Build uploaded using `kernels` (batch 9/10).

.gitattributes

@@ -45,3 +45,4 @@ build/torch210-cxx11-cu130-aarch64-linux/_deep_gemm_cuda_8546a43.abi3.so filter=
 build/torch211-cxx11-cu126-aarch64-linux/_deep_gemm_cuda_8546a43.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch211-cxx11-cu128-aarch64-linux/_deep_gemm_cuda_8546a43.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch211-cxx11-cu130-aarch64-linux/_deep_gemm_cuda_8546a43.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch29-cxx11-cu129-aarch64-linux/_deep_gemm_cuda_8546a43.abi3.so filter=lfs diff=lfs merge=lfs -text

build/torch29-cxx11-cu129-aarch64-linux/__init__.py

@@ -3,12 +3,42 @@ import subprocess
 import torch
 
 # Import the compiled extension
-from ._ops import ops
+from ._ops import ops, add_op_namespace_prefix
 from . import utils
 
 __version__ = "2.3.0"
 
 
+# ── Register fake tensor implementations for torch.compile ──────────────────
+# All GEMM ops mutate the output tensor `d` in-place and return void.
+# The fake implementations are no-ops since `d` is pre-allocated by the caller.
+
+
+for _op in [
+    "fp8_fp4_gemm_nt",
+    "fp8_fp4_gemm_nn",
+    "fp8_fp4_gemm_tn",
+    "fp8_fp4_gemm_tt",
+    "m_grouped_fp8_fp4_gemm_nt_contiguous",
+    "m_grouped_fp8_fp4_gemm_nn_contiguous",
+    "m_grouped_fp8_fp4_gemm_nt_masked",
+    "k_grouped_fp8_gemm_nt_contiguous",
+    "k_grouped_fp8_gemm_tn_contiguous",
+    "bf16_gemm_nt",
+    "bf16_gemm_nn",
+    "bf16_gemm_tn",
+    "bf16_gemm_tt",
+    "m_grouped_bf16_gemm_nt_contiguous",
+    "m_grouped_bf16_gemm_nn_contiguous",
+    "m_grouped_bf16_gemm_nt_masked",
+    "fp8_gemm_nt_skip_head_mid",
+]:
+
+    @torch.library.register_fake(add_op_namespace_prefix(_op))
+    def _fake(*args, **kwargs):
+        pass
+
+
 # Runtime
 
 
@@ -661,12 +691,14 @@ if "DG_CUTLASS_INCLUDE" not in os.environ:
         # Fall back to nvidia-cutlass pip package
         try:
             import nvidia.cutlass as _nc
+
             os.environ["DG_CUTLASS_INCLUDE"] = os.path.join(
                 os.path.dirname(_nc.__file__), "include"
             )
         except ImportError:
             pass
 
+
 def _ensure_initialized():
     global _initialized
     if _initialized:

build/torch29-cxx11-cu129-aarch64-linux/_deep_gemm_cuda_8546a43.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b062d4493fcefc08031fc464876d1ae5dd130cb3ce9cfca73af484225e0b5269
+size 2891904

build/torch29-cxx11-cu129-aarch64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _deep_gemm_cuda_bf91a5a
-ops = torch.ops._deep_gemm_cuda_bf91a5a
+from . import _deep_gemm_cuda_8546a43
+ops = torch.ops._deep_gemm_cuda_8546a43
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_deep_gemm_cuda_bf91a5a::{op_name}"
+    return f"_deep_gemm_cuda_8546a43::{op_name}"