS_SOFTCAPPING:
+                X_block = X_block * (1 - intermediate * intermediate)
+            tl.store(X_ptr + X_offsets, X_block, mask=X_offsets < n_cols)
+    # We need tl.debug_barrier() to ensure the new result of X_ptr is written as mentioned in
+    # https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/ops/cross_entropy.py#L34
+    tl.debug_barrier()
+    # 5. Calculate the loss
+    # loss = log (softmax(X_y)) = log ((e ^ (X_y - max(X)) / sum(e ^ (X - max(X))))
+    #      = (X_y - max(X)) - log(sum(e ^ (X - max(X))))
+    #      = X_y - m - log d = X_y - lse
+    # sum(e ^ (X - max(X))) must >= 1 because the max term is e ^ 0 = 1
+    # So we can safely calculate log (softmax(X_y)) without overflow
+    loss = lse - ori_X_y
+    if HAS_WEIGHT:
+        loss = weight_y * loss
+    # Original loss = H(q, p),  with label smoothing regularization = H(q', p) and (label_smoothing / V) = eps
+    # H(q', p) = (1 - label_smoothing) * H(q, p) + label_smoothing * H(u, p)
+    #          = (1 - label_smoothing) * H(q, p) + eps * sum(logsoftmax(x_i))
+    # By using m (global max of xi) and d (sum of e^(xi-m)), we can simplify as:
+    #          = (1 - label_smoothing) * H(q, p) + (sum(-eps * x_i) + label_smoothing * (m + logd))
+    # Refer to H(q', p) in section 7 of the paper: https://arxiv.org/pdf/1512.00567
+    # pytorch: https://github.com/pytorch/pytorch/blob/2981534f54d49fa3a9755c9b0855e7929c2527f0/aten/src/ATen/native/LossNLL.cpp#L516
+    # See full derivation at https://github.com/linkedin/Liger-Kernel/pull/198#issuecomment-2333753087
+    if label_smoothing > 0:
+        if HAS_WEIGHT:
+            smooth_loss = scaled_x_sum + eps * lse * weight_sum
+        else:
+            smooth_loss = scaled_x_sum + label_smoothing * lse
+        loss = loss * (1 - label_smoothing) + smooth_loss
+    # An auxiliary loss, z_loss
+    # Refer to Page14 Loss function section in the paper PaLM: https://www.jmlr.org/papers/v24/22-1144.html
+    z_loss = lse_square_scale * lse * lse
+    # Normalize the loss by the number of non-ignored elements if reduction is "mean"
+    if reduction == "mean":
+        if HAS_WEIGHT:
+            loss = loss / sum_non_ignore_weight
+        else:
+            loss = loss / n_non_ignore
+        # TODO: Implement weighted z_loss. Currently, z_loss is not scaled by weight.
+        z_loss = z_loss / n_non_ignore
+    loss += z_loss
+    tl.store(loss_ptr, loss)
+    if RETURN_Z_LOSS:
+        tl.store(z_loss_ptr, z_loss)
+    if RETURN_TOKEN_ACCURACY:
+        # Store 1.0 if prediction is correct, 0.0 otherwise
+        is_correct = 1.0 if argmax_idx == y else 0.0
+        tl.store(token_accuracy_ptr, is_correct)
+    if RETURN_PREDICTED_TOKENS:
+        tl.store(predicted_tokens_ptr, argmax_idx)
+# The hard limit of TRITON_MAX_TENSOR_NUMEL is 1048576 https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/language/core.py#L19
+# However, setting limit as 65536 as in LayerNorm tutorial is faster because of less register spilling
+# The optimal maximum block size depends on your hardware, your kernel, and your dtype
+# the best size we found by manually tuning on xpu and npu.
+if infer_device() == "xpu":
+    MAX_FUSED_SIZE = 4096
+elif infer_device() == "npu":
+    MAX_FUSED_SIZE = 2048
+else:
+    MAX_FUSED_SIZE = 65536 // 2
+def cross_entropy_forward(
+    _input,
+    target,
+    weight,
+    ignore_index,
+    lse_square_scale,
+    label_smoothing,
+    reduction,
+    softcap,
+    return_z_loss,
+    return_token_accuracy=False,
+    return_predicted_tokens=False,
+):
+    assert isinstance(return_z_loss, bool), f"return_z_loss must be True or False. Got: {return_z_loss}"
+    assert isinstance(return_token_accuracy, bool), (
+        f"return_token_accuracy must be True or False. Got: {return_token_accuracy}"
+    )
+    assert isinstance(return_predicted_tokens, bool), (
+        f"return_predicted_tokens must be True or False. Got: {return_predicted_tokens}"
+    )
+    BT, V = _input.shape
+    n_rows = BT
+    BLOCK_SIZE = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
+    # unreduced loss
+    loss_1d = torch.zeros(n_rows, dtype=_input.dtype, device=_input.device)
+    z_loss_1d = torch.zeros(n_rows, dtype=_input.dtype, device=_input.device) if return_z_loss else None
+    token_accuracy_1d = (
+        torch.zeros(n_rows, dtype=torch.float32, device=_input.device) if return_token_accuracy else None
+    )
+    predicted_tokens_1d = (
+        torch.full((n_rows,), -1, dtype=torch.int64, device=_input.device) if return_predicted_tokens else None
+    )
+    target_mask = target != ignore_index
+    n_non_ignore = target_mask.sum().item()
+    assert (target * target_mask).max() < _input.shape[-1], (
+        f"Target {target.max()} is out of bounds. Expected < {_input.shape[-1]}"
+    )
+    assert (target * target_mask).min() >= 0, f"Target {target.min()} is out of bounds. Expected >= 0"
+    sum_non_ignore_weight = n_non_ignore
+    weight_sum = 0.0
+    if weight is not None:
+        assert weight.shape[0] == V, f"If given, weight has to be a Tensor of size V. Got: {weight.shape}"
+        assert torch.is_floating_point(weight), (
+            f"If given, weight has to be a Tensor of floating point dtype. Got: {weight.dtype}"
+        )
+        sum_non_ignore_weight = torch.gather(weight, dim=0, index=target.masked_select(target_mask)).sum().item()
+        weight_sum = weight.sum().item()
+        # ensure weight is contiguous
+        if weight.stride(-1) != 1:
+            weight = weight.contiguous()
+    # ensure _input and target are contiguous in the last dimension
+    if _input.stride(-1) != 1:
+        _input = _input.contiguous()
+    if target.stride(-1) != 1:
+        target = target.contiguous()
+    # Here we use a trick to store X_ptr gradient in X_ptr so we can save memory
+    liger_cross_entropy_kernel[(n_rows,)](
+        X_ptr=_input,
+        X_stride=_input.stride(-2),
+        Y_ptr=target,
+        Y_stride=target.stride(-1),  # always 1
+        weight_ptr=weight,  # dummy if None
+        loss_ptr=loss_1d,
+        z_loss_ptr=z_loss_1d,
+        loss_stride=loss_1d.stride(-1),  # always 1
+        token_accuracy_ptr=token_accuracy_1d,
+        token_accuracy_stride=token_accuracy_1d.stride(-1)
+        if return_token_accuracy
+        else 0,  # always 1 if accuracy is enabled
+        predicted_tokens_ptr=predicted_tokens_1d,
+        predicted_tokens_stride=predicted_tokens_1d.stride(-1)
+        if return_predicted_tokens
+        else 0,  # always 1 if predicted tokens is enabled
+        n_cols=V,
+        n_non_ignore=n_non_ignore,
+        sum_non_ignore_weight=sum_non_ignore_weight,
+        ignore_index=ignore_index,
+        weight_sum=weight_sum,
+        lse_square_scale=lse_square_scale,
+        label_smoothing=label_smoothing,
+        reduction=reduction,
+        softcap=softcap,
+        RETURN_Z_LOSS=return_z_loss,
+        RETURN_TOKEN_ACCURACY=return_token_accuracy,
+        RETURN_PREDICTED_TOKENS=return_predicted_tokens,
+        BLOCK_SIZE=BLOCK_SIZE,
+        HAS_WEIGHT=True if weight is not None else False,
+        HAS_SOFTCAPPING=True if softcap is not None else False,
+        HAS_GRADIENTS=_input.requires_grad,
+        # TODO: 32 seems to give the best performance
+        # Performance is quite sensitive to num_warps
+        num_warps=32 if not is_hip() else 16,
+    )
+    if reduction == "none":
+        loss = loss_1d
+        z_loss = z_loss_1d if return_z_loss else None
+        token_accuracy = token_accuracy_1d if return_token_accuracy else None
+    else:
+        loss = torch.sum(loss_1d)
+        z_loss = torch.sum(z_loss_1d) if return_z_loss else None
+        # For accuracy, we compute the mean across all non-ignored tokens
+        token_accuracy = torch.sum(token_accuracy_1d) / n_non_ignore if return_token_accuracy else None
+    predicted_tokens = predicted_tokens_1d if return_predicted_tokens else None
+    return loss, z_loss, token_accuracy, predicted_tokens, _input
+def cross_entropy_backward(_input, grad_output):
+    # If cross entropy is the last layer, grad_output is 1.0. Skip the mul to save time
+    if torch.equal(grad_output, torch.tensor(1.0, device=grad_output.device)):
+        pass
+    # If reduction is 'none'
+    elif grad_output.ndim > 0:
+        _input = _input * grad_output.unsqueeze(dim=1)
+    # If reduction is ['mean', 'sum'], grad_output is just a scalar
+    # We use a Triton kernel instead of a PyTorch operation because modifying inputs in-place
+    # for gradient storage and backward multiple times causes anomalies with PyTorch but not with Triton.
+    else:
+        BT, V = _input.shape
+        n_rows = BT
+        BLOCK_SIZE = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
+        element_mul_kernel[(n_rows,)](
+            _input,
+            _input.stride(-2),
+            grad_output,
+            V,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=32 if not is_hip() else 16,
+        )
+    return _input
+class LigerCrossEntropyFunction(torch.autograd.Function):
+    """
+    This class implements a custom autograd function for the Liger Cross Entropy loss.
+    It overrides the forward and backward methods of the torch.autograd.Function class.
+    """
+    @staticmethod
+    def forward(
+        ctx,
+        _input: torch.Tensor,
+        target: torch.Tensor,
+        weight: Optional[torch.FloatTensor],
+        ignore_index: int = -100,
+        lse_square_scale: float = 0.0,
+        label_smoothing: float = 0.0,
+        reduction: str = "mean",
+        softcap: Optional[float] = None,
+        return_z_loss: bool = False,
+        return_token_accuracy: bool = False,
+        return_predicted_tokens: bool = False,
+    ):
+        """
+        The forward pass of the Liger Cross Entropy loss.
+        Parameters:
+        ctx : The context object.
+        _input (tensor): The input tensor of shape (BT, V) where B is batch size, T is sequence length, V is vocab size.
+        target (tensor): The target tensor of shape (BT) where each value is in [0, V-1].
+        weight(Tensor, optional): a manual rescaling weight given to each class. If given, has to be a Tensor of size V and floating point dtype
+        ignore_index (int): The index to ignore in the target.
+        lse_square_scale (float): The scaler of (logsumexp(_input)) ^ 2 adding to the loss for the stability of training.
+        label_smoothing (float): The amount of smoothing when computing the loss, where 0.0 means no smoothing.
+        reduction (str): The reduction to apply to the output: "none" | "mean | "sum".
+        softcap (Optional[float]): The upper threshold for scaling logits to the range (-softcap, +softcap).
+        return_z_loss (bool): When `return_z_loss` is `True`, returns (loss, z_loss, token_accuracy, predicted_tokens) instead of (loss, None, None, None). Default: `False`
+        return_token_accuracy (bool): When `return_token_accuracy` is `True`, computes and returns per-token accuracy without materializing logits. Default: `False`
+        return_predicted_tokens (bool): When `return_predicted_tokens` is `True`, returns per-token predicted class indices (argmax) without materializing logits. Default: `False`
+        Returns:
+        tuple: A tuple with the computed losses, accuracy, and predicted tokens: (loss, z_loss, token_accuracy, predicted_tokens). z_loss, token_accuracy, and predicted_tokens are None if not requested.
+        """
+        input_requires_grad = _input.requires_grad
+        loss, z_loss, token_accuracy, predicted_tokens, _input = cross_entropy_forward(
+            _input,
+            target,
+            weight,
+            ignore_index,
+            lse_square_scale,
+            label_smoothing,
+            reduction,
+            softcap,
+            return_z_loss,
+            return_token_accuracy,
+            return_predicted_tokens,
+        )
+        # TODO: investigation
+        # If we don't detach the _input tensor, the memory will double
+        # Not sure why but seems that there will be a time both grad and value exist but in different location
+        if input_requires_grad:
+            ctx.save_for_backward(_input.detach())
+        ctx.return_z_loss = return_z_loss
+        ctx.return_token_accuracy = return_token_accuracy
+        ctx.return_predicted_tokens = return_predicted_tokens
+        return loss, z_loss, token_accuracy, predicted_tokens
+    @staticmethod
+    def backward(ctx, grad_output, grad_output2, grad_output3, grad_output4):
+        """
+        The backward pass of the Liger Cross Entropy loss.
+        Parameters:
+        ctx : The context object with saved tensors.
+        grad_output (tensor): The tensor containing the gradient of the loss with respect to the output.
+        grad_output2 (tensor): No use. Gradient for z_loss (not used as z_loss is only for logging).
+        grad_output3 (tensor): No use. Gradient for token_accuracy (not used as token_accuracy is only for metrics).
+        grad_output4 (tensor): No use. Gradient for predicted_tokens (not used as predicted_tokens is only for metrics).
+        Returns:
+        tuple: A tuple with the gradients with respect to the inputs. The elements are tensors or None.
+        """
+        if ctx.return_z_loss:
+            del grad_output2  # z_loss is only for logging
+        if ctx.return_token_accuracy:
+            del grad_output3  # token_accuracy is only for metrics
+        if ctx.return_predicted_tokens:
+            del grad_output4  # predicted_tokens is only for metrics
+        (_input,) = ctx.saved_tensors
+        _input = cross_entropy_backward(_input, grad_output)
+        return (
+            _input,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+        )

build/torch-rocm/dyt.py ADDED Viewed

	@@ -0,0 +1,164 @@

+import operator
+import torch
+import triton
+import triton.language as tl
+from .utils import compare_version
+from .utils import ensure_contiguous
+from .utils import get_npu_core_count
+from .utils import infer_device
+from .utils import is_npu_available
+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
+    except ModuleNotFoundError:
+        # for working with NGC containers
+        from triton.language.extra.cuda.libdevice import tanh
+else:
+    from triton.language.math import tanh
+@triton.autotune(
+    configs=[
+        triton.Config({"BLOCK_N": bn}, num_stages=ns, num_warps=nw)
+        for bn in [1024, 2048, 4096]
+        for ns in [1, 2]
+        for nw in [4, 8, 16]
+    ],
+    key=["N"],
+)
+@triton.jit
+def _dyt_fwd_kernel(X, Y, Alpha, Gamma, Beta, HAVE_BETA: tl.constexpr, N: tl.constexpr, BLOCK_N: tl.constexpr):
+    col = tl.cast(tl.program_id(0), tl.int64) * BLOCK_N + tl.arange(0, BLOCK_N)
+    mask = col < N
+    row_id = tl.cast(tl.program_id(1), tl.int64)
+    X += row_id * N
+    Y += row_id * N
+    alpha = tl.load(Alpha).to(tl.float32)
+    gamma = tl.load(Gamma + col, mask=mask, other=0.0).to(tl.float32)
+    x = tl.load(X + col, mask=mask, other=0.0).to(tl.float32)
+    tanh_x = tanh(alpha * x)
+    y = tanh_x * gamma
+    if HAVE_BETA:
+        beta = tl.load(Beta + col, mask=mask, other=0.0).to(tl.float32)
+        y += beta
+    tl.store(Y + col, y, mask=mask)
+@triton.autotune(
+    configs=[
+        triton.Config({"BLOCK_N": bn}, num_stages=ns, num_warps=nw)
+        for bn in [1024, 2048, 4096]
+        for ns in [1, 2]
+        for nw in [4, 8, 16]
+    ],
+    key=["N"],
+    # DA is indexed by program_id(0), so different BLOCK_N configs write to
+    # different slot counts per SM. Autotune trials don't zero outputs between
+    # runs, so stale slots from a prior trial would leak into da.sum(). Reset
+    # DA between trials to isolate each config's writes.
+    reset_to_zero=["DA"],
+)
+@triton.jit
+def _dyt_bwd_kernel(
+    DY, DX, DA, DG, DB, X, Alpha, Gamma, HAVE_BETA: tl.constexpr, M, N: tl.constexpr, BLOCK_N: tl.constexpr
+):
+    col = tl.cast(tl.program_id(0), tl.int64) * BLOCK_N + tl.arange(0, BLOCK_N)
+    mask = col < N
+    start_row_id = tl.cast(tl.program_id(1), tl.int64)
+    alpha = tl.load(Alpha).to(tl.float32)
+    da = 0.0
+    gamma = tl.load(Gamma + col, mask=mask, other=0.0).to(tl.float32)
+    dg = tl.zeros((BLOCK_N,), dtype=tl.float32)
+    if HAVE_BETA:
+        db = tl.zeros((BLOCK_N,), dtype=tl.float32)
+    for row_id in range(start_row_id, M, tl.num_programs(1)):
+        x = tl.load(X + row_id * N + col, mask=mask, other=0.0).to(tl.float32)
+        dy = tl.load(DY + row_id * N + col, mask=mask, other=0.0).to(tl.float32)
+        tanh_x = tanh(alpha * x)
+        if HAVE_BETA:
+            db += dy
+        dg += dy * tanh_x
+        tmp = (1 - tanh_x * tanh_x) * dy * gamma
+        da += tl.sum(x * tmp, 0)
+        dx = alpha * tmp
+        tl.store(DX + row_id * N + col, dx, mask=mask)
+    tl.store(DG + start_row_id * N + col, dg, mask=mask)
+    if HAVE_BETA:
+        tl.store(DB + start_row_id * N + col, db, mask=mask)
+    tl.store(DA + start_row_id * tl.cdiv(N, 512) + tl.program_id(0), da)
+def liger_dyt_fwd(x, alpha, gamma, beta):
+    assert x.is_contiguous()
+    HAVE_BETA = True if beta is not None else False
+    input_shape = x.shape
+    x = x.view(-1, input_shape[-1])
+    M, N = x.shape
+    y = torch.empty_like(x)
+    grid = lambda meta: (triton.cdiv(N, meta["BLOCK_N"]), M)
+    _dyt_fwd_kernel[grid](
+        x,
+        y,
+        alpha,
+        gamma,
+        beta,
+        HAVE_BETA,
+        N,
+    )
+    return y.view(input_shape)
+def liger_dyt_bwd(dy, x, alpha, gamma, beta):
+    assert dy.is_contiguous()
+    input_shape = x.shape
+    x = x.view(-1, input_shape[-1])
+    M, N = x.shape
+    HAVE_BETA = True if beta is not None else False
+    device = infer_device()
+    if device == "cuda":
+        NUM_SMS = torch.cuda.get_device_properties(x.device).multi_processor_count
+    elif device == "xpu":
+        NUM_SMS = torch.xpu.get_device_properties(x.device).gpu_subslice_count
+    elif device == "npu":
+        NUM_SMS = get_npu_core_count()
+    da = torch.zeros(NUM_SMS, triton.cdiv(N, 512), dtype=torch.float32, device=x.device)
+    dg = torch.empty(NUM_SMS, N, dtype=torch.float32, device=x.device)
+    db = torch.empty(NUM_SMS, N, dtype=torch.float32, device=x.device) if HAVE_BETA else None
+    dx = torch.empty_like(dy)
+    grid = lambda meta: (triton.cdiv(N, meta["BLOCK_N"]), NUM_SMS)
+    _dyt_bwd_kernel[grid](dy, dx, da, dg, db, x, alpha, gamma, HAVE_BETA, M, N)
+    if HAVE_BETA:
+        db = db.sum(0).to(x.dtype)
+    dg = dg.sum(0).to(gamma.dtype)
+    da = da.sum().to(x.dtype).unsqueeze(0)
+    return dx.view(input_shape), da, dg, db
+class LigerDyTFunction(torch.autograd.Function):
+    @staticmethod
+    @ensure_contiguous
+    def forward(ctx, x, alpha, gamma, beta):
+        y = liger_dyt_fwd(x, alpha, gamma, beta)
+        ctx.save_for_backward(x, alpha, gamma, beta)
+        return y
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, dy):
+        x, alpha, gamma, beta = ctx.saved_tensors
+        dx, dalpha, dgamma, dbeta = liger_dyt_bwd(dy, x, alpha, gamma, beta)
+        return dx, dalpha, dgamma, dbeta

build/torch-rocm/fused_linear_cross_entropy.py ADDED Viewed

	@@ -0,0 +1,400 @@

+import torch
+import triton
+from .cross_entropy import liger_cross_entropy_kernel
+from .utils import amp_custom_bwd
+from .utils import amp_custom_fwd
+from .utils import element_mul_kernel
+from .utils import is_hip
+from .utils import infer_device
+# The hard limit of TRITON_MAX_TENSOR_NUMEL is 1048576 https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/language/core.py#L19
+# However, setting limit as 65536 as in LayerNorm tutorial is faster because of less register spilling
+# The optimal maximum block size depends on your hardware, your kernel, and your dtype
+MAX_FUSED_SIZE = 2048 if infer_device() == "npu" else 65536 // 2
+def fused_linear_cross_entropy_forward(
+    _input,
+    weight,
+    target,
+    ce_weight=None,
+    bias=None,
+    ignore_index=-100,
+    lse_square_scale=0.0,
+    label_smoothing=0.0,
+    reduction="mean",
+    softcap=None,
+    return_z_loss=False,
+    accum_dtype=None,
+    use_token_scaling=False,
+    return_token_accuracy=False,
+    return_predicted_tokens=False,
+):
+    assert isinstance(return_z_loss, bool), f"return_z_loss must be True or False. Got: {return_z_loss}"
+    assert isinstance(return_token_accuracy, bool), (
+        f"return_token_accuracy must be True or False. Got: {return_token_accuracy}"
+    )
+    assert isinstance(return_predicted_tokens, bool), (
+        f"return_predicted_tokens must be True or False. Got: {return_predicted_tokens}"
+    )
+    device = _input.device
+    input_requires_grad = _input.requires_grad
+    # inputs have shape: BT x H
+    # materialized activations will have shape: BT x V
+    # the increase in memory = BT x V
+    # reduction can be achieved by partitioning the number of tokens BT into smaller chunks.
+    # for ex: if we were to achieve the same memory consumption as BT x H, then the chunk size should be:
+    # inc_factor = (V+H-1)//H, chunk_size = (BT + inc_factor - 1)//inc_factor
+    # for ex: BT = 4096*4, V = 32000, H = 4096 ==> inc_factor = 8, chunk_size = 2048
+    BT, H = _input.shape
+    V = weight.shape[0]
+    BLOCK_SIZE = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
+    inc_factor = triton.cdiv(V, H)  # (V + H - 1) // H
+    chunk_size = triton.next_power_of_2(triton.cdiv(BT, inc_factor))  # (BT + inc_factor - 1) // inc_factor
+    num_chunks = triton.cdiv(BT, chunk_size)  # (BT + chunk_size - 1) // chunk_size
+    grad_input = torch.zeros_like(_input, device=device)
+    # we use fp32 for loss and gradients accumulator
+    if input_requires_grad:
+        if accum_dtype is None:
+            grad_weight = torch.zeros_like(weight, device=device) if weight.requires_grad else None
+            grad_bias = torch.zeros_like(bias, device=device) if bias is not None else None
+        else:
+            grad_weight = torch.zeros_like(weight, dtype=accum_dtype, device=device) if weight.requires_grad else None
+            grad_bias = torch.zeros_like(bias, dtype=accum_dtype, device=device) if bias is not None else None
+    else:
+        grad_weight = None
+        grad_bias = None
+    loss_1d = torch.zeros(BT, dtype=torch.float32, device=device)
+    z_loss_1d = torch.zeros(BT, dtype=_input.dtype, device=_input.device) if return_z_loss else None
+    token_accuracy_1d = torch.zeros(BT, dtype=torch.float32, device=device) if return_token_accuracy else None
+    predicted_tokens_1d = torch.full((BT,), -1, dtype=torch.int64, device=device) if return_predicted_tokens else None
+    # TODO: evaluate how CUDA synchronization caused by .item() affects the speed
+    target_mask = target != ignore_index
+    total_n_non_ignore = target_mask.sum().item()
+    total_sum_non_ignore_ce_weight = total_n_non_ignore
+    ce_weight_sum = 0.0
+    if ce_weight is not None:
+        assert ce_weight.shape[0] == V, f"If given, weight has to be a Tensor of size V. Got: {ce_weight.shape}"
+        assert torch.is_floating_point(ce_weight), (
+            f"If given, weight has to be a Tensor of floating point dtype. Got: {ce_weight.dtype}"
+        )
+        total_sum_non_ignore_ce_weight = (
+            torch.gather(ce_weight, dim=0, index=target.masked_select(target_mask)).sum().item()
+        )
+        ce_weight_sum = ce_weight.sum().item()
+        if ce_weight.stride(-1) != 1:
+            ce_weight = ce_weight.contiguous()
+    for chunk_id in range(num_chunks):
+        start_idx = chunk_id * chunk_size
+        end_idx = min((chunk_id + 1) * chunk_size, BT)
+        _input_chunk = _input[start_idx:end_idx]  # chunk_size x H
+        # when doing matmul, use the original precision
+        logits_chunk = _input_chunk @ weight.t()  # chunk_size x V
+        if bias is not None:
+            logits_chunk = logits_chunk + bias
+        target_chunk = target[start_idx:end_idx]  # chunk_size,
+        n_rows = logits_chunk.shape[0]
+        # Compute predicted probabilities for token scaling if needed
+        if use_token_scaling:
+            # Compute softmax probabilities for scaling
+            # We need to compute this before the cross entropy kernel modifies logits_chunk
+            logits_for_softmax = logits_chunk.detach().clone()  # Detach to avoid gradient flow
+            if softcap is not None:
+                logits_for_softmax = softcap * torch.tanh(logits_for_softmax / softcap)
+            # Compute softmax to get predicted probabilities
+            probs = torch.softmax(logits_for_softmax, dim=-1)
+            # Get predicted probabilities for token scaling, handling ignored targets
+            valid_target_mask = target_chunk != ignore_index
+            valid_targets = target_chunk[valid_target_mask]
+            if len(valid_targets) > 0:
+                # Gather probabilities only for valid targets
+                valid_probs = probs[valid_target_mask]
+                pred_probs_valid = torch.gather(valid_probs, -1, valid_targets.unsqueeze(-1)).squeeze(-1)
+                # Create full tensor with zeros for ignored targets
+                pred_probs = torch.zeros_like(target_chunk, dtype=probs.dtype, device=probs.device)
+                pred_probs[valid_target_mask] = pred_probs_valid
+            else:
+                # All targets are ignored
+                pred_probs = torch.zeros_like(target_chunk, dtype=probs.dtype, device=probs.device)
+            # Store the scaling factors
+            scaling_factors = pred_probs.detach()  # Detach to ensure no gradient flow
+        # unreduced loss
+        loss_1d_slice = loss_1d[start_idx:end_idx]  # chunk_size,
+        z_loss_1d_slice = z_loss_1d[start_idx:end_idx] if return_z_loss else None
+        token_accuracy_1d_slice = token_accuracy_1d[start_idx:end_idx] if return_token_accuracy else None
+        predicted_tokens_1d_slice = predicted_tokens_1d[start_idx:end_idx] if return_predicted_tokens else None
+        # ensure _input and target are contiguous
+        logits_chunk = logits_chunk.contiguous()
+        target_chunk = target_chunk.contiguous()
+        # Here we calculate the gradient of logits_chunk in place so we can save memory.
+        liger_cross_entropy_kernel[(n_rows,)](
+            X_ptr=logits_chunk,
+            X_stride=logits_chunk.stride(-2),
+            Y_ptr=target_chunk,
+            Y_stride=target_chunk.stride(-1),  # always 1
+            weight_ptr=ce_weight,
+            loss_ptr=loss_1d_slice,
+            z_loss_ptr=z_loss_1d_slice,
+            loss_stride=loss_1d_slice.stride(-1),  # always 1
+            token_accuracy_ptr=token_accuracy_1d_slice,
+            token_accuracy_stride=token_accuracy_1d_slice.stride(-1)
+            if return_token_accuracy
+            else 0,  # always 1 if accuracy is enabled
+            predicted_tokens_ptr=predicted_tokens_1d_slice,
+            predicted_tokens_stride=predicted_tokens_1d_slice.stride(-1)
+            if return_predicted_tokens
+            else 0,  # always 1 if predicted tokens is enabled
+            n_cols=V,
+            n_non_ignore=total_n_non_ignore,
+            sum_non_ignore_weight=total_sum_non_ignore_ce_weight,
+            weight_sum=ce_weight_sum,
+            ignore_index=ignore_index,
+            lse_square_scale=lse_square_scale,
+            label_smoothing=label_smoothing,
+            reduction=reduction,
+            softcap=softcap,
+            RETURN_Z_LOSS=return_z_loss,
+            RETURN_TOKEN_ACCURACY=return_token_accuracy,
+            RETURN_PREDICTED_TOKENS=return_predicted_tokens,
+            HAS_WEIGHT=True if ce_weight is not None else False,
+            HAS_SOFTCAPPING=True if softcap is not None else False,
+            HAS_GRADIENTS=input_requires_grad,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=32 if not is_hip() else 16,
+        )
+        # Apply token scaling if requested
+        if use_token_scaling:
+            loss_1d_slice = loss_1d_slice * scaling_factors
+            if return_z_loss:
+                z_loss_1d_slice = z_loss_1d_slice * scaling_factors
+        loss_1d[start_idx:end_idx] = loss_1d_slice
+        if return_z_loss:
+            z_loss_1d[start_idx:end_idx] = z_loss_1d_slice
+        if return_token_accuracy:
+            token_accuracy_1d[start_idx:end_idx] = token_accuracy_1d_slice
+        if return_predicted_tokens:
+            predicted_tokens_1d[start_idx:end_idx] = predicted_tokens_1d_slice
+        grad_logits_chunk = logits_chunk  # chunk_size x V
+        # Apply token scaling to gradients if requested
+        if use_token_scaling:
+            # Expand scaling factors to match gradient dimensions
+            scaling_factors_expanded = scaling_factors.unsqueeze(-1)  # chunk_size x 1
+            grad_logits_chunk = grad_logits_chunk * scaling_factors_expanded
+        if input_requires_grad:
+            grad_input[start_idx:end_idx] = grad_logits_chunk @ weight
+        if grad_weight is not None and input_requires_grad:
+            grad_weight += torch.mm(grad_logits_chunk.t(), _input_chunk).float()
+        if bias is not None and input_requires_grad:
+            torch.add(
+                input=grad_bias,
+                other=grad_logits_chunk.sum(dim=0),
+                out=grad_bias,
+                alpha=1.0,
+            )
+    # Need extra calculations for backward if reduction=='none'. Not supporting reduction='none' now.
+    # if reduction == "none":
+    #     loss = loss_1d
+    #     z_loss = z_loss_1d if return_z_loss else None
+    if reduction == "none":
+        # Return per-token losses
+        loss = loss_1d
+        z_loss = z_loss_1d if return_z_loss else None
+        token_accuracy = token_accuracy_1d if return_token_accuracy else None
+    else:
+        loss = torch.sum(loss_1d)
+        z_loss = torch.sum(z_loss_1d) if return_z_loss else None
+        # For accuracy, we compute the mean across all non-ignored tokens
+        token_accuracy = torch.sum(token_accuracy_1d) / total_n_non_ignore if return_token_accuracy else None
+    predicted_tokens = predicted_tokens_1d if return_predicted_tokens else None
+    # Cast back to original dtype
+    grad_weight = grad_weight.to(weight.dtype) if grad_weight is not None else None
+    grad_bias = grad_bias.to(bias.dtype) if grad_bias is not None else None
+    return loss, z_loss, token_accuracy, predicted_tokens, grad_input, grad_weight, grad_bias
+def fused_linear_cross_entropy_backward(grad_output, grad_input, grad_weight, grad_bias):
+    # If cross entropy is the last layer, grad_output is 1.0. Skip the mul to save time
+    if not torch.equal(grad_output, torch.tensor(1.0, device=grad_output.device)):
+        # We use a Triton kernel instead of a PyTorch operation because modifying inputs in-place
+        # for gradient storage and backward multiple times causes anomalies with PyTorch but not with Triton.
+        BT, H = grad_input.shape
+        n_rows = BT
+        BLOCK_SIZE = min(MAX_FUSED_SIZE, triton.next_power_of_2(H))
+        element_mul_kernel[(n_rows,)](
+            grad_input,
+            grad_input.stride(-2),
+            grad_output,
+            H,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=32 if not is_hip() else 16,
+        )
+        # handle grad_weight
+        if grad_weight is not None:
+            V, H = grad_weight.shape
+            n_rows = V
+            element_mul_kernel[(n_rows,)](
+                grad_weight,
+                grad_weight.stride(-2),
+                grad_output,
+                H,
+                BLOCK_SIZE=BLOCK_SIZE,
+                num_warps=32 if not is_hip() else 16,
+            )
+        if grad_bias is not None:
+            V = grad_bias.shape[0]
+            n_rows = V
+            element_mul_kernel[(n_rows,)](
+                grad_bias,
+                grad_bias.stride(-1),
+                grad_output,
+                1,
+                BLOCK_SIZE=BLOCK_SIZE,
+                num_warps=32 if not is_hip() else 16,
+            )
+    return grad_input, grad_weight, grad_bias
+class LigerFusedLinearCrossEntropyFunction(torch.autograd.Function):
+    @staticmethod
+    @amp_custom_fwd
+    def forward(
+        ctx,
+        _input,
+        weight,
+        target,
+        bias=None,
+        ce_weight=None,
+        ignore_index=-100,
+        lse_square_scale=0.0,
+        label_smoothing=0.0,
+        reduction="mean",
+        softcap=None,
+        return_z_loss: bool = False,
+        accum_dtype=None,
+        use_token_scaling: bool = False,
+        return_token_accuracy: bool = False,
+        return_predicted_tokens: bool = False,
+    ):
+        """
+        Fusing the last linear layer with cross-entropy loss
+            Reference: https://github.com/mgmalek/efficient_cross_entropy
+        Handle the forward and backward pass of the final linear layer via cross-entropy loss by avoiding
+        the materialization of the large logits tensor. Since Cross Entropy Loss is the last layer, we can
+        compute the gradient at the forward pass. By doing so, we don't have to store the _input and target
+        for the backward pass.
+        _input: (B*T, H) where B is batch size, T is sequence length, H is hidden dimension.
+        target: (B*T) where each value is in [0, V-1]
+        weight: (V, H) where V is the number of classes
+        bias: (V) where V is the number of classes
+        ce_weight: a manual rescaling weight given to each class. If given, has to be a Tensor of size V and floating point dtype
+        ignore_index: the index to ignore in the target
+        label_smoothing (float): The amount of smoothing when computing the loss, where 0.0 means no smoothing.
+        reduction: reduction to apply
+        accum_dtype (torch.dtype): the dtype of intermediate result buffers for weight and bias gradient accumulations.
+            Recommended to set `accum_dtype` to higher precision, e.g. `torch.float32`, if the training is unstable with original dtype. Default: `None`, performing accumulations in original dtype
+        use_token_scaling (bool): whether to scale each token's loss by its predicted probability (detached).
+            When True, each token's loss is multiplied by the model's predicted probability for that token's true class.
+            Default: False.
+        return_token_accuracy (bool): When `return_token_accuracy` is `True`, computes and returns per-token accuracy without materializing logits. Default: `False`
+        return_predicted_tokens (bool): When `return_predicted_tokens` is `True`, returns per-token predicted class indices (argmax) without materializing logits. Default: `False`
+        """
+        loss, z_loss, token_accuracy, predicted_tokens, grad_input, grad_weight, grad_bias = (
+            fused_linear_cross_entropy_forward(
+                _input=_input,
+                weight=weight,
+                target=target,
+                bias=bias,
+                ce_weight=ce_weight,
+                ignore_index=ignore_index,
+                lse_square_scale=lse_square_scale,
+                label_smoothing=label_smoothing,
+                reduction=reduction,
+                softcap=softcap,
+                return_z_loss=return_z_loss,
+                accum_dtype=accum_dtype,
+                use_token_scaling=use_token_scaling,
+                return_token_accuracy=return_token_accuracy,
+                return_predicted_tokens=return_predicted_tokens,
+            )
+        )
+        # downcast to dtype and store for backward
+        ctx.save_for_backward(
+            grad_input.detach(),
+            grad_weight.detach() if grad_weight is not None else None,
+            grad_bias.detach() if grad_bias is not None else None,
+        )
+        ctx.return_z_loss = return_z_loss
+        ctx.return_token_accuracy = return_token_accuracy
+        ctx.return_predicted_tokens = return_predicted_tokens
+        return loss, z_loss, token_accuracy, predicted_tokens
+    @staticmethod
+    @amp_custom_bwd
+    def backward(ctx, grad_output, grad_output2, grad_output3, grad_output4):
+        if ctx.return_z_loss:
+            del grad_output2  # z_loss is only for logging
+        if ctx.return_token_accuracy:
+            del grad_output3  # token_accuracy is only for metrics
+        if ctx.return_predicted_tokens:
+            del grad_output4  # predicted_tokens is only for metrics
+        (grad_input, grad_weight, grad_bias) = ctx.saved_tensors
+        grad_input, grad_weight, grad_bias = fused_linear_cross_entropy_backward(
+            grad_output, grad_input, grad_weight, grad_bias
+        )
+        return (
+            grad_input,
+            grad_weight,
+            None,
+            grad_bias,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,  # use_token_scaling
+            None,  # return_token_accuracy
+            None,  # return_predicted_tokens
+        )

build/torch-rocm/geglu.py ADDED Viewed

	@@ -0,0 +1,143 @@

+import operator
+import torch
+import triton
+import triton.language as tl
+from .utils import calculate_settings
+from .utils import compare_version
+from .utils import ensure_contiguous
+from .utils import is_npu_available
+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
+    except ModuleNotFoundError:
+        # for working with NGC containers
+        from triton.language.extra.cuda.libdevice import tanh
+else:
+    from triton.language.math import tanh
+@triton.jit
+def _geglu_tanh_forward_kernel(a, b, c, stride, n_cols: tl.constexpr, BLOCK_SIZE: tl.constexpr):
+    program_id = tl.program_id(0).to(tl.int64)
+    # locate start index
+    a += program_id * stride
+    b += program_id * stride
+    c += program_id * stride
+    col_offsets = 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(dc, a, b, stride, n_cols: tl.constexpr, BLOCK_SIZE: tl.constexpr):
+    program_id = tl.program_id(0).to(tl.int64)
+    # locate start index
+    dc += program_id * stride
+    a += program_id * stride
+    b += program_id * stride
+    col_offsets = 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):
+    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]
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+    _geglu_tanh_forward_kernel[(n_rows,)](
+        a,
+        b,
+        c,
+        c.stride(-2),
+        n_cols=n_cols,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+    )
+    return a, b, c.view(*ori_shape)
+def geglu_backward(a, b, dc):
+    ori_shape = dc.shape
+    n_cols = ori_shape[-1]
+    dc = dc.view(-1, n_cols)
+    n_rows = dc.shape[0]
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+    _geglu_tanh_backward_kernel[(n_rows,)](
+        dc,
+        a,
+        b,
+        dc.stride(-2),
+        n_cols=n_cols,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+    )
+    return a.view(*ori_shape), b.view(*ori_shape)
+class LigerGELUMulFunction(torch.autograd.Function):
+    @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

build/torch-rocm/group_norm.py ADDED Viewed

	@@ -0,0 +1,311 @@

+import operator
+import torch
+import triton
+import triton.language as tl
+from .utils import compare_version
+from .utils import ensure_contiguous
+from .utils import infer_device
+from .utils import is_npu_available
+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
+    except ModuleNotFoundError:
+        # for working with NGC containers
+        from triton.language.extra.cuda.libdevice import rsqrt
+else:
+    from triton.language.math import rsqrt
+if infer_device() == "npu":
+    MAX_FUSED_SIZE = 16384  # 8192
+else:
+    MAX_FUSED_SIZE = 65536
+@triton.jit
+def _group_norm_forward_kernel(
+    Y_ptr,  # pointer to output, shape (n_rows, n_groups, hidden_size)
+    Y_row_stride,  # stride of each row in output
+    Y_col_stride,  # stride of each column in output
+    X_ptr,  # pointer to input, shape (n_rows, n_groups, hidden_size)
+    X_row_stride,  # stride of each row in input
+    X_col_stride,  # stride of each column in input
+    Mean_ptr,  # pointer to mean, shape (n_rows, n_groups)
+    Mean_row_stride,  # stride of each row in mean
+    Mean_col_stride,  # stride of each column in mean
+    RSTD_ptr,  # pointer to rstd, shape (n_rows, n_groups)
+    RSTD_row_stride,  # stride of each row in rstd
+    RSTD_col_stride,  # stride of each column in rstd
+    W_ptr,  # pointer to W
+    B_ptr,  # pointer to B
+    hidden_size,  # hidden size of X
+    channels_per_group,  # the number of channels per group
+    eps,
+    BLOCK_SIZE: tl.constexpr,
+):
+    """
+    References:
+    https://nn.labml.ai/normalization/group_norm/index.html
+    """
+    batch_idx = tl.program_id(0)
+    group_idx = tl.program_id(1)
+    X_ptr += batch_idx * X_row_stride + group_idx * X_col_stride
+    Y_ptr += batch_idx * Y_row_stride + group_idx * Y_col_stride
+    block_range = tl.arange(0, BLOCK_SIZE)
+    # Compute mean and variance using the online algorithm
+    s = 0.0
+    squared_sum = 0.0
+    for i in tl.range(0, hidden_size, BLOCK_SIZE):
+        hidden_size_offsets = i + block_range
+        mask = hidden_size_offsets < hidden_size
+        X = tl.load(X_ptr + hidden_size_offsets, mask=mask, other=0.0)
+        s += tl.sum(X)
+        # X**2
+        squared_sum += tl.sum(X * X)
+    m = s / hidden_size
+    # variance = E[X**2] - E[X]**2
+    variance = (squared_sum / hidden_size) - (m * m)
+    # 1/std
+    rstd = rsqrt(variance + eps)
+    # Normalize — flat loop over full hidden_size (not per-channel)
+    # This avoids the nested channel × per_channel_hidden loop where
+    # BLOCK_SIZE >> hidden_size_per_channel causes massive padding waste.
+    hidden_size_per_channel = hidden_size // channels_per_group
+    for i in tl.range(0, hidden_size, BLOCK_SIZE):
+        hidden_size_offsets = i + block_range
+        mask = hidden_size_offsets < hidden_size
+        X = tl.load(X_ptr + hidden_size_offsets, mask=mask, other=m)
+        # Determine which channel each element belongs to, then load W/B
+        local_channel = hidden_size_offsets // hidden_size_per_channel
+        global_channel = group_idx * channels_per_group + local_channel
+        W = tl.load(W_ptr + global_channel, mask=mask)
+        B = tl.load(B_ptr + global_channel, mask=mask)
+        Y = (X - m) * rstd * W + B
+        tl.store(Y_ptr + 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)
+@triton.jit
+def _group_norm_backward_kernel(
+    X_ptr,  # pointer to input, shape (n_rows, n_channels, hidden_size)
+    X_row_stride,  # stride of each row in input
+    X_col_stride,  # stride of each column in input
+    W_ptr,  # pointer to weights, shape (n_channels)
+    Mean_ptr,  # pointer to mean, shape (n_rows, n_groups)
+    Mean_ptr_row_stride,  # stride of each column in mean
+    Mean_ptr_col_stride,  # stride of each column in mean
+    RSTD_ptr,  # pointer to rstd, shape (n_rows, n_groups)
+    DX_ptr,  # pointer to input grad, shape (n_rows, n_groups, hidden_size)
+    DW_ptr,  # pointer to weights grad, shape (n_channels)
+    DB_ptr,  # pointer to bias grad, shape (n_channels)
+    UPSTREAM_ptr,  # pointer to output grad, shape (n_rows, n_channels, hidden_size)
+    hidden_size: tl.constexpr,  # hidden size
+    channels_per_group: tl.constexpr,  # number of groups in group norm
+    BLOCK_SIZE: tl.constexpr,
+    dtype: tl.constexpr,
+):
+    """
+    References:
+    https://nn.labml.ai/normalization/group_norm/index.html
+    https://github.com/karpathy/llm.c/blob/master/doc/layernorm/layernorm.md
+    The backprop equations are the same for group_norm and layer_norm
+    the only difference here is that we load the Mean, Rstd corresponding to the
+    group we're computing gradients for and the mean and rstd are computed over n-channels
+    so the total number of elements we compute the mean over is num_channels_per_group * hidden_size
+    We also need to load the Weights corresponding to the current channel to compute the gradients.
+    """
+    batch_idx = tl.program_id(0)
+    group_idx = tl.program_id(1)
+    # Move the pointers to the correct batch
+    X_ptr += batch_idx * X_row_stride
+    DX_ptr += batch_idx * X_row_stride
+    UPSTREAM_ptr += batch_idx * X_row_stride
+    # Mean and rstd are the same shape so have the same strides
+    mean = tl.load(Mean_ptr + batch_idx * Mean_ptr_row_stride + group_idx * Mean_ptr_col_stride)
+    rstd = tl.load(RSTD_ptr + batch_idx * Mean_ptr_row_stride + group_idx * Mean_ptr_col_stride)
+    c1 = 0.0
+    c2 = 0.0
+    block_range = tl.arange(0, BLOCK_SIZE)
+    # We need to compute the sum terms of the backprop equations across all channels in the group
+    for channel_idx in range(group_idx * channels_per_group, (group_idx + 1) * channels_per_group):
+        dW = 0.0
+        dB = 0.0
+        # Move the pointers to the correct channel
+        W = tl.load(W_ptr + channel_idx)
+        for i in tl.range(0, hidden_size, BLOCK_SIZE):
+            hidden_size_offsets = i + block_range
+            mask = hidden_size_offsets < hidden_size
+            X = tl.load(
+                X_ptr + channel_idx * X_col_stride + hidden_size_offsets,
+                mask=mask,
+                other=0.0,
+            )
+            UPSTREAM_grad = tl.load(
+                UPSTREAM_ptr + channel_idx * X_col_stride + hidden_size_offsets,
+                mask=mask,
+                other=0.0,
+            )
+            x_hat = (X - mean) * rstd
+            dW += tl.sum(UPSTREAM_grad * x_hat)
+            dB += tl.sum(UPSTREAM_grad)
+            wdy = W * UPSTREAM_grad
+            c1 += tl.sum(x_hat * wdy)
+            c2 += tl.sum(wdy)
+        # Need to ensure additions to the same channel are atomic
+        tl.atomic_add(DW_ptr + channel_idx, dW.to(dtype))
+        tl.atomic_add(DB_ptr + channel_idx, dB.to(dtype))
+    N = hidden_size * channels_per_group
+    c1 = c1 / N
+    c2 = c2 / N
+    for channel_idx in tl.range(group_idx * channels_per_group, (group_idx + 1) * channels_per_group):
+        # Move the pointers to the correct channel
+        W = tl.load(W_ptr + channel_idx)
+        for i in range(0, hidden_size, BLOCK_SIZE):
+            hidden_size_offsets = i + block_range
+            mask = hidden_size_offsets < hidden_size
+            X = tl.load(
+                X_ptr + channel_idx * X_col_stride + hidden_size_offsets,
+                mask=mask,
+                other=0.0,
+            )
+            UPSTREAM_grad = tl.load(
+                UPSTREAM_ptr + channel_idx * X_col_stride + hidden_size_offsets,
+                mask=mask,
+                other=0.0,
+            )
+            x_hat = (X - mean) * rstd
+            wdy = W * UPSTREAM_grad
+            dx = (wdy - (x_hat * c1 + c2)) * rstd
+            tl.store(DX_ptr + channel_idx * X_col_stride + hidden_size_offsets, dx, mask=mask)
+def group_norm_forward(X, num_channels, num_groups, W, B, eps):
+    shape = X.shape
+    batch_size = shape[0]
+    channels_per_group = num_channels // num_groups
+    # Reshape X so that the mean and std are computed across the groups
+    X = X.view(batch_size, num_groups, -1).contiguous()
+    hidden_size = X.shape[-1]
+    BLOCK_SIZE = min(MAX_FUSED_SIZE, triton.next_power_of_2(hidden_size))
+    Y = torch.empty((batch_size, num_groups, hidden_size), dtype=X.dtype, device=X.device)
+    Mean = torch.zeros((batch_size, num_groups), dtype=X.dtype, device=X.device)
+    RSTD = torch.zeros((batch_size, num_groups), dtype=X.dtype, device=X.device)
+    _group_norm_forward_kernel[(batch_size, num_groups)](
+        Y,
+        Y.stride(0),
+        Y.stride(1),
+        X,
+        X.stride(0),
+        X.stride(1),
+        Mean,
+        Mean.stride(0),
+        Mean.stride(1),
+        RSTD,
+        RSTD.stride(0),
+        RSTD.stride(1),
+        W,
+        B,
+        hidden_size,
+        channels_per_group,
+        eps,
+        BLOCK_SIZE=BLOCK_SIZE,
+    )
+    # Return tensors in the original shape
+    return Y.view(*shape), X.view(*shape), Mean, RSTD, BLOCK_SIZE
+def group_norm_backward(dY, X, W, B, Mean, RSTD, num_channels, num_groups):
+    shape = dY.shape
+    batch_size = shape[0]
+    hidden_size = dY.shape[-1]
+    channels_per_group = num_channels // num_groups
+    dY = dY.view(batch_size, num_groups, -1)
+    DX = torch.empty(
+        (batch_size, num_groups, hidden_size * channels_per_group),
+        dtype=X.dtype,
+        device=X.device,
+    )
+    DW = torch.zeros((num_channels), dtype=W.dtype, device=W.device)
+    DB = torch.zeros((num_channels), dtype=B.dtype, device=B.device)
+    triton_dtype = tl.float32 if X.dtype == torch.float32 else tl.bfloat16
+    BLOCK_SIZE = min(MAX_FUSED_SIZE, triton.next_power_of_2(hidden_size))
+    _group_norm_backward_kernel[(batch_size, num_groups)](
+        X,
+        X.stride(0),
+        X.stride(1),
+        W,
+        Mean,
+        Mean.stride(0),
+        Mean.stride(1),
+        RSTD,
+        DX,
+        DW,
+        DB,
+        dY,
+        hidden_size,
+        channels_per_group,
+        BLOCK_SIZE=BLOCK_SIZE,
+        dtype=triton_dtype,
+    )
+    # Return tensors in the original shape
+    return DX.view(*shape), DW, DB
+class LigerGroupNormFunction(torch.autograd.Function):
+    @staticmethod
+    @ensure_contiguous
+    def forward(
+        ctx,
+        X,
+        affine_scaling_weight,
+        affine_shifting_bias,
+        num_channels,
+        num_groups,
+        eps,
+    ):
+        Y, X, Mean, RSTD, BLOCK_SIZE = group_norm_forward(
+            X,
+            num_channels,
+            num_groups,
+            affine_scaling_weight,
+            affine_shifting_bias,
+            eps,
+        )
+        ctx.num_channels = num_channels
+        ctx.num_groups = num_groups
+        ctx.save_for_backward(X, affine_scaling_weight, affine_shifting_bias, Mean, RSTD)
+        return Y
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, dY):
+        X, W, B, Mean, RSTD = ctx.saved_tensors
+        DX, DW, DB = group_norm_backward(dY, X, W, B, Mean, RSTD, ctx.num_channels, ctx.num_groups)
+        return DX, DW, DB, None, None, None

build/torch-rocm/jsd.py ADDED Viewed

	@@ -0,0 +1,201 @@

+from typing import Optional
+import torch
+import triton
+import triton.language as tl
+from .utils import ensure_contiguous
+from .utils import infer_device
+@triton.jit
+def _jsd_kernel(
+    X_ptr,  # input in logspace, X = log Q
+    X_stride,
+    Y_ptr,  # ground truth in logspace, Y = log P
+    Y_stride,
+    loss_ptr,
+    loss_stride,
+    dX_ptr,
+    dX_stride,
+    label_ptr,
+    beta: tl.constexpr,
+    n_non_ignore: int,
+    ignore_index: tl.constexpr,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+    HAS_LABEL: tl.constexpr,
+):
+    # JSD(P || Q) = (KL(P || M) + KL(Q || M)) / 2, M = (1/2) * (P + Q) = (1/2) * (e ^ Y + e ^ X)
+    #             = sum(P * log P + Q * log Q - 2 * M * log M) / 2
+    #             = sum(e ^ Y * Y + e ^ X * X - 2 * M * log M) / 2
+    # grad_x_i = 0.5 * Q * (X - log_M)
+    pid = tl.program_id(0).to(tl.int64)
+    X_ptr += pid * X_stride
+    dX_ptr += pid * dX_stride
+    Y_ptr += pid * Y_stride
+    loss_ptr += pid * loss_stride
+    label_ptr += pid
+    if HAS_LABEL:
+        label = tl.load(label_ptr)
+        if label == ignore_index:
+            for i in range(0, n_cols, BLOCK_SIZE):
+                offsets = i + tl.arange(0, BLOCK_SIZE)
+                tl.store(dX_ptr + offsets, 0.0, mask=offsets < n_cols)
+            return
+    for i in range(0, n_cols, BLOCK_SIZE):
+        offsets = i + tl.arange(0, BLOCK_SIZE)
+        mask = offsets < n_cols
+        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)
+        if beta == 0.0:  # forward KL
+            Y_max = tl.max(Y, axis=0)
+            Y_shifted = Y - Y_max
+            Y_prob = tl.exp(Y_shifted) * tl.exp(Y_max)  # Compensate for the shift
+            loss = Y_prob * (Y - X)
+            dX = -Y_prob
+        elif beta == 1.0:  # reverse KL
+            X_max = tl.max(X, axis=0)
+            X_shifted = X - X_max
+            X_prob = tl.exp(X_shifted) * tl.exp(X_max)  # Compensate for the shift
+            loss = X_prob * (X - Y)
+            dX = loss + X_prob
+        else:
+            max_val = tl.maximum(tl.max(X, axis=0), tl.max(Y, axis=0))
+            X_shifted = X - max_val
+            Y_shifted = Y - max_val
+            # Pre-compute exp(max_val) since it's used twice
+            exp_max = tl.exp(max_val)
+            # Compute exp terms with compensation
+            Q = tl.exp(X_shifted) * exp_max  # = exp(X)
+            P = tl.exp(Y_shifted) * exp_max  # = exp(Y)
+            # Pre-compute common terms
+            beta_P = beta * P
+            one_minus_beta_Q = (1 - beta) * Q
+            M = beta_P + one_minus_beta_Q
+            log_M = tl.log(M)  # No need to compensate as M is already in original scale
+            loss = beta_P * Y + one_minus_beta_Q * X - M * log_M
+            dX = one_minus_beta_Q * (X - log_M)
+        # Pre-compute scaling factor
+        scale = 1.0 / n_non_ignore
+        loss = loss * scale
+        dX = dX * scale
+        tl.store(loss_ptr + offsets, loss, mask=mask)
+        tl.store(dX_ptr + offsets, dX, mask=mask)
+MAX_FUSED_SIZE = 4096 if infer_device() == "xpu" else 65536
+def jsd_forward(_input, target, shift_labels, beta, ignore_index, has_label):
+    BT, V = _input.shape
+    n_rows = BT
+    BLOCK_SIZE = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
+    # non reduction loss
+    loss = torch.zeros(_input.shape, dtype=torch.float32, device=_input.device)
+    dX = torch.empty_like(_input)
+    if has_label:
+        n_non_ignore = (shift_labels != ignore_index).sum().item()
+    else:
+        n_non_ignore = BT
+    _jsd_kernel[(n_rows,)](
+        X_ptr=_input,  # input in logspace, X = log Q
+        X_stride=_input.stride(-2),
+        Y_ptr=target,  # ground truth in logspace, Y = log P
+        Y_stride=target.stride(-2),
+        loss_ptr=loss,
+        loss_stride=loss.stride(-2),
+        dX_ptr=dX,
+        dX_stride=dX.stride(-2),
+        label_ptr=(shift_labels if has_label else torch.empty(1, device=_input.device)),  # dummy ptr if no label
+        beta=beta,
+        n_non_ignore=n_non_ignore,
+        ignore_index=ignore_index,
+        n_cols=V,
+        BLOCK_SIZE=BLOCK_SIZE,
+        HAS_LABEL=has_label,
+    )
+    loss = torch.sum(loss)
+    return loss.to(_input.dtype), dX
+def jsd_backward(dX, grad_output):
+    # If jsd is the last layer, grad_output is 1.0. Skip the mul to save time
+    if torch.equal(grad_output, torch.tensor(1.0, device=grad_output.device)):
+        return dX
+    else:
+        return grad_output * dX
+class LigerJSDFunction(torch.autograd.Function):
+    r"""
+    This class implements the forward and backward pass for the generalized Jensen-Shannon Divergence.
+    .. math::
+        JSD(\beta)(P || Q)
+            = \beta * KLDiv(P || (\beta * P + (1 - \beta) * Q)) + (1 - \beta) * KLDiv(Q || (\beta * P + (1 - \beta) * Q))
+    .. note::
+        As all the other losses in PyTorch, this function expects the first argument,
+        :attr:`_input`, to be the predictions, the output of the student model, in log-space
+        and the second, :attr:`target`, to be the observations, the output of the teacher model, in log-space.
+        This differs from the standard mathematical notation :math:`JSD(P || Q)` where
+        :math:`P` denotes the teacher model and :math:`Q` denotes the student model.
+    """
+    @staticmethod
+    @ensure_contiguous
+    def forward(
+        ctx,
+        _input: torch.Tensor,
+        target: torch.Tensor,
+        shift_labels: Optional[torch.Tensor] = None,
+        beta: float = 0.5,
+        ignore_index: int = -100,
+    ) -> torch.Tensor:
+        """
+        Args:
+            _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 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:
+            loss (torch.Tensor): generalized JSD
+        """
+        has_label = False
+        if shift_labels is not None:
+            assert shift_labels.shape == (_input.shape[0],), (
+                f"the shape of shift_labels must be (BT,). Got: {shift_labels.shape}"
+            )
+            shift_labels = shift_labels.contiguous()
+            has_label = True
+        loss, dX = jsd_forward(_input, target, shift_labels, beta, ignore_index, has_label)
+        ctx.save_for_backward(dX)
+        return loss
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, grad_output: torch.Tensor) -> torch.Tensor:
+        (dX,) = ctx.saved_tensors
+        dX = jsd_backward(dX, grad_output)
+        return (
+            dX,
+            None,
+            None,
+            None,
+            None,
+        )

build/torch-rocm/kl_div.py ADDED Viewed

	@@ -0,0 +1,259 @@

+from typing import Literal
+import torch
+import triton
+import triton.language as tl
+from .utils import ensure_contiguous
+from .utils import is_hip
+from .utils import infer_device
+def get_num_warps(BLOCK_SIZE):
+    num_warps = 4
+    if BLOCK_SIZE >= 32768:
+        num_warps = 32 if not is_hip() else 16
+    elif BLOCK_SIZE >= 8192:
+        num_warps = 16
+    elif BLOCK_SIZE >= 2048:
+        num_warps = 8
+    return num_warps
+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"]
+_REDUCTION_MODE_NONE: tl.constexpr = tl.constexpr(0)
+_REDUCTION_MODE_SUM: tl.constexpr = tl.constexpr(1)
+_REDUCTION_MODE_MEAN: tl.constexpr = tl.constexpr(2)
+_REDUCTION_MODE_BATCHMEAN: tl.constexpr = tl.constexpr(3)
+_str_to_reduction_mode = {
+    "none": _REDUCTION_MODE_NONE.value,
+    "sum": _REDUCTION_MODE_SUM.value,
+    "mean": _REDUCTION_MODE_MEAN.value,
+    "batchmean": _REDUCTION_MODE_BATCHMEAN.value,
+}
+@triton.jit
+def _kldiv_kernel_forward(
+    y_ptr,  # [B, S], prediction ptr, the kernel expects the prediction in log-space
+    y_stride,  # int, prediction stride
+    gt_ptr,  # [B, S], ground truth ptr
+    gt_stride,  # int, ground truth stride
+    loss_ptr,  # [B] or [B, S] if reduction == _REDUCTION_MODE_NONE, output ptr
+    loss_stride,  # int, output stride
+    n_cols,  # int, number of columns in the input tensor
+    eps,
+    BLOCK_SIZE: tl.constexpr,
+    log_target: tl.constexpr = False,
+    reduction: tl.constexpr = _REDUCTION_MODE_BATCHMEAN,
+):
+    pid = tl.program_id(0).to(tl.int64)
+    y_ptr += pid * y_stride
+    gt_ptr += pid * gt_stride
+    loss_ptr += pid * loss_stride
+    base_offsets = tl.arange(0, BLOCK_SIZE)
+    loss_sum = 0.0
+    for i in range(0, n_cols, BLOCK_SIZE):
+        offsets = i + base_offsets
+        mask = offsets < n_cols
+        y = tl.load(y_ptr + offsets, mask=mask, other=0.0)
+        y_true = tl.load(gt_ptr + offsets, mask=mask, other=0.0)
+        # KL(y_true || y) = y_true * (log(y_true) - log(y))
+        # We compute KL(y_true || y) with y in the log-space
+        if not log_target:
+            loss = y_true * (tl.log(tl.maximum(y_true, eps)) - y)
+        else:
+            loss = tl.exp(y_true) * (y_true - y)
+        if reduction == _REDUCTION_MODE_NONE:
+            tl.store(loss_ptr + offsets, loss, mask=mask)
+        else:
+            loss_sum += tl.sum(loss, axis=0)
+    if reduction != _REDUCTION_MODE_NONE:
+        tl.store(loss_ptr, loss_sum)
+@triton.jit
+def _kldiv_kernel_backward(
+    target_ptr,
+    target_stride,
+    new_grads_ptr,
+    new_grads_stride,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+    log_target: tl.constexpr = False,
+):
+    pid = tl.program_id(0).to(tl.int64)
+    target_ptr += pid * target_stride
+    new_grads_ptr += pid * new_grads_stride
+    offsets = tl.arange(0, BLOCK_SIZE)
+    mask = offsets < n_cols
+    for i in range(0, n_cols, BLOCK_SIZE):
+        offsets = i + tl.arange(0, BLOCK_SIZE)
+        mask = offsets < n_cols
+        target = tl.load(target_ptr + offsets, mask=mask, other=0.0)
+        if not log_target:
+            res = target * -1
+        else:
+            res = -tl.exp(target)
+        tl.store(new_grads_ptr + offsets, res, mask=mask)
+def kldiv_forward_triton(y_pred, y_true, log_target, reduction, eps):  # [BT, V]
+    BT, V = y_pred.shape
+    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,)
+    reduction = _str_to_reduction_mode[reduction]
+    out_size = (BT, V) if reduction == _REDUCTION_MODE_NONE.value else (BT,)
+    output_tensor = torch.zeros(out_size, device=y_pred.device, dtype=torch.float32)
+    _kldiv_kernel_forward[grid](
+        y_pred,
+        y_pred.stride(0),
+        y_true,
+        y_true.stride(0),
+        output_tensor,
+        output_tensor.stride(0),
+        V,
+        eps=eps,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+        log_target=log_target,
+        reduction=reduction,
+    )
+    # calculated according to the reduction mode same as in Pytorch. In the later versions, `mean` will be changed to the same behavior as `batchmean`
+    # https://pytorch.org/docs/stable/generated/torch.nn.KLDivLoss.html
+    # https://github.com/pytorch/pytorch/blob/d7b57c4d63edb42e1deeeba9497fcb5f1f748ff2/torch/nn/functional.py#L3372
+    if reduction == _REDUCTION_MODE_BATCHMEAN.value:
+        return output_tensor.sum() / BT
+    elif reduction == _REDUCTION_MODE_SUM.value:
+        return output_tensor.sum(dim=0)
+    elif reduction == _REDUCTION_MODE_MEAN.value:
+        return output_tensor.sum() / (BT * V)
+    else:
+        return output_tensor
+def kldiv_backward_triton(target, grad_output, new_grads, log_target):
+    BT, V = target.shape
+    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,)
+    # We store the gradients in-place in the input tensor
+    _kldiv_kernel_backward[grid](
+        target,
+        target.stride(0),
+        new_grads,
+        new_grads.stride(0),
+        V,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+        log_target=log_target,
+    )
+    # If cross entropy is the last layer, grad_output is 1.0. Skip the mul then.
+    if torch.equal(grad_output, torch.tensor(1.0, device=grad_output.device)):
+        return new_grads
+    return new_grads * grad_output
+class LigerKLDivLossFunction(torch.autograd.Function):
+    """
+    Class implementing the forward and backward pass for the KL Divergence Loss using Triton, as defined by the following formula:
+    ```python
+    if log_target:
+        loss = target.exp() * (target - input)
+    else:
+        loss = target * (target.log() - input)
+    ```,
+    then the loss is reduced according to the `reduction` parameter.
+    as defined in the PyTorch documentation: https://pytorch.org/docs/stable/generated/torch.nn.KLDivLoss.html
+    """
+    @staticmethod
+    @ensure_contiguous
+    def forward(
+        ctx,
+        y_pred: torch.Tensor,
+        y_true: torch.Tensor,
+        reduction: REDUCTION_LITERAL = "batchmean",
+        log_target: bool = False,
+        eps: float = 1e-10,
+    ) -> torch.Tensor:
+        """A forward pass for the KL Divergence Loss.
+        Args:
+            ctx: Torch autograd context
+            y_pred (torch.Tensor): A tensor of shape (BT, V) containing the predicted values, expected to be log-probabilities.
+            y_true (torch.Tensor): A tensor of shape (BT, V) containing the target values, expected to be either probabilities or log-probabilities, depending on the value of `log_target`.
+            reduction (REDUCTION_LITERAL, optional): Reduction to be used. Defaults to "batchmean".
+            log_target (bool, optional): If set to true, expects the ground truth to already be log-probabilities. Defaults to False.
+            eps: (float, optional): A small value to avoid division by zero. Defaults to 1e-10.
+        Returns:
+            torch.Tensor: The computed KL Divergence Loss, with shape (BT, V) if `reduction` is "none", else a scalar.
+        """
+        ctx.save_for_backward(y_true)
+        ctx.reduction = reduction
+        ctx.log_target = log_target
+        return kldiv_forward_triton(y_pred, y_true, log_target=log_target, reduction=reduction, eps=eps)
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, grad_output: torch.Tensor) -> torch.Tensor:
+        """A backward pass for the KL Divergence Loss.
+        Args:
+            ctx: Torch autograd context
+            grad_output (torch.Tensor): The gradient of the loss with respect to the output.
+        Returns:
+            tuple[torch.Tensor, None, None, None, None]: The gradient of the loss with respect to the inputs and None for the other arguments of the forward method.
+        """
+        (y_true,) = ctx.saved_tensors
+        new_grads = torch.empty_like(y_true)
+        derivative = kldiv_backward_triton(y_true, grad_output, new_grads, ctx.log_target)
+        if ctx.reduction == "batchmean":
+            derivative = derivative / y_true.shape[0]
+        elif ctx.reduction == "sum" or ctx.reduction == "none":
+            pass
+        elif ctx.reduction == "mean":
+            derivative = derivative / (y_true.shape[0] * y_true.shape[1])
+        return (
+            derivative,
+            None,
+            None,
+            None,
+            None,
+        )

build/torch-rocm/layer_norm.py ADDED Viewed

	@@ -0,0 +1,320 @@

+import math
+import operator
+import torch
+import triton
+import triton.language as tl
+from .utils import calculate_settings
+from .utils import compare_version
+from .utils import ensure_contiguous
+from .utils import get_npu_core_count
+from .utils import set_large_grf_mode
+from .utils import is_npu_available
+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
+    except ModuleNotFoundError:
+        # for working with NGC containers
+        from triton.language.extra.cuda.libdevice import rsqrt
+else:
+    from triton.language.math import rsqrt
+@triton.jit
+def _layer_norm_forward_kernel(
+    Y_ptr,  # pointer to output, shape (n_rows, n_cols)
+    Y_row_stride,  # stride of each row in output
+    X_ptr,  # pointer to input, shape (n_rows, n_cols)
+    X_row_stride,  # stride of each row in input
+    W_ptr,  # pointer to weights, shape (n_cols,)
+    W_row_stride,  # stride of each row in weights
+    B_ptr,  # pointer to bias, shape (n_cols,)
+    B_row_stride,  # stride of each row in bias
+    Mean_ptr,  # pointer to mean, shape (n_rows,)
+    Mean_row_stride,  # stride of each row in mean
+    RSTD_ptr,  # pointer to rstd, shape (n_rows,)
+    RSTD_row_stride,  # stride of each row in rstd
+    n_cols,
+    eps,
+    BLOCK_SIZE: tl.constexpr,
+):
+    """
+    References:
+    https://arxiv.org/abs/1607.06450
+    https://github.com/karpathy/llm.c/blob/master/doc/layernorm/layernorm.md
+    """
+    row_idx = tl.program_id(0).to(tl.int64)
+    col_offsets = tl.arange(0, BLOCK_SIZE)
+    mask = col_offsets < 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)
+    # 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))
+    # 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
+    # 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)
+    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,
+):
+    """
+    References:
+    https://arxiv.org/abs/1607.06450
+    https://github.com/karpathy/llm.c/blob/master/doc/layernorm/layernorm.md
+    """
+    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)
+    db_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+    # 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)
+    for row_idx in range(row_start, row_end):
+        # Calculate pointers for this specific row
+        row_X_ptr = X_ptr + row_idx * stride_x
+        row_DX_ptr = DX_ptr + row_idx * stride_dx
+        row_DY_ptr = DY_ptr + row_idx * stride_dy
+        row_Mean_ptr = Mean_ptr + row_idx * stride_mean
+        row_RSTD_ptr = RSTD_ptr + row_idx * stride_rstd
+        # 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_f32
+        # Store input gradient
+        tl.store(row_DX_ptr + cols, dx, mask=mask)
+        # 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
+    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]}) "
+            f"must match weight size (W.shape[0]={W.shape[0]})"
+        )
+    # XPU-specific optimization
+    kernel_args = {}
+    if X.device.type == "xpu":
+        set_large_grf_mode(kernel_args)
+    # Launch kernel with one thread block per row for optimal performance
+    grid = (n_rows,)
+    _layer_norm_forward_kernel[grid](
+        Y,
+        Y.stride(0),
+        X,
+        X.stride(0),
+        W,
+        W.stride(0),
+        B,
+        B.stride(0),
+        Mean,
+        Mean.stride(0),
+        RSTD,
+        RSTD.stride(0),
+        n_cols,
+        eps,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+        **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)
+    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_eu_count
+    elif X.device.type == "npu":
+        sm_count = get_npu_core_count()
+    # 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}.")
+    rows_per_program = math.ceil(n_rows / sm_count)
+    grid = (sm_count,)
+    # 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
+    if X.device.type == "xpu":
+        kernel_args.update({"num_warps": 32, "num_stages": 4})
+        set_large_grf_mode(kernel_args)
+    # 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,
+        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,
+        BLOCK_SIZE=BLOCK_SIZE,
+        **kernel_args,
+    )
+    DX = DX.view(*shape)
+    DW = _DW.sum(dim=0).to(W.dtype)
+    DB = _DB.sum(dim=0).to(B.dtype)
+    return DX, DW, DB
+class LigerLayerNormFunction(torch.autograd.Function):
+    @staticmethod
+    @ensure_contiguous
+    def forward(ctx, X, W, B, eps):
+        Y, X, Mean, RSTD, BLOCK_SIZE, num_warps = layer_norm_forward(X, W, B, eps)
+        ctx.save_for_backward(X, W, B, Mean, RSTD)
+        return Y
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, dY):
+        X, W, B, Mean, RSTD = ctx.saved_tensors
+        DX, DW, DB = layer_norm_backward(dY, X, W, B, Mean, RSTD)
+        return DX, DW, DB, None

build/torch-rocm/layers.py ADDED Viewed

	@@ -0,0 +1,342 @@

+import inspect
+from dataclasses import dataclass
+from typing import Optional
+import torch
+import torch.nn as nn
+from .fused_linear_cross_entropy import LigerFusedLinearCrossEntropyFunction
+from .geglu import LigerGELUMulFunction
+from .rms_norm import LigerRMSNormFunction
+from .rope import LigerRopeFunction
+from .swiglu import LigerSiLUMulFunction
+from .tiled_mlp import apply_tiled_mlp
+# NOTE: Not compile-friendly --> large deviations to the original implementation under compile
+class LigerRMSNorm(nn.Module):
+    weight: nn.Parameter
+    variance_epsilon: float
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        return LigerRMSNormFunction.apply(
+            hidden_states,
+            self.weight,
+            self.variance_epsilon,
+            0,
+            "llama",
+            True,
+            None,
+        )
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+class LigerLinear(nn.Module):
+    weight: nn.Parameter
+    bias: nn.Parameter | None
+    can_torch_compile = True
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        def _forward(module, x):
+            return nn.functional.linear(x, module.weight, module.bias)
+        compute_params = [p for p in self.parameters() if p.requires_grad]
+        return apply_tiled_mlp(
+            fn=_forward,
+            mlp_module=self,
+            x=input,
+            num_shards=None,
+            compute_params=compute_params,
+        )
+class LigerSwiGLUMLP(nn.Module):
+    gate_proj: nn.Linear
+    up_proj: nn.Linear
+    down_proj: nn.Linear
+    can_torch_compile = True
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.down_proj(LigerSiLUMulFunction.apply(self.gate_proj(x), self.up_proj(x)))
+class LigerGEGLUMLP(nn.Module):
+    gate_proj: nn.Linear
+    up_proj: nn.Linear
+    down_proj: nn.Linear
+    can_torch_compile = True
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.down_proj(LigerGELUMulFunction.apply(self.gate_proj(x), self.up_proj(x)))
+class LigerTiledSwiGLUMLP(nn.Module):
+    gate_proj: nn.Linear
+    up_proj: nn.Linear
+    down_proj: nn.Linear
+    can_torch_compile = True
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        def _mlp_forward(module, x):
+            """Internal MLP forward function for tiled computation."""
+            gate = module.gate_proj(x)
+            up = module.up_proj(x)
+            return module.down_proj(LigerSiLUMulFunction.apply(gate, up))
+        compute_params = [p for p in self.parameters() if p.requires_grad]
+        return apply_tiled_mlp(
+            fn=_mlp_forward,
+            mlp_module=self,
+            x=x,
+            num_shards=None,
+            compute_params=compute_params,
+        )
+class LigerTiledGEGLUMLP(nn.Module):
+    gate_proj: nn.Linear
+    up_proj: nn.Linear
+    down_proj: nn.Linear
+    can_torch_compile = True
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        def _mlp_forward(module, x):
+            """Internal MLP forward function for tiled computation."""
+            gate = module.gate_proj(x)
+            up = module.up_proj(x)
+            return module.down_proj(LigerGELUMulFunction.apply(gate, up))
+        compute_params = [p for p in self.parameters() if p.requires_grad]
+        return apply_tiled_mlp(
+            fn=_mlp_forward,
+            mlp_module=self,
+            x=x,
+            num_shards=None,
+            compute_params=compute_params,
+        )
+def liger_rotary_pos_emb(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    unsqueeze_dim: int = 1,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Apply standard rotary positional embedding to ``q`` and ``k``."""
+    return LigerRopeFunction.apply(q, k, cos, sin, None, unsqueeze_dim)
+@dataclass
+class CrossEntropyOutput:
+    loss: torch.Tensor
+    z_loss: Optional[torch.Tensor] = None
+    token_accuracy: Optional[torch.Tensor] = None
+    predicted_tokens: Optional[torch.Tensor] = None
+def liger_fused_linear_cross_entropy(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+    target: torch.Tensor,
+    bias: Optional[torch.Tensor] = None,
+    ce_weight: Optional[torch.Tensor] = None,
+    ignore_index: int = -100,
+    lse_square_scale: float = 0.0,
+    label_smoothing: float = 0.0,
+    reduction: str = "mean",
+    softcap: Optional[float] = None,
+    return_z_loss: bool = False,
+    accum_dtype: Optional[torch.dtype] = None,
+    use_token_scaling: bool = False,
+    return_token_accuracy: bool = False,
+    return_predicted_tokens: bool = False,
+):
+    loss, z_loss, token_accuracy, predicted_tokens = LigerFusedLinearCrossEntropyFunction.apply(
+        input,
+        weight,
+        target,
+        bias,
+        ce_weight,
+        ignore_index,
+        lse_square_scale,
+        label_smoothing,
+        reduction,
+        softcap,
+        return_z_loss,
+        accum_dtype,
+        use_token_scaling,
+        return_token_accuracy,
+        return_predicted_tokens,
+    )
+    if not return_z_loss and not return_token_accuracy and not return_predicted_tokens:
+        return loss
+    return CrossEntropyOutput(
+        loss=loss,
+        z_loss=z_loss,
+        token_accuracy=token_accuracy,
+        predicted_tokens=predicted_tokens,
+    )
+# NOTE: We have this intentional graph break as we encounter issues such as IMAs and Cublas errors.
+#       We know that this is an optimized kernel already so there is less ways to
+#       fuse it either way; we rely on torch compile to go through the base model to optimize.
+@torch.compiler.disable
+def LigerForCausalLMLoss(
+    logits: None,  # to match transformers signature
+    labels: torch.Tensor,
+    vocab_size: int,  # to match transformers signature
+    num_items_in_batch: Optional[int] = None,
+    ignore_index: int = -100,
+    shift_labels: Optional[torch.Tensor] = None,
+    hidden_states: torch.Tensor | None = None,
+    lm_head_weight: torch.Tensor | None = None,
+    lm_head_bias: torch.Tensor | None = None,
+    **kwargs,
+):
+    # To match signature we hide these behind the kwargs but we expect a few kwargs to exist
+    hidden_size = kwargs.pop("hidden_size", None)
+    final_logit_softcapping = kwargs.pop("final_logit_softcapping", None)
+    if hidden_size is None or hidden_states is None or lm_head_weight is None:
+        raise ValueError(
+            f"`LigerForCausalLMLoss` requires the LLM's weight (found `{lm_head_weight is not None}`),"
+            f"the last hidden state (found `{hidden_states is not None}`), and the `hidden_size`"
+            f"(found `{hidden_size is not None}`). Please make sure to pass the necessary kwargs."
+        )
+    applicable_params = inspect.signature(liger_fused_linear_cross_entropy).parameters
+    kwargs = {k: v for k, v in kwargs.items() if k in applicable_params}
+    if shift_labels is None:
+        labels = nn.functional.pad(labels, (0, 1), value=ignore_index)
+        shift_labels = labels[..., 1:].contiguous()
+    hidden_states = hidden_states.view(-1, hidden_size)
+    shift_labels = shift_labels.view(-1).to(hidden_states.device)
+    reduction = "sum" if num_items_in_batch is not None else "mean"
+    loss = liger_fused_linear_cross_entropy(
+        hidden_states,
+        lm_head_weight,
+        shift_labels,
+        bias=lm_head_bias,
+        reduction=reduction,
+        ignore_index=ignore_index,
+        softcap=final_logit_softcapping,
+        return_token_accuracy=False,
+        return_predicted_tokens=False,
+        **kwargs,
+    )
+    if reduction == "sum":
+        loss = loss / num_items_in_batch
+    return loss
+# Add torch compile support for functions - in this case it's to allow this to be used
+# but the function itself will not be compiled (see the note at the function)
+liger_rotary_pos_emb.can_torch_compile = True
+LigerForCausalLMLoss.can_torch_compile = True
+class liger_rotary_pos_emb_layer(nn.Module):
+    can_torch_compile = True
+    def forward(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        cos: torch.Tensor,
+        sin: torch.Tensor,
+        unsqueeze_dim: int = 1,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """Apply standard rotary positional embedding to ``q`` and ``k``."""
+        return LigerRopeFunction.apply(q, k, cos, sin, None, unsqueeze_dim)
+class LigerForCausalLMLossLayer(nn.Module):
+    can_torch_compile = True
+    # NOTE: We have this intentional graph break as we encounter issues such as IMAs and Cublas errors.
+    #       We know that this is an optimized kernel already so there is less ways to
+    #       fuse it either way; we rely on torch compile to go through the base model to optimize.
+    @torch.compiler.disable
+    def forward(
+        self,
+        logits: None,  # to match transformers signature
+        labels: torch.Tensor,
+        vocab_size: int,  # to match transformers signature
+        num_items_in_batch: Optional[int] = None,
+        ignore_index: int = -100,
+        shift_labels: Optional[torch.Tensor] = None,
+        hidden_states: torch.Tensor | None = None,
+        lm_head_weight: torch.Tensor | None = None,
+        lm_head_bias: torch.Tensor | None = None,
+        **kwargs,
+    ):
+        # To match signature we hide these behind the kwargs but we expect a few kwargs to exist
+        hidden_size = kwargs.pop("hidden_size", None)
+        final_logit_softcapping = kwargs.pop("final_logit_softcapping", None)
+        if hidden_size is None or hidden_states is None or lm_head_weight is None:
+            raise ValueError(
+                f"`LigerForCausalLMLoss` requires the LLM's weight (found `{lm_head_weight is not None}`),"
+                f"the last hidden state (found `{hidden_states is not None}`), and the `hidden_size`"
+                f"(found `{hidden_size is not None}`). Please make sure to pass the necessary kwargs."
+            )
+        applicable_params = inspect.signature(liger_fused_linear_cross_entropy).parameters
+        kwargs = {k: v for k, v in kwargs.items() if k in applicable_params}
+        if shift_labels is None:
+            labels = nn.functional.pad(labels, (0, 1), value=ignore_index)
+            shift_labels = labels[..., 1:].contiguous()
+        hidden_states = hidden_states.view(-1, hidden_size)
+        shift_labels = shift_labels.view(-1).to(hidden_states.device)
+        reduction = "sum" if num_items_in_batch is not None else "mean"
+        loss = liger_fused_linear_cross_entropy(
+            hidden_states,
+            lm_head_weight,
+            shift_labels,
+            bias=lm_head_bias,
+            reduction=reduction,
+            ignore_index=ignore_index,
+            softcap=final_logit_softcapping,
+            return_token_accuracy=False,
+            return_predicted_tokens=False,
+            **kwargs,
+        )
+        if reduction == "sum":
+            loss = loss / num_items_in_batch
+        return loss
+__all__ = [
+    "LigerRMSNorm",
+    "LigerLinear",
+    "LigerSwiGLUMLP",
+    "LigerGEGLUMLP",
+    "LigerTiledSwiGLUMLP",
+    "LigerTiledGEGLUMLP",
+    "liger_rotary_pos_emb",
+    "liger_rotary_pos_emb_layer",
+    "LigerForCausalLMLoss",
+    "LigerForCausalLMLossLayer",
+]

build/torch-rocm/liger_kernels/__init__.py ADDED Viewed

	@@ -0,0 +1,26 @@

+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch-rocm/metadata.json ADDED Viewed

	@@ -0,0 +1,45 @@

+{
+  "name": "liger-kernels",
+  "id": "_liger_kernels_rocm_4d9f798",
+  "version": 3,
+  "license": "BSD-2-Clause",
+  "python-depends": [],
+  "backend": {
+    "type": "rocm"
+  },
+  "digest": {
+    "algorithm": "sha256",
+    "files": {
+      "__init__.py": "DSZMiK0xOBiMb0JdCc2K9QH4Z6lV7L8D/fTowstf/og=",
+      "_ops.py": "pdeXrc8M9KCvCpe0mtloEEFT3iJWZab9CMNsgT2CK10=",
+      "cross_entropy.py": "LsQN0OI8VHqgUObWF0qqTgNA9z7SYMaZNb2I7FIYnO4=",
+      "dyt.py": "806Tai0T+Yi7Ctf5YGmtwAc87MUjlzjsLky7b1T+oWM=",
+      "fused_linear_cross_entropy.py": "yjtZHWyBcXCBJ57NbbrQ1wMmTr9nARqDBrdjXVYpR74=",
+      "geglu.py": "Lmw0I38IBIHkNROTf9V3bP6PZLktCWCqNAxzVRaipbk=",
+      "group_norm.py": "xdY/tCr4rHOb4PhD827259XB7QDb22z2psUKzTYTKBE=",
+      "jsd.py": "CftyIN36AU63e1GdtuA7hNhs66I3TjS9lttcFriBnGA=",
+      "kl_div.py": "VQG/89mefLifPeVEv+1O8hmsXgGujMLA0JXDwO5QvtU=",
+      "layer_norm.py": "INpxK7Ac27f+eckcMfOOSgZErzDhNF5iGt207eV9aoM=",
+      "layers.py": "+50F9xmnpQXbU1K/lZxdJwbnTotkksJlhoLOWHjo+/M=",
+      "liger_kernels/__init__.py": "DFYPlrhXwYjEqCl/8n0SmWGZV8NFml5DPhMjKfv98GY=",
+      "qwen2vl_mrope.py": "3GExhYpLgB4VUtyZyjRk8XjEur3W4EWF6HQ67ML5vBU=",
+      "rms_norm.py": "GLzIZjDP74QxaeKQ5B/VDXKTGxD6Xo9cVaW/iqmsc4c=",
+      "rope.py": "v+7JHRrv+5ImoROkpKfl30WwWI4qTa2tAl7zQeB4ml4=",
+      "swiglu.py": "TLhyvpr7nbVbMqWbx8uz+iU2GxS6q0QrVxWAd1eQCBo=",
+      "tiled_mlp.py": "PJw2R8YdHHxQRSwbFJWjcHxgzssEqsRWqyogkY8pKR8=",
+      "tvd.py": "VbWUtpJ79E2atqjBoVyTSO785aS+idHiQBIrJwifDAw=",
+      "utils.py": "Y8wQsplz6ed7oPmBfX7yGGcX74VB3t/e5VFvYv/X7wU="
+    }
+  },
+  "provenance": {
+    "kernel-builder": {
+      "version": "0.17.0-dev0",
+      "sha": "81580bb92577f2f7228661ca2a221fc052375709",
+      "dirty": false
+    },
+    "kernel": {
+      "sha": "4d9f798b95ed280b770213048b8d3f41509d0c9e",
+      "dirty": false
+    }
+  }
+}

build/torch-rocm/metadata.json.sigstore ADDED Viewed

	@@ -0,0 +1 @@

+ {"mediaType":"application/vnd.dev.sigstore.bundle.v0.3+json", "verificationMaterial":{"certificate":{"rawBytes":"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"}, "tlogEntries":[{"logIndex":"2167855782", "logId":{"keyId":"wNI9atQGlz+VWfO6LRygH4QUfY/8W4RFwiT5i5WRgB0="}, "kindVersion":{"kind":"hashedrekord", "version":"0.0.1"}, "integratedTime":"1784030571", "inclusionPromise":{"signedEntryTimestamp":"MEQCIFPwhR+59GwljQ+ZxD1rhaNTV3Sh4vPwQeU/9JrQoO2xAiBPUFa1Dck8XvuXhEpd/t74KU9P4qaGczgDzl//4hAk8w=="}, "inclusionProof":{"logIndex":"2045951520", "rootHash":"Re/Ph87BOM7jhvV4VE+3WsyQb72tumWdlqTxAOe0qEM=", "treeSize":"2045951527", "hashes":["nUla+MAuN3xIPh3Fw90o+gJvejDEBaNrSVpncG6U3kM=", "dEagCZIabgWBH6ViR5uwn8xXp3WdvaR3vEy+Q8US9b8=", "4tMZ/2/4pBxGx11gMAczICVOHyul021sgz6LWWWRpIQ=", "m+Vk8U0/4EcekuSwkXtYOJcYVvuBCDXI7XChKnMHSec=", "iMBom1ahASM6ZAxymZoy+ISJTz5W3t4BNlJ+0yiYrnc=", "IuHNjKNGzZgw8Ec1vaQ4yxMC8m5mfRHL/sO7dTeEGIM=", "T/vXn6bo+SCijMlRZwcNRxHpWcvInutRlkDvk1Y7oEM=", "TXWJIUMj+J24/w5Nh86FuokZClxMAnvB6sia+ArBfNI=", "pfPn/gRvLHfxa/LNybYhqG78qHZ4QD3PI+zQ35oYweM=", "sMaCrs8erYNQ48jT2FC56F0RAbpTkuSXZ+cs1zcSObw=", "zGam2EiRYwS9IjemppqM/acVfU15hw3eAqn9Fb8+8vY=", "qQGtNWYnFxNQ1OmqNBR2y92EFrgo5lyaKNQLaE/Dl1c=", "/UAkaaq9D8BPMw5qLAxaf+ur4Sfbv3FfLWB2M/py3wM=", "56+1wf4OgrBZEzfW9n13hPfm0bRvMRgerPv4TPE+3Ho=", "mTSsSZdzVhRgljy9CvDq3GfjxxeOTHrwgpOWK4rRj3Q=", "5sxNZDoxEj6DMmQATisX9bQXdFmeRHYfC8BjyDIIgec=", "Rq/A2aTC9e54ldNjcpsJ26rX/h8JlN6ZHxgw7yEa6C0=", "+/VZ56MsIPxMiyLAodzKXo5TEWdQp36z89qLhpzloAo=", "daxmZaajRpZV+JxHiOYZhJBiSKN5ucqjh2WnGbHhirw=", "DOCeoSMovIvLExkhIvisow9AuNXgeWs4ECkyR6EcqYU="], "checkpoint":{"envelope":"rekor.sigstore.dev - 1193050959916656506\n2045951527\nRe/Ph87BOM7jhvV4VE+3WsyQb72tumWdlqTxAOe0qEM=\n\n— rekor.sigstore.dev wNI9ajBEAiAxviZPcApj0ZJ1BUlQmvWCHwh/I05iSQg0PsjajUUu5wIgfTmpusfaD0Tu2YE2XuUcR3X1uFqLW7goWine86qQDnQ=\n"}}, "canonicalizedBody":"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"}], "timestampVerificationData":{"rfc3161Timestamps":[{"signedTimestamp":"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"}]}}, "messageSignature":{"messageDigest":{"algorithm":"SHA2_256", "digest":"tr0SClqM2qQrkZGvDA7UNDEBmebN3YXBpRhQGBfHZJs="}, "signature":"MEYCIQCG2a5cjD/Oscu+vZqZ5gmpzKlULhWdxrFeRHM4vTIkJwIhAK+mZA5ZPYPl+frTHWKbNSeLEhXyN7sguB6jL1lALHGL"}}

build/torch-rocm/qwen2vl_mrope.py ADDED Viewed

	@@ -0,0 +1,222 @@

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

build/torch-rocm/rms_norm.py ADDED Viewed

	@@ -0,0 +1,654 @@

+"""
+This file incorporates code from Unsloth licensed under the Apache License, Version 2.0.
+See the original Unsloth repository at https://github.com/unslothai/unsloth.
+The following line
+https://github.com/linkedin/Liger-Kernel/blob/7382a8761f9af679482b968f9348013d933947c7/src/liger_kernel/ops/rms_norm.py#L30
+is based on code from Unsloth, located at:
+https://github.com/unslothai/unsloth/blob/fd753fed99ed5f10ef8a9b7139588d9de9ddecfb/unsloth/kernels/rms_layernorm.py#L22
+Modifications made by Yanning Chen, 2024.
+"""
+import math
+import operator
+import torch
+import triton
+import triton.language as tl
+from .utils import calculate_settings
+from .utils import compare_version
+from .utils import ensure_contiguous
+from .utils import get_npu_core_count
+from .utils import set_large_grf_mode
+from .utils import torch_to_triton_dtype
+from .utils import is_npu_available
+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
+    except ModuleNotFoundError:
+        # for working with NGC containers
+        from triton.language.extra.cuda.libdevice import rsqrt
+else:
+    from triton.language.math import rsqrt
+_CASTING_MODE_NONE: tl.constexpr = tl.constexpr(-1)
+_CASTING_MODE_LLAMA: tl.constexpr = tl.constexpr(0)
+_CASTING_MODE_GEMMA: tl.constexpr = tl.constexpr(1)
+@triton.jit
+def _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_cols,
+    eps,
+    offset,
+    casting_mode: tl.constexpr,  # constexpr so the `if` blocks can be optimized out
+    elementwise_affine: tl.constexpr,
+    BLOCK_SIZE: 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).to(tl.int64)
+    col_offsets = tl.arange(0, BLOCK_SIZE)
+    mask = col_offsets < n_cols
+    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_base + col_offsets, mask=mask, other=0)
+    X_row_dtype = X_row.dtype
+    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:
+        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=0) / 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_base, rstd)
+    X_row = X_row * rstd
+    # 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)
+    else:
+        Y_row = X_row
+    if casting_mode == _CASTING_MODE_GEMMA:
+        Y_row = Y_row.to(X_row_dtype)
+    tl.store(y_base + col_offsets, Y_row, mask=mask)
+@triton.jit
+def _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,
+    rows_per_program,
+    casting_mode: tl.constexpr,
+    elementwise_affine: tl.constexpr,
+    BLOCK_SIZE: 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
+    """
+    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
+    if elementwise_affine:
+        dW_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+    if elementwise_affine:
+        W_row = tl.load(W_ptr + col_offsets, mask=mask, other=0.0)
+        W_row = W_row + offset
+    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
+        x_base = X_ptr + row_idx * X_row_stride
+        rstd_base = RSTD_ptr + row_idx * RSTD_row_stride
+        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_base)
+        X_row = X_row.to(tl.float32)
+        # Different bacward graphs for different casting modes
+        if casting_mode == _CASTING_MODE_LLAMA:
+            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)
+            if elementwise_affine:
+                m = dY_row * W_row
+            else:
+                m = dY_row
+        else:
+            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)
+        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:
+            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:
+            if elementwise_affine:
+                m = dY_row * W_row[None, :]
+            else:
+                m = dY_row
+        dX_row = rstd_row[:, None] * m
+        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, :],
+        )
+    if elementwise_affine:
+        tl.store(dW_ptr + pid * dW_row_stride + col_offsets, dW_row, mask=col_mask)
+_str_to_casting_mode = {
+    "llama": _CASTING_MODE_LLAMA.value,
+    "gemma": _CASTING_MODE_GEMMA.value,
+    "none": _CASTING_MODE_NONE.value,
+}
+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]
+    else:
+        assert casting_mode in _str_to_casting_mode.values(), f"Invalid casting mode: {casting_mode}"
+    shape = X.shape
+    dim = shape[-1]
+    X = X.view(-1, dim)
+    n_rows, n_cols = X.shape
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+    Y = torch.empty((n_rows, n_cols), dtype=X.dtype, device=X.device)
+    # RSTD is to cache rstd for each row
+    # RSTD is always computed/stored in fp32 if we are using Llama or Gemma 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)
+    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":
+        set_large_grf_mode(kernel_args)
+    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, row_mode):
+    shape = dY.shape
+    dim = shape[-1]
+    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_eu_count
+    elif X.device.type == "npu":
+        sm_count = get_npu_core_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.")
+    rows_per_program = math.ceil(n_rows / sm_count)
+    grid = (sm_count,)
+    if in_place is True:
+        dX = dY
+    else:
+        dX = torch.zeros_like(dY)
+    # XPU-specific optimization
+    kernel_args = {}
+    if X.device.type == "xpu":
+        set_large_grf_mode(kernel_args)
+    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)
+    if elementwise_affine:
+        dW = _dW.sum(dim=0).to(W.dtype)
+    else:
+        dW = None
+    return dX, dW
+class LigerRMSNormFunction(torch.autograd.Function):
+    """
+    Performs RMSNorm (Root Mean Square Normalization), which normalizes the input tensor `X` using the
+    weight tensor `W`, with an optional offset and casting mode.
+    Some models use an 'offset' to shift the weight tensor `W` by a constant value. For example, Gemma
+    uses an offset of 1.0, so the computation becomes `(X / RMS(X)) * (W + 1.0)` instead of the usual
+    `(X / RMS(X)) * W`. You can pass the offset value as an argument to the forward function.
+    In addition, different models cast their inputs at different places during RMSNorm computation. For
+    example, Gemma casts everything to fp32 nefore starting the computation, while Llama casts only the
+    inverse RMS to fp32. You can specify the casting mode using the `casting_mode` argument. We currently
+    support the following casting modes (they match HuggingFace Transformers' implementations):
+    - 'llama': matches the Llama implementation, where only the inverse RMS is computed on fp32.
+    - 'gemma': matches the Gemma implementation, where everything is cast to fp32, then computed, then cast back to the original dtype.
+    - 'none': no casting is done. The computation is done in the original dtype. This saves memory and is slightly faster, but has more error w.r.t. the original implementation.
+    `in_place` option means whether to in_place modify dY to store dX. This is default to `True` to save memory. However, under certain cases, it can produce incorrect inputs.
+        For example, gemma2 uses two rmsnorm sequentially with residual in between. The resesidual part needs dY so it cannot be modified in-place.
+        Therefore, for the patching of RMSNorm in gemma2, we set `in_place` to `False`
+    """
+    @staticmethod
+    @ensure_contiguous
+    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,)
+        """
+        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.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
+    @ensure_contiguous
+    def backward(ctx, dY):
+        """
+        Y: (B, T, H) or (BxT, H)
+        """
+        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, ctx.row_mode
+        )
+        return dX, dW, None, None, None, None, None

build/torch-rocm/rope.py ADDED Viewed

	@@ -0,0 +1,239 @@

+import torch
+import triton
+import triton.language as tl
+@triton.jit
+def _triton_rope(
+    q_ptr,
+    q_row_stride,
+    k_ptr,
+    k_row_stride,
+    cos,
+    cos_row_stride,
+    sin,
+    sin_row_stride,
+    sl,
+    bs: tl.constexpr,
+    cos_bs: tl.constexpr,
+    n_qh: tl.constexpr,
+    n_kh: tl.constexpr,
+    hd: tl.constexpr,
+    pad_n_qh: tl.constexpr,
+    pad_n_kh: tl.constexpr,
+    pad_hd: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    BACKWARD_PASS: tl.constexpr = False,
+):
+    # q size: (bsz, seq_len, num_q_heads, head_dim)
+    # q stride: (seq_len * num_q_heads * head_dim, num_q_heads * head_dim, head_dim, 1)
+    # k size: (bsz, seq_len, num_kv_heads, head_dim)
+    # k stride: (seq_len * num_kv_heads * head_dim, num_kv_heads * head_dim, head_dim, 1)
+    # 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).to(tl.int64)
+    # locate start address
+    q_ptr = q_ptr + pid * q_row_stride
+    k_ptr = k_ptr + pid * k_row_stride
+    # ####################################################################
+    # 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.
+    batch_idx = pid // sl
+    cos_row_idx = pid % sl
+    cos = cos + tl.where(
+        cos_bs == 1,
+        cos_row_idx * cos_row_stride,
+        batch_idx * (sl * cos_row_stride) + cos_row_idx * cos_row_stride,
+    )
+    sin = sin + tl.where(
+        cos_bs == 1,
+        cos_row_idx * sin_row_stride,
+        batch_idx * (sl * sin_row_stride) + cos_row_idx * sin_row_stride,
+    )
+    cos_offsets = tl.arange(0, pad_hd // 2)
+    cos_mask = cos_offsets < hd // 2
+    cos_row = tl.load(cos + cos_offsets, mask=cos_mask, other=0)
+    sin_row = tl.load(sin + cos_offsets, mask=cos_mask, other=0)
+    # ####################################################################
+    # 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 rope_forward(q, k, cos, sin):
+    # 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()
+    cos_batch_size = cos.shape[0]
+    _triton_rope[(n_row,)](
+        q,
+        q.stride(1),
+        k,
+        k.stride(1),
+        cos,
+        cos.stride(-2),
+        sin,
+        sin.stride(-2),
+        seq_len,
+        batch_size,
+        cos_batch_size,
+        n_q_head,
+        n_kv_head,
+        head_dim,
+        pad_n_q_head,
+        pad_n_kv_head,
+        pad_hd,
+        BLOCK_SIZE=BLOCK_SIZE,
+        BACKWARD_PASS=False,
+    )
+    return q.transpose(1, 2), k.transpose(1, 2), cos, sin
+def rope_backward(dq, dk, cos, sin):
+    dq = dq.transpose(1, 2)
+    dk = dk.transpose(1, 2)
+    batch_size, seq_len, n_q_head, head_dim = dq.shape
+    cos_batch_size = cos.shape[0]
+    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_rope[(n_row,)](
+        dq,
+        dq.stride(1),
+        dk,
+        dk.stride(1),
+        cos,
+        cos.stride(-2),
+        sin,
+        sin.stride(-2),
+        seq_len,
+        batch_size,
+        cos_batch_size,
+        n_q_head,
+        n_kv_head,
+        head_dim,
+        pad_n_q_head,
+        pad_n_kv_head,
+        pad_hd,
+        BLOCK_SIZE=BLOCK_SIZE,
+        BACKWARD_PASS=True,
+    )
+    return dq.transpose(1, 2), dk.transpose(1, 2)
+class LigerRopeFunction(torch.autograd.Function):
+    """
+    Triton implementation of the Rotary Positional Embedding (RoPE) operation. Please note that
+    this implements the HuggingFace Llama & Mistral version, whose rotation matrix is slightly different
+    than the original RoPE paper.
+    Please find the corresponding HuggingFace implementation here:
+    https://github.com/huggingface/transformers/blob/v4.40.2/src/transformers/models/llama/modeling_llama.py#L184
+    For more details about the rotation matrix used here, please refer to:
+    https://discuss.huggingface.co/t/is-llama-rotary-embedding-implementation-correct/44509/2
+    """
+    @staticmethod
+    def forward(ctx, q, k, cos, sin, position_ids=None, 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: (1, seq_len, head_dim) or (bsz, seq_len, head_dim)
+        sin size: (1, seq_len, head_dim) or (bsz, seq_len, head_dim)
+        """
+        q, k, cos, sin = rope_forward(q, k, cos, sin)
+        ctx.save_for_backward(cos, sin)
+        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: (1, seq_len, head_dim) or (bsz, seq_len, head_dim)
+        sin size: (1, seq_len, head_dim) or (bsz, seq_len, head_dim)
+        """
+        cos, sin = ctx.saved_tensors
+        dq, dk = rope_backward(dq, dk, cos, sin)
+        return dq, dk, None, None, None, None

build/torch-rocm/swiglu.py ADDED Viewed

	@@ -0,0 +1,176 @@

+import torch
+import triton
+import triton.language as tl
+from .utils import calculate_settings
+from .utils import ensure_contiguous
+@triton.jit
+def silu(x):
+    return x * tl.sigmoid(x)
+@triton.jit
+def _swiglu_forward_kernel(
+    a_ptr, b_ptr, c_ptr, stride, gate_multiplier, n_cols: tl.constexpr, BLOCK_SIZE: tl.constexpr
+):
+    program_id = tl.program_id(0).to(tl.int64)
+    # locate start index
+    a_ptr += program_id * stride
+    b_ptr += program_id * stride
+    c_ptr += program_id * stride
+    col_offsets = tl.arange(0, BLOCK_SIZE)
+    mask = col_offsets < n_cols
+    # sigmoid requires type float32
+    a_row = tl.load(a_ptr + col_offsets, mask=mask, other=0).to(tl.float32) * gate_multiplier
+    b_row = tl.load(b_ptr + col_offsets, mask=mask, other=0)
+    c_row = silu(a_row).cast(b_row.dtype) * b_row
+    tl.store(c_ptr + col_offsets, c_row, mask=mask)
+@triton.jit
+def _swiglu_backward_kernel(
+    dc_ptr, a_ptr, b_ptr, stride, gate_multiplier, n_cols: tl.constexpr, BLOCK_SIZE: tl.constexpr
+):
+    program_id = tl.program_id(0).to(tl.int64)
+    # locate start index
+    dc_ptr += program_id * stride
+    a_ptr += program_id * stride
+    b_ptr += program_id * stride
+    col_offsets = tl.arange(0, BLOCK_SIZE)
+    mask = col_offsets < n_cols
+    dc_row = tl.load(dc_ptr + col_offsets, mask=mask, other=0)
+    # sigmoid requires type float32
+    a_row = tl.load(a_ptr + col_offsets, mask=mask, other=0).to(tl.float32) * gate_multiplier
+    b_row = tl.load(b_ptr + col_offsets, mask=mask, other=0)
+    # recomputation to save memory. a_row already holds a * gate_multiplier.
+    sig_a = tl.sigmoid(a_row)
+    silu_a = a_row * sig_a
+    db_row = dc_row * silu_a
+    # chain rule pulls an extra factor of gate_multiplier through the pre-activation scaling
+    da_row = dc_row * (silu_a * (1 - sig_a) + sig_a) * b_row * gate_multiplier
+    tl.store(a_ptr + col_offsets, da_row, mask=mask)
+    tl.store(b_ptr + col_offsets, db_row, mask=mask)
+def swiglu_forward(a, b, gate_multiplier: float = 1.0):
+    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]
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+    _swiglu_forward_kernel[(n_rows,)](
+        a,
+        b,
+        c,
+        c.stride(-2),
+        float(gate_multiplier),
+        n_cols=n_cols,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+    )
+    return a, b, c.view(*ori_shape)
+def swiglu_backward(a, b, dc, gate_multiplier: float = 1.0):
+    ori_shape = dc.shape
+    n_cols = ori_shape[-1]
+    dc = dc.view(-1, n_cols)
+    n_rows = dc.shape[0]
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+    _swiglu_backward_kernel[(n_rows,)](
+        dc,
+        a,
+        b,
+        dc.stride(-2),
+        float(gate_multiplier),
+        n_cols=n_cols,
+        BLOCK_SIZE=BLOCK_SIZE,
+        num_warps=num_warps,
+    )
+    return a.view(*ori_shape), b.view(*ori_shape)
+class LigerSiLUMulFunction(torch.autograd.Function):
+    @staticmethod
+    @ensure_contiguous
+    def forward(ctx, a, b, gate_multiplier: float = 1.0, down_multiplier: float = 1.0):
+        gate_multiplier = float(gate_multiplier)
+        down_multiplier = float(down_multiplier)
+        ctx.gate_multiplier = gate_multiplier
+        ctx.down_multiplier = down_multiplier
+        if isinstance(a, torch.distributed.tensor.DTensor) or isinstance(b, torch.distributed.tensor.DTensor):
+            device_mesh, placements = (
+                (a.device_mesh, a.placements)
+                if isinstance(a, torch.distributed.tensor.DTensor)
+                else (b.device_mesh, b.placements)
+            )
+            # Assume that full tensors are gathered before and identical across
+            # the associated process groups.
+            if not isinstance(a, torch.distributed.tensor.DTensor):
+                a = torch.distributed.tensor.distribute_tensor(a, device_mesh=device_mesh, placements=placements)
+            if not isinstance(b, torch.distributed.tensor.DTensor):
+                b = torch.distributed.tensor.distribute_tensor(b, device_mesh=device_mesh, placements=placements)
+            a_local, b_local, c_local = swiglu_forward(a.to_local(), b.to_local(), gate_multiplier)
+            if down_multiplier != 1.0:
+                c_local = c_local * down_multiplier
+            ctx.save_for_backward(a_local, b_local)
+            ctx.dtensor_metadata = (device_mesh, placements)
+            return torch.distributed.tensor.DTensor.from_local(c_local, device_mesh, placements)
+        else:
+            a, b, c = swiglu_forward(a, b, gate_multiplier)
+            if down_multiplier != 1.0:
+                c = c * down_multiplier
+            ctx.save_for_backward(a, b)
+            ctx.dtensor_metadata = None
+            return c
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, dc):
+        a, b = ctx.saved_tensors
+        gate_multiplier = ctx.gate_multiplier
+        down_multiplier = ctx.down_multiplier
+        if ctx.dtensor_metadata is not None:
+            device_mesh, placements = ctx.dtensor_metadata
+            # Assume that full tensors are gathered before and identical across
+            # the associated process groups.
+            dc_local = (
+                dc.to_local()
+                if isinstance(dc, torch.distributed.tensor.DTensor)
+                else torch.distributed.tensor.distribute_tensor(dc, device_mesh=device_mesh, placements=placements)
+            )
+            if down_multiplier != 1.0:
+                dc_local = dc_local * down_multiplier
+            a_local, b_local = swiglu_backward(a, b, dc_local, gate_multiplier)
+            return (
+                torch.distributed.tensor.DTensor.from_local(a_local, device_mesh, placements),
+                torch.distributed.tensor.DTensor.from_local(b_local, device_mesh, placements),
+                None,
+                None,
+            )
+        if down_multiplier != 1.0:
+            dc = dc * down_multiplier
+        a, b = swiglu_backward(a, b, dc, gate_multiplier)
+        return a, b, None, None

build/torch-rocm/tiled_mlp.py ADDED Viewed

	@@ -0,0 +1,159 @@

+import math
+from typing import Callable
+from typing import List
+from typing import Optional
+import torch
+import torch.distributed as dist
+from .utils import ensure_contiguous
+class LigerTiledMLPFunction(torch.autograd.Function):
+    """
+    Based on DeepSpeed's TiledMLP:
+    https://github.com/deepspeedai/DeepSpeed/blob/v0.18.2/deepspeed/runtime/sequence_parallel/ulysses_sp.py#L838
+    Perform a tiled MLP computation to massively reduce memory usage needed to compute MLP
+    when using very long sequence lengths.
+    This module re-computes `forward` in the `backward`. So the `forward` occurs twice each iteration.
+    And if you're using activation checkpointing it then occurs thrice.
+    Args:
+        fn: the function to call on sharded inputs (e.g., mlp.forward)
+        mlp_module: the MLP nn.Module object
+        x: the input to MLP.forward (hidden_states)
+        shards: how many shards to use
+        compute_params: a list of weights engaged in the compute
+    Returns:
+        the computed hidden_states
+    """
+    @staticmethod
+    @ensure_contiguous
+    def forward(
+        ctx,
+        fn: Callable,
+        mlp_module: torch.nn.Module,
+        x: torch.Tensor,
+        shards: int,
+        compute_params: Optional[List[torch.nn.Parameter]] = None,
+    ) -> torch.Tensor:
+        ctx.fn = fn
+        ctx.mlp_module = mlp_module
+        ctx.shards = shards
+        ctx.compute_params = compute_params
+        ctx.save_for_backward(x)
+        # x.shape could be [bs, seqlen, hidden_size] or [seqlen, hidden_size] (moe experts)
+        x_shards = list(torch.chunk(x, chunks=shards, dim=-2))
+        with torch.no_grad():
+            output_shards = [fn(mlp_module, x_shard) for x_shard in x_shards]
+        output_unsharded = torch.cat(output_shards, dim=-2)
+        return output_unsharded
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, *grads) -> tuple:
+        fn = ctx.fn
+        (x,) = ctx.saved_tensors
+        mlp_module = ctx.mlp_module
+        shards = ctx.shards
+        compute_params = ctx.compute_params
+        grad_output = grads[0]
+        x_requires_grad = x.requires_grad
+        x = x.detach()
+        # detach() unsets x.requires_grad, so restore it
+        x.requires_grad_(x_requires_grad)
+        # x.shape could be [bs, seqlen, in_features] or [seqlen, in_features] (moe experts)
+        in_features = x.shape[-1]
+        out_features = grad_output.shape[-1]
+        x_shape_orig = x.shape
+        # flatten bs+seqlen to avoid having stride issues when narrowing into seqlen w/ bs>1
+        # NOTE: input and output feature dimensions may differ, e.g.
+        # Linear(in_features, out_features), so flatten x and grad_output with their own last dims.
+        x = x.view(-1, in_features)
+        incoming_grad = grad_output.view(-1, out_features)
+        x_grad = torch.zeros_like(x)
+        x_shards = list(torch.chunk(x, chunks=shards, dim=0))
+        incoming_grad_shards = list(torch.chunk(incoming_grad, chunks=shards, dim=0))
+        # ZeRO-3 partitioned parameters carry a ds_id; collect them once so the per-shard loop only
+        # flips the ready flag. Parameters on other backends have no ds_id and are left untouched.
+        ds_params = [p for p in compute_params if hasattr(p, "ds_id")] if compute_params else []
+        for i, x_shard in enumerate(x_shards):
+            x_shard.requires_grad_(x_requires_grad)
+            # if seqlen is not exactly divisible by shards the last step will be shorter than shard_step
+            shard_step = x_shards[i].shape[0]
+            shard_offset = i * x_shards[0].shape[0]
+            x_shard.grad = x_grad.narrow(0, shard_offset, shard_step).view_as(x_shard)
+            incoming_grad_shard = incoming_grad_shards[i]
+            # Defer DeepSpeed's reduction until the last shard has accumulated into param.grad; the flag
+            # is read by ZeRO's hook during each shard's backward, so it must be set per shard.
+            for param in ds_params:
+                param.ds_grad_is_ready = i + 1 == len(x_shards)
+            with torch.enable_grad():
+                output = fn(mlp_module, x_shard)
+            torch.autograd.backward(output, incoming_grad_shard)
+        # unflatten
+        x_grad = x_grad.view(x_shape_orig)
+        return (None, None, x_grad, None, None)
+def apply_tiled_mlp(
+    fn: Callable,
+    mlp_module: torch.nn.Module,
+    x: torch.Tensor,
+    num_shards: Optional[int] = None,
+    compute_params: Optional[List[torch.nn.Parameter]] = None,
+) -> torch.Tensor:
+    """
+    Apply tiled MLP computation for memory efficiency.
+    Args:
+        fn: the function to call on sharded inputs (e.g., lambda module, x: module(x))
+        mlp_module: the MLP nn.Module object
+        x: the input tensor with shape [bs, seqlen, hidden_size] or [seqlen, hidden_size]
+        num_shards: number of shards to use. If None, automatically calculated as ceil(seqlen / hidden_size)
+        compute_params: list of parameters for DeepSpeed ZeRO optimization
+    """
+    if num_shards is None:
+        # x.shape could be [bs, seqlen, hidden_size] or [seqlen, hidden_size]
+        hidden_size = x.shape[-1]
+        seqlen = x.shape[-2]
+        num_shards = math.ceil(seqlen / hidden_size)
+    # Ensure num_shards is at least 1
+    num_shards = max(1, num_shards)
+    # All ranks must run the same number of shards: a sharded-parameter backend (DeepSpeed ZeRO-3, FSDP)
+    # gathers weights inside each shard's recompute, so a rank that runs fewer shards stops participating
+    # in those collectives and deadlocks the others. Harmonize on the per-rank maximum.
+    if dist.is_available() and dist.is_initialized() and dist.get_world_size() > 1:
+        num_shards_tensor = torch.tensor(num_shards, device=x.device)
+        dist.all_reduce(num_shards_tensor, op=dist.ReduceOp.MAX)
+        num_shards = int(num_shards_tensor.item())
+    return LigerTiledMLPFunction.apply(
+        fn,
+        mlp_module,
+        x,
+        num_shards,
+        compute_params,
+    )

build/torch-rocm/tvd.py ADDED Viewed

	@@ -0,0 +1,218 @@

+from typing import Literal
+from typing import Optional
+import torch
+import triton
+import triton.language as tl
+from .utils import ensure_contiguous
+MAX_FUSED_SIZE = 65536 // 4
+REDUCTION_LITERAL = Literal["none", "sum", "mean", "batchmean"]
+_REDUCTION_MODE_NONE = tl.constexpr(0)
+_REDUCTION_MODE_SUM = tl.constexpr(1)
+_REDUCTION_MODE_MEAN = tl.constexpr(2)
+_REDUCTION_MODE_BATCHMEAN = tl.constexpr(3)
+_str_to_reduction_mode = {
+    "none": _REDUCTION_MODE_NONE.value,
+    "sum": _REDUCTION_MODE_SUM.value,
+    "mean": _REDUCTION_MODE_MEAN.value,
+    "batchmean": _REDUCTION_MODE_BATCHMEAN.value,
+}
+def get_num_warps(BLOCK_SIZE):
+    num_warps = 4
+    if BLOCK_SIZE >= 32768:
+        num_warps = 32
+    elif BLOCK_SIZE >= 8192:
+        num_warps = 16
+    elif BLOCK_SIZE >= 2048:
+        num_warps = 8
+    return num_warps
+@triton.jit
+def _tv_distance_kernel(
+    p_ptr,
+    p_stride,
+    q_ptr,
+    q_stride,
+    loss_ptr,
+    loss_stride,
+    grads_ptr,
+    grads_stride,
+    label_ptr,
+    ignore_index: tl.constexpr,
+    n_cols,
+    scale,  # pre-computed reduction scale for gradients (fused into kernel)
+    BLOCK_SIZE: tl.constexpr,
+    HAS_LABEL: tl.constexpr,
+    reduction: tl.constexpr = _REDUCTION_MODE_BATCHMEAN,
+):
+    pid = tl.program_id(0).to(tl.int64)
+    p_ptr += pid * p_stride
+    q_ptr += pid * q_stride
+    loss_ptr += pid * loss_stride
+    grads_ptr += pid * grads_stride
+    label_ptr += pid
+    base_offsets = tl.arange(0, BLOCK_SIZE)
+    if HAS_LABEL:
+        label = tl.load(label_ptr)
+        if label == ignore_index:
+            for i in range(0, n_cols, BLOCK_SIZE):
+                offsets = i + base_offsets
+                mask = offsets < n_cols
+                tl.store(grads_ptr + offsets, 0.0, mask=mask)
+                if reduction == _REDUCTION_MODE_NONE:
+                    tl.store(loss_ptr + offsets, 0.0, mask=mask)
+            return
+    loss_sum = 0.0
+    for i in range(0, n_cols, BLOCK_SIZE):
+        offsets = i + base_offsets
+        mask = offsets < n_cols
+        p = tl.load(p_ptr + offsets, mask=mask, other=0.0)
+        q = tl.load(q_ptr + offsets, mask=mask, other=0.0)
+        # TVD(P || Q) = 0.5 * |P - Q|
+        tv_loss = 0.5 * tl.abs(p - q)
+        # Fuse reduction scaling into gradient computation (eliminates separate Python division)
+        grad_res = tl.where(p > q, 0.5 * scale, -0.5 * scale)
+        tl.store(grads_ptr + offsets, grad_res, mask=mask)
+        if reduction == _REDUCTION_MODE_NONE:
+            tl.store(loss_ptr + offsets, tv_loss, mask=mask)
+        else:
+            loss_sum += tl.sum(tv_loss, axis=0)
+    if reduction != _REDUCTION_MODE_NONE:
+        # Fuse reduction scaling into loss (same scale as gradients; avoids Python division)
+        tl.store(loss_ptr, loss_sum * scale)
+def tv_distance_forward_triton(p, q, shift_labels, reduction, ignore_index, has_label):
+    BT, V = p.shape
+    BLOCK_SIZE = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))
+    num_warps = get_num_warps(BLOCK_SIZE)
+    grid = (BT,)
+    reduction = _str_to_reduction_mode[reduction]
+    out_size = (BT, V) if reduction == _REDUCTION_MODE_NONE.value else (BT,)
+    output_tensor = torch.zeros(out_size, device=p.device, dtype=torch.float32)
+    grads = torch.empty_like(p)
+    n_non_ignore = (shift_labels != ignore_index).sum().item() if has_label else BT
+    # Pre-compute gradient scale factor (fused into kernel to avoid separate division)
+    if reduction == _REDUCTION_MODE_BATCHMEAN.value:
+        scale = 1.0 / n_non_ignore
+    elif reduction == _REDUCTION_MODE_MEAN.value:
+        scale = 1.0 / (n_non_ignore * V)
+    else:
+        scale = 1.0
+    _tv_distance_kernel[grid](
+        p,
+        p.stride(0),
+        q,
+        q.stride(0),
+        output_tensor,
+        output_tensor.stride(0),
+        grads,
+        grads.stride(0),
+        shift_labels if has_label else torch.empty(1, device=p.device),
+        ignore_index,
+        V,
+        scale,
+        BLOCK_SIZE=BLOCK_SIZE,
+        HAS_LABEL=has_label,
+        num_warps=num_warps,
+        reduction=reduction,
+    )
+    # Loss and gradients are already scaled inside the kernel — no separate division needed
+    if reduction in (_REDUCTION_MODE_BATCHMEAN.value, _REDUCTION_MODE_MEAN.value):
+        return output_tensor.sum(), grads
+    elif reduction == _REDUCTION_MODE_SUM.value:
+        return output_tensor.sum(dim=0), grads
+    else:
+        return output_tensor, grads
+def tvd_backward_triton(grad_output, grads):
+    # If cross entropy is the last layer, grad_output is 1.0. Skip the mul then.
+    if torch.equal(grad_output, torch.tensor(1.0, device=grad_output.device)):
+        return grads
+    return grads * grad_output
+class LigerTVDLossFunction(torch.autograd.Function):
+    """
+    Class implementing the forward and backward pass for the Total Variation Distance Loss using Triton.
+    """
+    @staticmethod
+    @ensure_contiguous
+    def forward(
+        ctx,
+        p: torch.Tensor,
+        q: torch.Tensor,
+        shift_labels: Optional[torch.Tensor] = None,
+        reduction: REDUCTION_LITERAL = "batchmean",
+        ignore_index: int = -100,
+    ) -> torch.Tensor:
+        """A forward pass for the Total Variation Distance Loss.
+        Args:
+            ctx: Torch autograd context
+            p (torch.Tensor): A tensor of shape (BT, V) containing the first distribution.
+            q (torch.Tensor): A tensor of shape (BT, V) containing the second distribution.
+            shift_labels (Optional[torch.Tensor]): A tensor of shape (BT,) containing the labels.
+            reduction (REDUCTION_LITERAL, optional): The reduction method to be applied. Defaults to "batchmean".
+            ignore_index (int, optional): The index to ignore during loss calculation. Defaults to -100.
+        Returns:
+            torch.Tensor: The computed Total Variation Distance Loss.
+        """
+        has_label = False
+        if shift_labels is not None:
+            assert shift_labels.shape == (p.shape[0],), (
+                f"the shape of shift_labels must be (BT,). Got: {shift_labels.shape}"
+            )
+            shift_labels = shift_labels.contiguous()
+            has_label = True
+        loss, grads = tv_distance_forward_triton(p, q, shift_labels, reduction, ignore_index, has_label)
+        ctx.save_for_backward(grads)
+        return loss
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, grad_output: torch.Tensor) -> torch.Tensor:
+        """A backward pass for the Total Variation Distance Loss.
+        Args:
+            ctx: Torch autograd context
+            grad_output (torch.Tensor): The gradient of the loss with respect to the output.
+        Returns:
+            tuple[torch.Tensor, None, None, None, None]: The gradient of the loss with respect to the inputs.
+        """
+        (grads,) = ctx.saved_tensors
+        grads = tvd_backward_triton(grad_output, grads)
+        return grads, None, None, None, None

build/torch-rocm/utils.py ADDED Viewed

	@@ -0,0 +1,176 @@

+"""
+This file incorporates code from Unsloth licensed under the Apache License, Version 2.0.
+See the original Unsloth repository at https://github.com/unslothai/unsloth.
+The following line
+https://github.com/linkedin/Liger-Kernel/blob/7382a8761f9af679482b968f9348013d933947c7/src/liger_kernel/ops/utils.py#L23
+is based on code from Unsloth, located at:
+https://github.com/unslothai/unsloth/blob/fd753fed99ed5f10ef8a9b7139588d9de9ddecfb/unsloth/kernels/utils.py#L43
+Modifications made by Yanning Chen, 2024.
+"""
+import functools
+import importlib
+import operator
+from typing import Callable
+import torch
+import triton
+import triton.language as tl
+from packaging.version import Version
+def is_npu_available() -> bool:
+    """Detect Ascend NPU availability."""
+    try:
+        from transformers.utils import is_torch_npu_available
+        return is_torch_npu_available()
+    except Exception:
+        return False
+def infer_device():
+    """
+    Get current device name based on available devices
+    """
+    if torch.cuda.is_available():  # Works for both Nvidia and AMD
+        return "cuda"
+    # Use Ascend NPU if available (torch.npu)
+    elif is_npu_available():
+        return "npu"
+    # XPU (Intel) if available
+    elif torch.xpu.is_available():
+        return "xpu"
+    else:
+        return "cpu"
+def is_hip() -> bool:
+    return torch.version.hip is not None
+def ensure_contiguous(fn):
+    @functools.wraps(fn)
+    def wrapper(ctx, *args, **kwargs):
+        def maybe_to_contiguous(x):
+            return x.contiguous() if isinstance(x, torch.Tensor) else x
+        args = [maybe_to_contiguous(arg) for arg in args]
+        kwargs = {k: maybe_to_contiguous(v) for k, v in kwargs.items()}
+        return fn(ctx, *args, **kwargs)
+    return wrapper
+def calculate_settings(n):
+    # reference: https://github.com/unslothai/unsloth/blob/fd753fed99ed5f10ef8a9b7139588d9de9ddecfb/unsloth/kernels/utils.py#L43
+    MAX_FUSED_SIZE = 65536
+    BLOCK_SIZE = triton.next_power_of_2(n)
+    if BLOCK_SIZE > MAX_FUSED_SIZE:
+        raise RuntimeError(
+            f"Cannot launch Triton kernel since n = {n} exceeds the recommended Triton blocksize = {MAX_FUSED_SIZE}."
+        )
+    num_warps = 4
+    if BLOCK_SIZE >= 32768:
+        num_warps = 32 if not is_hip() else 16
+    elif BLOCK_SIZE >= 8192:
+        num_warps = 16
+    elif BLOCK_SIZE >= 2048:
+        num_warps = 8
+    return BLOCK_SIZE, num_warps
+def compare_version(package: str, operator: Callable, target: str):
+    try:
+        pkg = importlib.import_module(package)
+    except ImportError:
+        return False
+    pkg_version = Version(pkg.__version__)
+    return operator(pkg_version, Version(target))
+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=device),
+            functools.partial(torch.amp.custom_bwd, device_type=device),
+        )
+    if hasattr(torch, "npu") and getattr(torch.npu, "amp", None) is not None:
+        return torch.npu.amp.custom_fwd, torch.npu.amp.custom_bwd
+    return torch.cuda.amp.custom_fwd, torch.cuda.amp.custom_bwd
+amp_custom_fwd, amp_custom_bwd = get_amp_custom_fwd_bwd()
+torch_to_triton_dtype = {
+    torch.float32: tl.float32,
+    torch.float16: tl.float16,
+    torch.bfloat16: tl.bfloat16,
+}
+@triton.jit
+def element_mul_kernel(
+    X_ptr,
+    X_stride,
+    grad_output_ptr,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+):
+    """
+    This function multiplies each element of the tensor pointed by X_ptr with the value pointed by grad_output_ptr.
+    The multiplication is performed in-place on the tensor pointed by X_ptr.
+    Parameters:
+    X_ptr: Pointer to the input tensor.
+    X_stride (int): The stride of the input tensor.
+    grad_output_ptr: Pointer to the gradient output value.
+    n_cols (int): The number of columns in the input tensor.
+    BLOCK_SIZE (int): The block size for Triton operations.
+    """
+    # Get the program ID and convert it to int64 to avoid overflow
+    program_id = tl.program_id(0).to(tl.int64)
+    # Locate the start index
+    X_ptr += program_id * X_stride
+    # Load the gradient output value
+    grad_output = tl.load(grad_output_ptr)
+    # Perform the element-wise multiplication
+    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)
+        tl.store(X_ptr + X_offsets, X_block * grad_output, mask=X_offsets < n_cols)
+def get_npu_core_count(default: int = 20) -> int:
+    """Return NPU vector core count.
+    Fallback to `default` if Triton runtime or NPU device is unavailable.
+    """
+    try:
+        utils = triton.runtime.driver.active.utils
+        props = utils.get_device_properties(0)
+        return int(props.get("num_vectorcore", default))
+    except Exception:
+        return default
+def set_large_grf_mode(kernel_args: dict):
+    """Set large GRF mode for XPU devices."""
+    # On XPU triton installed along with pytorch-xpu will be called `pytorch-triton-xpu`,
+    # triton XPU installed from source will be called `triton`.
+    if compare_version("pytorch-triton-xpu", operator.ge, "3.6.0") or compare_version("triton", operator.ge, "3.6.0"):
+        kernel_args["grf_mode"] = "256"
+    else:
+        # API was changed in https://github.com/intel/intel-xpu-backend-for-triton/pull/5430
+        kernel_args["grf_mode"] = "large"