rebuild metal for get_rows, and track artifacts by glob

The published `.so` predates `get_rows`, so a client that asks for it gets an AttributeError and
falls back to unpacking whole tensors. Rebuilt from a clean tree, so the module carries the commit
it was built from rather than a `_dirty` name.

`.gitattributes` listed every artifact by path, one line per rebuild, and the entries outlived the
files. One glob covers whatever hash the next build stamps in.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

.gitattributes

@@ -33,28 +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-cu126-x86_64-linux/_gguf_kernels_20260805124229.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch210-cxx11-cu128-x86_64-linux/_gguf_kernels_20260805124229.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch210-cxx11-cu130-x86_64-linux/_gguf_kernels_20260805124229.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch28-cxx11-cu126-x86_64-linux/_gguf_kernels_20260805124229.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch28-cxx11-cu128-x86_64-linux/_gguf_kernels_20260805124229.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch28-cxx11-cu129-x86_64-linux/_gguf_kernels_20260805124229.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch29-cxx11-cu126-x86_64-linux/_gguf_kernels_20260805124229.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch29-cxx11-cu128-x86_64-linux/_gguf_kernels_20260805124229.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch29-cxx11-cu130-x86_64-linux/_gguf_kernels_20260805124229.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch211-cxx11-cu126-x86_64-linux/_gguf_kernels_cuda_e4e2164.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch211-cxx11-cu128-x86_64-linux/_gguf_kernels_cuda_e4e2164.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch211-cxx11-cu130-x86_64-linux/_gguf_kernels_cuda_e4e2164.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch212-cxx11-cu126-x86_64-linux/_gguf_kernels_cuda_e4e2164.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch212-cxx11-cu130-x86_64-linux/_gguf_kernels_cuda_e4e2164.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch212-cxx11-cu132-x86_64-linux/_gguf_kernels_cuda_e4e2164.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch212-metal-aarch64-darwin/_gguf_kernels_metal_0becb0a.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch213-metal-aarch64-darwin/_gguf_kernels_metal_0becb0a.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch212-cxx11-cu126-x86_64-linux/_gguf_kernels_cuda_1df3b99.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch212-cxx11-cu130-x86_64-linux/_gguf_kernels_cuda_1df3b99.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch212-cxx11-cu132-x86_64-linux/_gguf_kernels_cuda_1df3b99.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch213-cxx11-cu126-x86_64-linux/_gguf_kernels_cuda_1df3b99.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch213-cxx11-cu130-x86_64-linux/_gguf_kernels_cuda_1df3b99.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch213-cxx11-cu132-x86_64-linux/_gguf_kernels_cuda_1df3b99.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch212-metal-aarch64-darwin/_gguf_kernels_metal_7b10656.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch213-metal-aarch64-darwin/_gguf_kernels_metal_7b10656.abi3.so filter=lfs diff=lfs merge=lfs -text
+# Every built artifact, whatever hash the build stamps into its name.
+build/**/*.abi3.so filter=lfs diff=lfs merge=lfs -text

build/torch212-metal-aarch64-darwin/__init__.py

@@ -14,7 +14,7 @@ import torch
 from ._ops import add_op_namespace_prefix, ops
 
 
-__all__ = ["GEMV_TYPES", "MAX_GEMV_ROWS", "dequantize", "mul_mat_vec"]
+__all__ = ["GEMV_TYPES", "MAX_GEMV_ROWS", "dequantize", "get_rows", "mul_mat_vec"]
 
 # Upstream's MMVQ_MAX_BATCH_SIZE: `mul_mat_vec` has no implementation beyond this many rows, so a
 # caller with more (prefill) dequantizes and uses an ordinary matmul.
@@ -32,6 +32,17 @@ except AttributeError:
     GEMV_TYPES = frozenset({2, 3, 6, 7, 8, 10, 11, 12, 13, 14, 16, 17, 18, 19, 20, 21, 22, 23, 29, 39, 40, 41, 42})
 
 
+def get_rows(
+    blocks: torch.Tensor, indices: torch.Tensor, ggml_type: int, cols: int, dtype: torch.dtype
+) -> torch.Tensor:
+    """The rows `indices` names, unpacked: `(rows, bytes_per_row)` uint8 -> `(len(indices), cols)`.
+
+    ggml's `get_rows`, which dequantizes as it gathers -- so reading a few rows out of a large table
+    never touches the rest.
+    """
+    return ops.get_rows(blocks, indices, ggml_type, cols, dtype)
+
+
 def dequantize(
     blocks: torch.Tensor, ggml_type: int, rows: int, cols: int, dtype: torch.dtype
 ) -> torch.Tensor:
@@ -52,6 +63,11 @@ def mul_mat_vec(
 
 
 # Without these, torch.compile cannot trace the ops and breaks the graph at every call.
+@torch.library.register_fake(add_op_namespace_prefix("get_rows"))
+def _get_rows_fake(blocks, indices, ggml_type, cols, dtype):
+    return blocks.new_empty((indices.numel(), cols), dtype=dtype)
+
+
 @torch.library.register_fake(add_op_namespace_prefix("dequantize"))
 def _dequantize_fake(blocks, ggml_type, rows, cols, dtype):
     return blocks.new_empty((rows, cols), dtype=dtype)

build/torch212-metal-aarch64-darwin/{_gguf_kernels_metal_7b10656.abi3.so → _ggml_quantization_metal_1f05408.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d5ea02aab35301ff5a378d86de62477becd6604cd56ef6000802c15677c9375d
-size 814816
+oid sha256:d731041e5f08c864b02363a29e9431fc8343d20c3e809456e83aedcb0108d609
+size 817976

build/torch212-metal-aarch64-darwin/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _gguf_kernels_metal_7b10656
-ops = torch.ops._gguf_kernels_metal_7b10656
+from . import _ggml_quantization_metal_1f05408
+ops = torch.ops._ggml_quantization_metal_1f05408
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_gguf_kernels_metal_7b10656::{op_name}"
+    return f"_ggml_quantization_metal_1f05408::{op_name}"

build/torch212-metal-aarch64-darwin/metadata.json

@@ -1,6 +1,6 @@
 {
-  "name": "gguf-kernels",
-  "id": "_gguf_kernels_metal_7b10656",
+  "name": "ggml-quantization",
+  "id": "_ggml_quantization_metal_1f05408",
   "version": 1,
   "license": "MIT",
   "python-depends": [],
@@ -10,9 +10,9 @@
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "__init__.py": "GoFcHJbbJ9C8b40XNznMxsTcC+CV9kaFE8a1DP763kU=",
-      "_gguf_kernels_metal_7b10656.abi3.so": "1eoCqrNTAf9aN42G3mJHe+zWYEzVbvYACALBVnfJN10=",
-      "_ops.py": "nwCRMS2OmSxach0Edkf8TLdxZXmrrMhMJntgxLN2MAI="
+      "__init__.py": "68gDnPDp55W5ZSVOVBPJqML9O/YHq9mzcdCvNs4rnEg=",
+      "_ggml_quantization_metal_1f05408.abi3.so": "1zEEHl8IyGSwI2OinpQx/IND0gw+gJRW6DrtywEI1gk=",
+      "_ops.py": "jWnZSZuckU1udyY5AEqxJlvw8JDXMqm58DgbvdbdlpI="
     }
   },
   "provenance": {
@@ -22,7 +22,7 @@
       "dirty": false
     },
     "kernel": {
-      "sha": "7b106564513dead0e4416ea0c7d3898eb23e0ca2",
+      "sha": "1f0540807daf50af68e13f6e6db55bf3440101b3",
       "dirty": false
     }
   }

build/torch213-metal-aarch64-darwin/__init__.py

@@ -14,7 +14,7 @@ import torch
 from ._ops import add_op_namespace_prefix, ops
 
 
-__all__ = ["GEMV_TYPES", "MAX_GEMV_ROWS", "dequantize", "mul_mat_vec"]
+__all__ = ["GEMV_TYPES", "MAX_GEMV_ROWS", "dequantize", "get_rows", "mul_mat_vec"]
 
 # Upstream's MMVQ_MAX_BATCH_SIZE: `mul_mat_vec` has no implementation beyond this many rows, so a
 # caller with more (prefill) dequantizes and uses an ordinary matmul.
@@ -32,6 +32,17 @@ except AttributeError:
     GEMV_TYPES = frozenset({2, 3, 6, 7, 8, 10, 11, 12, 13, 14, 16, 17, 18, 19, 20, 21, 22, 23, 29, 39, 40, 41, 42})
 
 
+def get_rows(
+    blocks: torch.Tensor, indices: torch.Tensor, ggml_type: int, cols: int, dtype: torch.dtype
+) -> torch.Tensor:
+    """The rows `indices` names, unpacked: `(rows, bytes_per_row)` uint8 -> `(len(indices), cols)`.
+
+    ggml's `get_rows`, which dequantizes as it gathers -- so reading a few rows out of a large table
+    never touches the rest.
+    """
+    return ops.get_rows(blocks, indices, ggml_type, cols, dtype)
+
+
 def dequantize(
     blocks: torch.Tensor, ggml_type: int, rows: int, cols: int, dtype: torch.dtype
 ) -> torch.Tensor:
@@ -52,6 +63,11 @@ def mul_mat_vec(
 
 
 # Without these, torch.compile cannot trace the ops and breaks the graph at every call.
+@torch.library.register_fake(add_op_namespace_prefix("get_rows"))
+def _get_rows_fake(blocks, indices, ggml_type, cols, dtype):
+    return blocks.new_empty((indices.numel(), cols), dtype=dtype)
+
+
 @torch.library.register_fake(add_op_namespace_prefix("dequantize"))
 def _dequantize_fake(blocks, ggml_type, rows, cols, dtype):
     return blocks.new_empty((rows, cols), dtype=dtype)

build/torch213-metal-aarch64-darwin/{_gguf_kernels_metal_7b10656.abi3.so → _ggml_quantization_metal_1f05408.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e6dbb2eaefd58d768f8a9e30e89b08f1f7a139648b12fb8924a34caccbd8d066
-size 814816
+oid sha256:fd4351ef7e2f79914badfece93f4d372e5146f8bbb1b4b7399fadffd8634f90c
+size 817976

build/torch213-metal-aarch64-darwin/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _gguf_kernels_metal_7b10656
-ops = torch.ops._gguf_kernels_metal_7b10656
+from . import _ggml_quantization_metal_1f05408
+ops = torch.ops._ggml_quantization_metal_1f05408
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_gguf_kernels_metal_7b10656::{op_name}"
+    return f"_ggml_quantization_metal_1f05408::{op_name}"

build/torch213-metal-aarch64-darwin/metadata.json

@@ -1,6 +1,6 @@
 {
-  "name": "gguf-kernels",
-  "id": "_gguf_kernels_metal_7b10656",
+  "name": "ggml-quantization",
+  "id": "_ggml_quantization_metal_1f05408",
   "version": 1,
   "license": "MIT",
   "python-depends": [],
@@ -10,9 +10,9 @@
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "__init__.py": "GoFcHJbbJ9C8b40XNznMxsTcC+CV9kaFE8a1DP763kU=",
-      "_gguf_kernels_metal_7b10656.abi3.so": "5tuy6u/VjXaPip4w6JsI8fehOWSLEvuJJKNMrMvY0GY=",
-      "_ops.py": "nwCRMS2OmSxach0Edkf8TLdxZXmrrMhMJntgxLN2MAI="
+      "__init__.py": "68gDnPDp55W5ZSVOVBPJqML9O/YHq9mzcdCvNs4rnEg=",
+      "_ggml_quantization_metal_1f05408.abi3.so": "/UNR734veZFLrf7Ok/TTcuUUb4u7G0tzmfrf/YY0+Qw=",
+      "_ops.py": "jWnZSZuckU1udyY5AEqxJlvw8JDXMqm58DgbvdbdlpI="
     }
   },
   "provenance": {
@@ -22,7 +22,7 @@
       "dirty": false
     },
     "kernel": {
-      "sha": "7b106564513dead0e4416ea0c7d3898eb23e0ca2",
+      "sha": "1f0540807daf50af68e13f6e6db55bf3440101b3",
       "dirty": false
     }
   }