liger-kernels: sync with liger-kernels upstream

AGENTS.md

@@ -35,3 +35,44 @@ steps:
 
 If the user did not specify the version tag, stop and ask which tag to sync
 from.
+
+## liger-kernels
+
+When the user asks to sync a Liger-Kernel release, carry out the following
+steps:
+
+- Fetch the upstream Git repository from https://github.com/linkedin/Liger-Kernel.git
+- Check out the tag that the user specified (e.g. `v0.8.0`).
+- Liger-Kernel ops are in the directory `src/liger_kernel/ops` of the upstream
+  repo. There is no `cute` / `quack` dependency — only the Triton ops are
+  mirrored.
+- For each file already present in `liger-kernels/torch-ext/liger_kernels/`,
+  except `__init__.py` and `layers.py` (which are local additions, not
+  upstream), copy the matching file from `src/liger_kernel/ops/` upstream,
+  overwriting the local copy. Do not introduce new files that were not
+  previously synced — the local tree intentionally tracks a subset of upstream
+  ops.
+- Some helpers used by the ops live in the upstream top-level module
+  `src/liger_kernel/utils.py` (e.g. `is_npu_available`) rather than in
+  `src/liger_kernel/ops/utils.py`. If a synced op imports such a helper, fold
+  the helper into `liger-kernels/torch-ext/liger_kernels/utils.py` instead of
+  adding a separate file.
+- Now make all imports of Liger-Kernel in
+  `liger-kernels/torch-ext/liger_kernels` relative imports. In particular:
+  - `from liger_kernel.ops.<name> import ...` → `from .<name> import ...`
+  - `from liger_kernel.ops.utils import ...` → `from .utils import ...`
+  - `from liger_kernel.utils import ...` → `from .utils import ...` (after
+    folding the helper into the local `utils.py`).
+- Do not modify `liger-kernels/torch-ext/liger_kernels/__init__.py` or
+  `layers.py` — these are local `torch.nn.Module` wrappers that are not
+  derived from upstream.
+- Check whether any Torch custom ops are defined in the synced files (look
+  for `torch.library.custom_op`, `torch.library.define`, etc.). If any are
+  found, update them to use `add_op_namespace_prefix` for the op name,
+  importing it from `._ops` (see
+  `flash-attn3/torch-ext/flash_attn3/flash_attn_interface.py` for an
+  example). Upstream Liger-Kernel does not currently define any such ops, so
+  this step is usually a no-op.
+
+If the user did not specify the version tag, stop and ask which tag to sync
+from.

liger-kernels/flake.nix

@@ -2,7 +2,7 @@
   description = "Flake for Unsloth Kernels";
 
   inputs = {
-    kernel-builder.url = "github:huggingface/kernels/torch-2.12";
+    kernel-builder.url = "github:huggingface/kernels";
   };
 
   outputs =

liger-kernels/torch-ext/liger_kernels/__init__.py

@@ -1,3 +1,4 @@
 from . import layers
+from .layers import CrossEntropyOutput, LigerForCausalLMLoss
 
-__all__ = ["layers"]
+__all__ = ["layers", "LigerForCausalLMLoss", "CrossEntropyOutput"]