Promote latest kernel artifacts to main

README.md

@@ -1,9 +0,0 @@
-# flashrt/flashrt-nvfp4
-
-This repository is a compatibility mirror for older `kernels` clients
-that resolve repositories through the default Hugging Face model repo API.
-
-Canonical Kernel Hub repo: https://huggingface.co/kernels/flashrt/flashrt-nvfp4
-
-Do not edit this mirror by hand. It is generated from the Kernel Hub
-`vN` branches and contains the same `build/**` artifacts.

build/{torch211-cxx11-cu130-x86_64-linux/_flashrt_nvfp4_cuda_c4d802d.abi3.so → torch211-cxx11-cu128-x86_64-linux/_flashrt_nvfp4_cuda_9e33734.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2d8e6c0b823b0bebe75d49f7a9bd4806378107c013e043df41ca3d587c55e913
-size 100344
+oid sha256:f1c2e72487799c2955fffa6ae64245e6a690211a52c1abde76f5f7e189fca764
+size 95168

build/torch211-cxx11-cu128-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flashrt_nvfp4_cuda_c4d802d
-ops = torch.ops._flashrt_nvfp4_cuda_c4d802d
+from . import _flashrt_nvfp4_cuda_9e33734
+ops = torch.ops._flashrt_nvfp4_cuda_9e33734
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flashrt_nvfp4_cuda_c4d802d::{op_name}"
+    return f"_flashrt_nvfp4_cuda_9e33734::{op_name}"

build/torch211-cxx11-cu128-x86_64-linux/metadata.json

@@ -1,13 +1,13 @@
 {
   "name": "flashrt-nvfp4",
-  "id": "_flashrt_nvfp4_cuda_c4d802d",
+  "id": "_flashrt_nvfp4_cuda_9e33734",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],
   "backend": {
     "type": "cuda",
     "archs": [
-      "12.0"
+      "12.0a"
     ]
   }
 }

build/torch211-cxx11-cu130-aarch64-linux/__init__.py

@@ -0,0 +1,67 @@
+"""FlashRT NVFP4 layout kernels."""
+
+from __future__ import annotations
+
+from typing import Optional
+
+import torch
+
+from ._ops import add_op_namespace_prefix, ops
+
+
+@torch.library.register_fake(add_op_namespace_prefix("nvfp4_sf_linear_to_swizzled"))
+def _nvfp4_sf_linear_to_swizzled_fake(
+    scales: torch.Tensor,
+    out: torch.Tensor,
+    D: int,
+    is_sfb: bool = False,
+) -> None:
+    if scales.dim() != 2:
+        raise RuntimeError("scales must have shape (rows, D / 16)")
+    return None
+
+
+def nvfp4_sf_swizzled_bytes(rows: int, D: int) -> int:
+    """Return byte count for a CUTLASS Sm1xx NVFP4 swizzled SF buffer."""
+
+    if rows <= 0:
+        raise ValueError("rows must be positive")
+    if D <= 0 or D % 16 != 0:
+        raise ValueError("D must be positive and divisible by 16")
+    n_blocks = D // 16
+    n_row_super = (rows + 127) // 128
+    n_col_super = (n_blocks + 3) // 4
+    return n_row_super * n_col_super * 512
+
+
+def nvfp4_sf_linear_to_swizzled(
+    scales: torch.Tensor,
+    *,
+    out: Optional[torch.Tensor] = None,
+    is_sfb: bool = False,
+) -> torch.Tensor:
+    """Convert linear NVFP4 scale bytes to CUTLASS Sm1xx swizzled layout.
+
+    ``scales`` must be contiguous CUDA ``torch.uint8`` with shape
+    ``(rows, D / 16)``. If ``out`` is omitted, a flat ``torch.uint8`` output
+    tensor with ``nvfp4_sf_swizzled_bytes(rows, D)`` bytes is allocated.
+    """
+
+    if scales.dim() != 2:
+        raise ValueError("scales must have shape (rows, D / 16)")
+    rows = scales.shape[0]
+    D = scales.shape[1] * 16
+    if out is None:
+        out = torch.zeros(
+            (nvfp4_sf_swizzled_bytes(rows, D),),
+            device=scales.device,
+            dtype=torch.uint8,
+        )
+    ops.nvfp4_sf_linear_to_swizzled(scales, out, D, is_sfb)
+    return out
+
+
+__all__ = [
+    "nvfp4_sf_linear_to_swizzled",
+    "nvfp4_sf_swizzled_bytes",
+]

build/{torch212-cxx11-cu132-x86_64-linux → torch211-cxx11-cu130-aarch64-linux}/_flashrt_nvfp4_cuda_c4d802d.abi3.so

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:91f6297b7871d47faf300ef1308b9a340666ab3afaa6b46b9a2d4d314b77255f
-size 111104
+oid sha256:1c206040d7bd2911cca52e0c788a9c981e8161991db90f2870c906d210fdc628
+size 167824

build/torch211-cxx11-cu130-aarch64-linux/_ops.py

@@ -0,0 +1,6 @@
+import torch
+from . import _flashrt_nvfp4_cuda_c4d802d
+ops = torch.ops._flashrt_nvfp4_cuda_c4d802d
+
+def add_op_namespace_prefix(op_name: str):
+    return f"_flashrt_nvfp4_cuda_c4d802d::{op_name}"

build/torch211-cxx11-cu130-aarch64-linux/flashrt_nvfp4/__init__.py

@@ -0,0 +1,14 @@
+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+
+def _import_from_path(file_path: Path):
+    path_hash = '{:x}'.format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    spec = importlib.util.spec_from_file_location(path_hash, file_path)
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[path_hash] = module
+    spec.loader.exec_module(module)
+    return module
+
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / '__init__.py')))

build/torch211-cxx11-cu130-aarch64-linux/metadata.json

@@ -0,0 +1,32 @@
+{
+  "name": "flashrt-nvfp4",
+  "id": "_flashrt_nvfp4_cuda_c4d802d",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "11.0a"
+    ]
+  },
+  "digest": {
+    "algorithm": "sha256",
+    "files": {
+      "__init__.py": "aB242czTWLzSGSTrkJY75ln0cEILoGwF8bHScJPgvMc=",
+      "_flashrt_nvfp4_cuda_c4d802d.abi3.so": "HCBgQNe9KRHMpS4MeIqcmB6BYZkduQ8ocMkG0hD9xig=",
+      "_ops.py": "9FbAFa8lhnvONtloChdas+yY4tDVqcicGsvmPNAKMr4=",
+      "flashrt_nvfp4/__init__.py": "v6p5XMfQzddhi1fLSAw4HX9CyS0rQsidvu9VsT01xi4="
+    }
+  },
+  "provenance": {
+    "kernel": {
+      "sha": "9e33734",
+      "dirty": false
+    },
+    "validation": {
+      "torch": "2.11.0+cu130",
+      "cuda": "13.0"
+    }
+  }
+}

build/{torch212-cxx11-cu130-x86_64-linux/_flashrt_nvfp4_cuda_c4d802d.abi3.so → torch211-cxx11-cu130-x86_64-linux/_flashrt_nvfp4_cuda_9e33734.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8b0a24f8bbe4406e273fb631b33cabc777ef02ee23b23f1be618b3441bdc2ec9
-size 111104
+oid sha256:17c97cca9d88bd01b62199b514c722bb955c23a6290f40c875aa6a2eb31d9aa2
+size 104424

build/torch211-cxx11-cu130-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flashrt_nvfp4_cuda_c4d802d
-ops = torch.ops._flashrt_nvfp4_cuda_c4d802d
+from . import _flashrt_nvfp4_cuda_9e33734
+ops = torch.ops._flashrt_nvfp4_cuda_9e33734
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flashrt_nvfp4_cuda_c4d802d::{op_name}"
+    return f"_flashrt_nvfp4_cuda_9e33734::{op_name}"

build/torch211-cxx11-cu130-x86_64-linux/metadata.json

@@ -1,13 +1,14 @@
 {
   "name": "flashrt-nvfp4",
-  "id": "_flashrt_nvfp4_cuda_c4d802d",
+  "id": "_flashrt_nvfp4_cuda_9e33734",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],
   "backend": {
     "type": "cuda",
     "archs": [
-      "12.0"
+      "11.0a",
+      "12.0a"
     ]
   }
 }

build/torch212-cxx11-cu130-x86_64-linux/_flashrt_nvfp4_cuda_9e33734.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bc3ea1097842b95184b411c5ecad33a36e1d289a1ef65c67a60c48a71229220
+size 119280

build/torch212-cxx11-cu130-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flashrt_nvfp4_cuda_c4d802d
-ops = torch.ops._flashrt_nvfp4_cuda_c4d802d
+from . import _flashrt_nvfp4_cuda_9e33734
+ops = torch.ops._flashrt_nvfp4_cuda_9e33734
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flashrt_nvfp4_cuda_c4d802d::{op_name}"
+    return f"_flashrt_nvfp4_cuda_9e33734::{op_name}"

build/torch212-cxx11-cu130-x86_64-linux/metadata.json

@@ -1,13 +1,14 @@
 {
   "name": "flashrt-nvfp4",
-  "id": "_flashrt_nvfp4_cuda_c4d802d",
+  "id": "_flashrt_nvfp4_cuda_9e33734",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],
   "backend": {
     "type": "cuda",
     "archs": [
-      "12.0"
+      "11.0a",
+      "12.0a"
     ]
   }
 }

build/torch212-cxx11-cu132-x86_64-linux/_flashrt_nvfp4_cuda_9e33734.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf1db69d0f2ea2d97282d2999ec69cb339559e855eae29d2009463a50237fe20
+size 119280

build/torch212-cxx11-cu132-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flashrt_nvfp4_cuda_c4d802d
-ops = torch.ops._flashrt_nvfp4_cuda_c4d802d
+from . import _flashrt_nvfp4_cuda_9e33734
+ops = torch.ops._flashrt_nvfp4_cuda_9e33734
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flashrt_nvfp4_cuda_c4d802d::{op_name}"
+    return f"_flashrt_nvfp4_cuda_9e33734::{op_name}"

build/torch212-cxx11-cu132-x86_64-linux/metadata.json

@@ -1,13 +1,14 @@
 {
   "name": "flashrt-nvfp4",
-  "id": "_flashrt_nvfp4_cuda_c4d802d",
+  "id": "_flashrt_nvfp4_cuda_9e33734",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],
   "backend": {
     "type": "cuda",
     "archs": [
-      "12.0"
+      "11.0a",
+      "12.0a"
     ]
   }
 }