Add build/ variants to main

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-rocm70-x86_64-linux/_tail_hip_20260703003713.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-rocm71-x86_64-linux/_tail_hip_20260703003713.abi3.so filter=lfs diff=lfs merge=lfs -text

build/torch210-cxx11-rocm70-x86_64-linux/__init__.py

@@ -0,0 +1,47 @@
+"""tail_hip — native HIP "tail" elementwise kernels for gfx1201 (RDNA4).
+
+The kernel-builder / `kernels`-hub packaging of the kernel formerly vendored at minisgl-rdna4/tail_hip.
+Framework-agnostic (shared by minisgl-rdna4 and vLLM), AOT-compiled once (no Triton) to cut per-op
+launch/alloc overhead on the launch-sensitive decode path. Load with `kernels.get_kernel("<repo-id>")`
+or import the built package directly. Ops are exposed both as callables here and (for torch.compile)
+as opaque custom ops with registered fake/meta impls.
+
+Ops: rms_norm, rms_norm_add, silu_and_mul, rope. Conventions match minisgl-rdna4 (plus_one RMSNorm
+gain, gated silu over [...,2D], NeoX partial RoPE with cat(cos,sin) cache).
+"""
+import torch
+
+from ._ops import add_op_namespace_prefix, ops
+
+# ---- fake/meta impls: keep torch.compile/Inductor from graph-breaking on the opaque custom ops ----
+
+
+@torch.library.register_fake(add_op_namespace_prefix("rms_norm"))
+def _rms_norm_fake(x, w, eps, plus_one):
+    return torch.empty_like(x)
+
+
+@torch.library.register_fake(add_op_namespace_prefix("rms_norm_add"))
+def _rms_norm_add_fake(x, residual, w, eps, plus_one):
+    return torch.empty_like(x)
+
+
+@torch.library.register_fake(add_op_namespace_prefix("silu_and_mul"))
+def _silu_and_mul_fake(x):
+    shape = list(x.shape)
+    shape[-1] //= 2
+    return x.new_empty(shape)
+
+
+@torch.library.register_fake(add_op_namespace_prefix("rope"))
+def _rope_fake(x, pos, cache, head_size, rotary_dim):
+    return torch.empty_like(x)
+
+
+# ---- op wrappers ----
+rms_norm = ops.rms_norm
+rms_norm_add = ops.rms_norm_add
+silu_and_mul = ops.silu_and_mul
+rope = ops.rope
+
+__all__ = ["rms_norm", "rms_norm_add", "silu_and_mul", "rope"]

build/torch210-cxx11-rocm70-x86_64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _tail_hip_20260703003713
+ops = torch.ops._tail_hip_20260703003713
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_tail_hip_20260703003713::{op_name}"

build/torch210-cxx11-rocm70-x86_64-linux/_tail_hip_20260703003713.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a896ac92a978c91556d94b2979612e5c87672be9bd2cf9c9a2b2ca075041d4ca
+size 1971408

build/torch210-cxx11-rocm70-x86_64-linux/metadata.json

@@ -0,0 +1,4 @@
+{
+  "version": 1,
+  "python-depends": []
+}

build/torch210-cxx11-rocm70-x86_64-linux/tail_hip/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import sys
+
+import importlib
+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/torch210-cxx11-rocm71-x86_64-linux/__init__.py

@@ -0,0 +1,47 @@
+"""tail_hip — native HIP "tail" elementwise kernels for gfx1201 (RDNA4).
+
+The kernel-builder / `kernels`-hub packaging of the kernel formerly vendored at minisgl-rdna4/tail_hip.
+Framework-agnostic (shared by minisgl-rdna4 and vLLM), AOT-compiled once (no Triton) to cut per-op
+launch/alloc overhead on the launch-sensitive decode path. Load with `kernels.get_kernel("<repo-id>")`
+or import the built package directly. Ops are exposed both as callables here and (for torch.compile)
+as opaque custom ops with registered fake/meta impls.
+
+Ops: rms_norm, rms_norm_add, silu_and_mul, rope. Conventions match minisgl-rdna4 (plus_one RMSNorm
+gain, gated silu over [...,2D], NeoX partial RoPE with cat(cos,sin) cache).
+"""
+import torch
+
+from ._ops import add_op_namespace_prefix, ops
+
+# ---- fake/meta impls: keep torch.compile/Inductor from graph-breaking on the opaque custom ops ----
+
+
+@torch.library.register_fake(add_op_namespace_prefix("rms_norm"))
+def _rms_norm_fake(x, w, eps, plus_one):
+    return torch.empty_like(x)
+
+
+@torch.library.register_fake(add_op_namespace_prefix("rms_norm_add"))
+def _rms_norm_add_fake(x, residual, w, eps, plus_one):
+    return torch.empty_like(x)
+
+
+@torch.library.register_fake(add_op_namespace_prefix("silu_and_mul"))
+def _silu_and_mul_fake(x):
+    shape = list(x.shape)
+    shape[-1] //= 2
+    return x.new_empty(shape)
+
+
+@torch.library.register_fake(add_op_namespace_prefix("rope"))
+def _rope_fake(x, pos, cache, head_size, rotary_dim):
+    return torch.empty_like(x)
+
+
+# ---- op wrappers ----
+rms_norm = ops.rms_norm
+rms_norm_add = ops.rms_norm_add
+silu_and_mul = ops.silu_and_mul
+rope = ops.rope
+
+__all__ = ["rms_norm", "rms_norm_add", "silu_and_mul", "rope"]

build/torch210-cxx11-rocm71-x86_64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _tail_hip_20260703003713
+ops = torch.ops._tail_hip_20260703003713
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_tail_hip_20260703003713::{op_name}"

build/torch210-cxx11-rocm71-x86_64-linux/_tail_hip_20260703003713.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27b38cec13446ffeb8b65ad46081fa2fc13c417d8830db7a26a6dc47ffb6082e
+size 1971312

build/torch210-cxx11-rocm71-x86_64-linux/metadata.json

@@ -0,0 +1,4 @@
+{
+  "version": 1,
+  "python-depends": []
+}

build/torch210-cxx11-rocm71-x86_64-linux/tail_hip/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import sys
+
+import importlib
+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")))