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

liger_kernel/chunked_loss/jsd_loss.py

@@ -1,3 +1,8 @@
+import math
+
+from typing import Tuple
+from typing import Union
+
 import torch
 import torch.nn.functional as F
 
@@ -6,34 +11,50 @@ from liger_kernel.chunked_loss.fused_linear_distillation import LigerFusedLinear
 
 class LigerFusedLinearJSDFunction(LigerFusedLinearDistillationBase):
     @staticmethod
-    def distillation_loss_fn(student_logits, teacher_logits, beta=0.5):
+    def distillation_loss_fn(student_logits, teacher_logits, beta=0.5, target=None, ignore_index=-100):
         """
         Compute JSD loss (Jensen-Shannon Divergence Loss).
         Args:
             student_logits (torch.Tensor): Logits of student tokens. Shape: (batch_size * seq_len,).
             teacher_logits (torch.Tensor): Logits of teacher tokens. Shape: (batch_size * seq_len,).
             beta (float): Coefficient beta of generalized JSD in the interval [0, 1]. Default: `0.5`.
+            target (torch.Tensor): Target labels for masking. Shape: (chunk_size,).
+            ignore_index (int): Index to ignore in loss computation.
         Returns:
             torch.Tensor: Jensen-Shannon Divergence loss
+        Note:
+            - Uses reduction="none" to preserve per-token losses for masking
+            - KL divergence requires summing over vocab dimension (not mean)
+            - Masking excludes padding/prompt tokens from loss computation
         """
         student_log_probs = F.log_softmax(student_logits, dim=-1)
         teacher_log_probs = F.log_softmax(teacher_logits, dim=-1)
 
         if beta == 0:
-            jsd_loss = F.kl_div(student_log_probs, teacher_log_probs, reduction="sum", log_target=True)
+            jsd_loss = F.kl_div(student_log_probs, teacher_log_probs, reduction="none", log_target=True)
         elif beta == 1:
-            jsd_loss = F.kl_div(teacher_log_probs, student_log_probs, reduction="sum", log_target=True)
+            jsd_loss = F.kl_div(teacher_log_probs, student_log_probs, reduction="none", log_target=True)
         else:
             # Compute probabilities (only required for mean calculation)
-            mean_probs = (1 - beta) * student_log_probs.exp() + beta * teacher_log_probs.exp()
-            log_mean_probs = mean_probs.log()
+            log_mean_probs = torch.logsumexp(
+                torch.stack([student_log_probs + math.log(1 - beta), teacher_log_probs + math.log(beta)], dim=0), dim=0
+            )
 
-            student_kl = F.kl_div(log_mean_probs, student_log_probs, reduction="sum", log_target=True)
-            teacher_kl = F.kl_div(log_mean_probs, teacher_log_probs, reduction="sum", log_target=True)
+            student_kl = F.kl_div(log_mean_probs, student_log_probs, reduction="none", log_target=True)
+            teacher_kl = F.kl_div(log_mean_probs, teacher_log_probs, reduction="none", log_target=True)
 
             # JSD is the weighted average of the KL divergences
             jsd_loss = beta * teacher_kl + (1 - beta) * student_kl
-        return jsd_loss
+
+        # Sum over vocab dimension (KL divergence definition)
+        jsd_loss = jsd_loss.sum(dim=-1)  # (chunk_size,)
+
+        # Apply ignore_index mask
+        if target is not None:
+            mask = target != ignore_index
+            jsd_loss = jsd_loss.masked_fill(~mask, 0.0)
+
+        return jsd_loss.sum()
 
     @classmethod
     def forward(
@@ -53,6 +74,7 @@ class LigerFusedLinearJSDFunction(LigerFusedLinearDistillationBase):
         temperature: float = 1.0,
         compiled: bool = True,
         chunk_size: int = 1024,
+        return_soft_hard_loss: bool = False,
     ):
         """
         Fused linear layer with JSD distillation loss.
@@ -69,8 +91,9 @@ class LigerFusedLinearJSDFunction(LigerFusedLinearDistillationBase):
             temperature (float): Temperature for softening/sharpening distributions
             compiled (bool): Whether to use torch compile
             chunk_size (int): Size of chunks for processing.
+            return_soft_hard_loss (bool): Whether to return soft and hard losses separately. Default: False.
         Returns:
-            torch.Tensor: Computed loss
+            torch.Tensor: Computed loss, or tuple (loss, soft_loss, hard_loss) if return_soft_hard_loss=True
         """
         return super().forward(
             cls=cls,
@@ -89,11 +112,12 @@ class LigerFusedLinearJSDFunction(LigerFusedLinearDistillationBase):
             ignore_index=ignore_index,
             temperature=temperature,
             compiled=compiled,
+            return_soft_hard_loss=return_soft_hard_loss,
         )
 
     @staticmethod
-    def backward(ctx, grad_output):
-        grads = LigerFusedLinearDistillationBase.backward(ctx, grad_output)[:6]
+    def backward(ctx, grad_output, *args):
+        grads = LigerFusedLinearDistillationBase.backward(ctx, grad_output, *args)[:6]
 
         return (
             *grads,
@@ -105,6 +129,7 @@ class LigerFusedLinearJSDFunction(LigerFusedLinearDistillationBase):
             None,  # temperature
             None,  # compiled
             None,  # chunk_size
+            None,  # return_soft_hard_loss
         )
 
 
@@ -122,6 +147,7 @@ class LigerFusedLinearJSDLoss(torch.nn.Module):
         temperature: float = 1.0,
         compiled: bool = True,
         chunk_size: int = 1024,
+        return_soft_hard_loss: bool = False,
     ):
         """
         Args:
@@ -132,6 +158,7 @@ class LigerFusedLinearJSDLoss(torch.nn.Module):
             compiled (bool): Whether to use torch compile
             beta (float): Coefficient beta of generalized JSD in the interval [0, 1]. Default: `0.5`.
             chunk_size (int): Size of chunks for processing.
+            return_soft_hard_loss (bool): Whether to return soft and hard losses separately. Default: False.
         """
         super().__init__()
         assert temperature != 0, "Temperature cannot be 0."
@@ -142,6 +169,7 @@ class LigerFusedLinearJSDLoss(torch.nn.Module):
         self.compiled = compiled
         self.beta = beta
         self.chunk_size = chunk_size
+        self.return_soft_hard_loss = return_soft_hard_loss
 
     def forward(
         self,
@@ -150,9 +178,9 @@ class LigerFusedLinearJSDLoss(torch.nn.Module):
         teacher_input: torch.Tensor,
         teacher_weight: torch.Tensor,
         true_labels: torch.LongTensor,
-        student_bias: torch.Tensor,
-        teacher_bias: torch.Tensor,
-    ) -> torch.Tensor:
+        student_bias: torch.Tensor = None,
+        teacher_bias: torch.Tensor = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor, torch.Tensor]]:
         """
         Compute the JSD distillation loss.
 
@@ -164,7 +192,9 @@ class LigerFusedLinearJSDLoss(torch.nn.Module):
             true_labels (torch.LongTensor): Target labels tensor
 
         Returns:
-            torch.Tensor: Computed loss
+            torch.Tensor or Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+                If return_soft_hard_loss is False: Computed combined loss
+                If return_soft_hard_loss is True: Tuple of (combined_loss, soft_loss, hard_loss)
         """
         return LigerFusedLinearJSDFunction.apply(
             student_input,
@@ -181,4 +211,5 @@ class LigerFusedLinearJSDLoss(torch.nn.Module):
             self.temperature,
             self.compiled,
             self.chunk_size,
+            self.return_soft_hard_loss,
         )

liger_kernel/ops/__init__.py

@@ -0,0 +1,141 @@
+"""
+Liger-Kernel operators with automatic vendor-specific replacement.
+
+This module provides two ways to import operators:
+
+1. Import from this package (recommended for Function classes):
+       from liger_kernel.ops import LigerGELUMulFunction
+
+   This automatically uses vendor-specific implementation if available.
+
+2. Import from submodules (for kernel functions or specific access):
+       from liger_kernel.ops.geglu import geglu_forward, geglu_backward
+
+   This always uses the default implementation (no auto-replacement).
+
+The replacement mechanism:
+1. Default implementations are imported from individual modules (e.g., geglu.py)
+2. On module load, device is detected via infer_device()
+3. If running on a supported vendor device (npu, xpu, etc.), the default
+   implementations are replaced with vendor-specific ones
+4. All subsequent imports from this package get the replaced versions
+
+Note: Direct imports from submodules (e.g., from liger_kernel.ops.geglu import ...)
+      are NOT affected by the replacement mechanism.
+"""
+
+# =============================================================================
+# Import default implementations
+# Both Function classes and kernel functions are imported here.
+# All of these can be replaced by vendor-specific implementations.
+# =============================================================================
+
+from liger_kernel.ops.cross_entropy import LigerCrossEntropyFunction  # noqa: F401
+from liger_kernel.ops.cross_entropy import cross_entropy_backward  # noqa: F401
+from liger_kernel.ops.cross_entropy import cross_entropy_forward  # noqa: F401
+from liger_kernel.ops.dyt import LigerDyTFunction  # noqa: F401
+from liger_kernel.ops.experimental.embedding import LigerEmbeddingFunction  # noqa: F401
+from liger_kernel.ops.fused_add_rms_norm import LigerFusedAddRMSNormFunction  # noqa: F401
+from liger_kernel.ops.fused_add_rms_norm import fused_add_rms_norm_backward  # noqa: F401
+from liger_kernel.ops.fused_add_rms_norm import fused_add_rms_norm_forward  # noqa: F401
+from liger_kernel.ops.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyFunction  # noqa: F401
+from liger_kernel.ops.fused_linear_cross_entropy import fused_linear_cross_entropy_backward  # noqa: F401
+from liger_kernel.ops.fused_linear_cross_entropy import fused_linear_cross_entropy_forward  # noqa: F401
+from liger_kernel.ops.fused_linear_jsd import LigerFusedLinearJSDFunction  # noqa: F401
+from liger_kernel.ops.fused_linear_jsd import fused_linear_jsd_backward  # noqa: F401
+from liger_kernel.ops.fused_linear_jsd import fused_linear_jsd_forward  # noqa: F401
+from liger_kernel.ops.fused_neighborhood_attention import LigerFusedNeighborhoodAttentionFunction  # noqa: F401
+from liger_kernel.ops.geglu import LigerGELUMulFunction  # noqa: F401
+from liger_kernel.ops.geglu import geglu_backward  # noqa: F401
+from liger_kernel.ops.geglu import geglu_forward  # noqa: F401
+from liger_kernel.ops.group_norm import LigerGroupNormFunction  # noqa: F401
+from liger_kernel.ops.group_norm import group_norm_backward  # noqa: F401
+from liger_kernel.ops.group_norm import group_norm_forward  # noqa: F401
+from liger_kernel.ops.grpo_loss import GrpoLossFunction  # noqa: F401
+from liger_kernel.ops.jsd import LigerJSDFunction  # noqa: F401
+from liger_kernel.ops.jsd import jsd_backward  # noqa: F401
+from liger_kernel.ops.jsd import jsd_forward  # noqa: F401
+from liger_kernel.ops.kl_div import LigerKLDivLossFunction  # noqa: F401
+from liger_kernel.ops.layer_norm import LigerLayerNormFunction  # noqa: F401
+from liger_kernel.ops.layer_norm import layer_norm_backward  # noqa: F401
+from liger_kernel.ops.layer_norm import layer_norm_forward  # noqa: F401
+from liger_kernel.ops.llama4_rope import LigerLlama4RopeFunction  # noqa: F401
+from liger_kernel.ops.multi_token_attention import LigerMultiTokenAttentionFunction  # noqa: F401
+from liger_kernel.ops.poly_norm import LigerPolyNormFunction  # noqa: F401
+from liger_kernel.ops.poly_norm import poly_norm_backward  # noqa: F401
+from liger_kernel.ops.poly_norm import poly_norm_forward  # noqa: F401
+from liger_kernel.ops.qwen2vl_mrope import LigerQwen2VLMRopeFunction  # noqa: F401
+from liger_kernel.ops.rms_norm import LigerRMSNormFunction  # noqa: F401
+from liger_kernel.ops.rms_norm import rms_norm_backward  # noqa: F401
+from liger_kernel.ops.rms_norm import rms_norm_forward  # noqa: F401
+from liger_kernel.ops.rope import LigerRopeFunction  # noqa: F401
+from liger_kernel.ops.rope import rope_backward  # noqa: F401
+from liger_kernel.ops.rope import rope_forward  # noqa: F401
+from liger_kernel.ops.softmax import LigerSoftmaxFunction  # noqa: F401
+from liger_kernel.ops.sparsemax import LigerSparsemaxFunction  # noqa: F401
+from liger_kernel.ops.swiglu import LigerSiLUMulFunction  # noqa: F401
+from liger_kernel.ops.swiglu import swiglu_backward  # noqa: F401
+from liger_kernel.ops.swiglu import swiglu_forward  # noqa: F401
+from liger_kernel.ops.tiled_mlp import LigerTiledMLPFunction  # noqa: F401
+from liger_kernel.ops.tiled_mlp import apply_tiled_mlp  # noqa: F401
+from liger_kernel.ops.tvd import LigerTVDLossFunction  # noqa: F401
+
+# NOTE: __all__ is intentionally NOT defined.
+# - Import from this package (liger_kernel.ops) -> subject to vendor replacement
+# - Import from submodules (liger_kernel.ops.geglu) -> always use default implementation
+
+
+# =============================================================================
+# Vendor-specific replacement logic
+# =============================================================================
+
+
+def _replace_with_vendor_ops():
+    """
+    Replace/add vendor-specific operator implementations.
+
+    This function is called automatically on module load. It:
+    1. Detects the current device (cuda, npu, xpu, etc.)
+    2. Looks up the vendor for that device via VENDOR_REGISTRY
+    3. Loads and applies vendor-specific implementations
+
+    Vendor implementations should be placed in:
+        liger_kernel/ops/backends/_<vendor>/ops/
+
+    If the vendor module defines __all__, only those symbols are exported.
+    Otherwise, all public symbols (not starting with _) are auto-discovered.
+
+    Note: Vendor can both override existing ops AND add new vendor-specific ops.
+    """
+    from liger_kernel.ops.backends import get_vendor_for_device
+    from liger_kernel.utils import infer_device
+
+    device = infer_device()
+
+    # Look up vendor info for this device
+    vendor_info = get_vendor_for_device(device)
+    if vendor_info is None:
+        return
+
+    try:
+        import importlib
+
+        vendor_ops = importlib.import_module(vendor_info.module_path)
+
+        # Get names to export: use __all__ if defined, otherwise auto-discover
+        names_to_export = getattr(vendor_ops, "__all__", None)
+
+        if names_to_export is None:
+            # Auto-discover: find all public symbols (classes and functions)
+            names_to_export = [name for name in dir(vendor_ops) if not name.startswith("_")]
+
+        # Replace or add to this module's globals
+        for name in names_to_export:
+            globals()[name] = getattr(vendor_ops, name)
+
+    except ImportError:
+        # Vendor module not available, use default implementations
+        pass
+
+
+_replace_with_vendor_ops()

liger_kernel/ops/backends/README.md

@@ -0,0 +1,151 @@
+# Adding a New Vendor Backend
+
+This directory contains vendor-specific operator implementations that automatically replace the default (CUDA) implementations when running on the corresponding device.
+
+## Concepts
+
+- **Vendor**: Chip manufacturer (e.g., `ascend`, `intel`, `nvidia`)
+- **Device**: Device type (e.g., `npu`, `xpu`, `cuda`)
+- **VendorInfo**: Defines the mapping between vendor and device
+
+## Directory Structure
+
+```
+backends/
+├── README.md          
+├── __init__.py         
+├── registry.py         # VendorInfo, register_vendor(), VENDOR_REGISTRY
+├── _ascend/            # Ascend (Huawei) vendor - supports NPU
+│   ├── __init__.py     # Registers VendorInfo for NPU
+│   └── ops/
+│       ├── __init__.py # Exports vendor-specific implementations
+│       └── geglu.py    # NPU-specific GEGLU implementation
+└── _<vendor>/          # Your new vendor backend
+    └── ...
+```
+
+## How It Works
+
+1. When `liger_kernel.ops.backends` is imported, it imports all vendor packages (e.g., `_ascend`)
+2. Each vendor's `__init__.py` calls `register_vendor()` to register itself
+3. When `liger_kernel.ops` is imported, `_replace_with_vendor_ops()` is called
+4. It detects the current device via `infer_device()` and looks up the vendor
+5. Vendor implementations replace/add to the `liger_kernel.ops` namespace
+
+## Adding a New Vendor
+
+### Step 1: Create Directory Structure
+
+```bash
+mkdir -p backends/_<vendor>/ops
+touch backends/_<vendor>/__init__.py
+touch backends/_<vendor>/ops/__init__.py
+```
+
+### Step 2: Register Your Vendor
+
+In `backends/_<vendor>/__init__.py`, register your vendor:
+
+```python
+"""
+<Vendor> backend for Liger-Kernel.
+"""
+
+from liger_kernel.ops.backends.registry import VendorInfo, register_vendor
+
+register_vendor(
+    VendorInfo(
+        vendor="<vendor>",
+        device="<device>",
+    )
+)
+```
+
+
+### Step 3: Ensure Device Detection Works
+
+Make sure `infer_device()` in `liger_kernel/utils.py` can detect your device:
+
+```python
+def infer_device():
+    if torch.cuda.is_available():
+        return "cuda"
+    if is_npu_available():
+        return "npu"
+    # Add your device detection here
+    if is_<device>_available():
+        return "<device>"
+    return "cpu"
+```
+
+### Step 4: Implement Vendor-Specific Operators
+
+Create operator files in `backends/_<vendor>/ops/`. For example, `geglu.py`:
+
+```python
+import torch
+
+class LigerGELUMulFunction(torch.autograd.Function):
+    """
+    Vendor-specific LigerGELUMulFunction implementation.
+    """
+    @staticmethod
+    def forward(ctx, a, b):
+        # Your vendor-specific forward implementation
+        ...
+
+    @staticmethod
+    def backward(ctx, dc):
+        # Your vendor-specific backward implementation
+        ...
+
+# Optional: vendor-specific kernel functions
+def geglu_forward_vendor(a, b):
+    ...
+
+def geglu_backward_vendor(a, b, dc):
+    ...
+```
+
+### Step 5: Export in `ops/__init__.py`
+
+In `backends/_<vendor>/ops/__init__.py`, export your implementations:
+
+```python
+"""
+<Vendor>-specific operator implementations.
+"""
+
+from .<module> import (
+    LigerGELUMulFunction,
+    geglu_forward_vendor as geglu_forward,   # Rename to match default API
+    geglu_backward_vendor as geglu_backward,
+)
+
+# Explicitly declare what to export (recommended)
+__all__ = [
+    "LigerGELUMulFunction",
+    "geglu_forward",
+    "geglu_backward",
+]
+```
+
+## Key Points
+
+### Incremental Override
+
+You **don't need to implement all operators**. Only implement the ones that require vendor-specific adaptations. Unimplemented operators will automatically fall back to the default (CUDA) implementation.
+
+### Vendor-Specific Additions
+
+Vendors can also **add new operators** that don't exist in the default implementation. These will be exported to `liger_kernel.ops` namespace for users to import.
+
+### Naming Convention
+
+- Use the **same class/function names** as the default implementations for overrides
+- This allows seamless replacement without changing user code
+- Use `as` imports to rename if your internal naming differs
+
+## Example: Ascend NPU Backend
+
+See `_ascend/` directory for a complete example of the Ascend NPU backend implementation.

liger_kernel/ops/backends/__init__.py

@@ -0,0 +1,13 @@
+import importlib
+import pkgutil
+
+from liger_kernel.ops.backends.registry import VENDOR_REGISTRY  # noqa: F401
+from liger_kernel.ops.backends.registry import VendorInfo  # noqa: F401
+from liger_kernel.ops.backends.registry import get_vendor_for_device  # noqa: F401
+from liger_kernel.ops.backends.registry import register_vendor  # noqa: F401
+
+# Auto-import all _<vendor> subpackages to trigger registration
+# Each vendor's __init__.py calls register_vendor() when imported
+for _, modname, ispkg in pkgutil.iter_modules(__path__):
+    if ispkg and modname.startswith("_"):
+        importlib.import_module(f"{__name__}.{modname}")

liger_kernel/ops/backends/_ascend/__init__.py

@@ -0,0 +1,5 @@
+from liger_kernel.ops.backends.registry import VendorInfo
+from liger_kernel.ops.backends.registry import register_vendor
+
+# Register Ascend vendor for NPU device
+register_vendor(VendorInfo(vendor="ascend", device="npu"))