Refresh the Metal builds for the 20-type table

Built from a clean tree at 0becb0a, so the variants carry that revision rather than a dirty stamp.
104 tests pass against torch212-metal on this machine.

.gitattributes

@@ -48,7 +48,5 @@ build/torch211-cxx11-cu130-x86_64-linux/_gguf_kernels_cuda_e4e2164.abi3.so filte
 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_8eddc6f.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch213-metal-aarch64-darwin/_gguf_kernels_metal_8eddc6f.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch212-metal-aarch64-darwin/_gguf_kernels_metal_49187c6.abi3.so filter=lfs diff=lfs merge=lfs -text
-build/torch213-metal-aarch64-darwin/_gguf_kernels_metal_49187c6.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-metal-aarch64-darwin/__init__.py

@@ -20,9 +20,16 @@ __all__ = ["GEMV_TYPES", "MAX_GEMV_ROWS", "dequantize", "mul_mat_vec"]
 # caller with more (prefill) dequantizes and uses an ordinary matmul.
 MAX_GEMV_ROWS = 8
 
-# ggml type ids `mul_mat_vec` implements: Q4_0/Q4_1/Q5_0/Q5_1/Q8_0, the K quants, and the IQ quants.
-# `dequantize` covers more, so check this one before choosing the fused path.
-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})
+# ggml type ids `mul_mat_vec` implements, as this build reports them -- it is backend-specific, and
+# a type routed into a gemv that has no kernel for it is a fault, not a fallback. CUDA covers
+# Q4_0/Q4_1/Q5_0/Q5_1/Q8_0, the K quants, the IQ quants, MXFP4/NVFP4 and Q1_0/Q2_0; Metal is the
+# same minus NVFP4/Q1_0/Q2_0, which its metallib has no kernels for. `dequantize` covers more, so
+# check this before choosing the fused path.
+try:
+    GEMV_TYPES = frozenset(ops.gemv_types())
+except AttributeError:
+    # A build published before `gemv_types` existed; those are CUDA-only, so its list is theirs.
+    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 dequantize(

build/torch212-metal-aarch64-darwin/{_gguf_kernels_metal_49187c6.abi3.so → _gguf_kernels_metal_0becb0a.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f12c8c93dcb56534b8a0f0a25c15917de5529897d0113fb415545a0619f0819a
-size 7378688
+oid sha256:b369cc5b468e2fd3b9ac84f5c63897a52bc25e0381c37104d11d0956f5ad4dd9
+size 7403104

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

@@ -1,9 +1,9 @@
 import torch
-from . import _gguf_kernels_metal_49187c6
-ops = torch.ops._gguf_kernels_metal_49187c6
+from . import _gguf_kernels_metal_0becb0a
+ops = torch.ops._gguf_kernels_metal_0becb0a
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_gguf_kernels_metal_49187c6::{op_name}"
+    return f"_gguf_kernels_metal_0becb0a::{op_name}"

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

@@ -1,6 +1,6 @@
 {
   "name": "gguf-kernels",
-  "id": "_gguf_kernels_metal_49187c6",
+  "id": "_gguf_kernels_metal_0becb0a",
   "version": 1,
   "license": "MIT",
   "python-depends": [],
@@ -10,9 +10,9 @@
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "__init__.py": "Xl1/3Qvp7VYrhYXje6GLBwPa9H7ICY5jHKLc7frZvlo=",
-      "_gguf_kernels_metal_49187c6.abi3.so": "8SyMk9y1ZTS4oPCiXBWRfeVSmJfQET+0FVRaBhnwgZo=",
-      "_ops.py": "3/oAViDUMhdbtPnSJwwu0q4n/n9y0JN9OdT5E9FDY5o="
+      "__init__.py": "GoFcHJbbJ9C8b40XNznMxsTcC+CV9kaFE8a1DP763kU=",
+      "_gguf_kernels_metal_0becb0a.abi3.so": "s2nMW0aOL9O5rIT1xjiXpSvCXgOBw3EE0R0JVvWtTdk=",
+      "_ops.py": "3htLDPFBO8zVroZkTqPmryb9pDPjZ5OuMItEBg94Jp4="
     }
   },
   "provenance": {
@@ -22,7 +22,7 @@
       "dirty": false
     },
     "kernel": {
-      "sha": "49187c66b4f3baba2041bf2df9258c3bf8473110",
+      "sha": "0becb0aafe12b27da05fa08f4b6cc7c72b288fbb",
       "dirty": false
     }
   }

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

@@ -20,9 +20,16 @@ __all__ = ["GEMV_TYPES", "MAX_GEMV_ROWS", "dequantize", "mul_mat_vec"]
 # caller with more (prefill) dequantizes and uses an ordinary matmul.
 MAX_GEMV_ROWS = 8
 
-# ggml type ids `mul_mat_vec` implements: Q4_0/Q4_1/Q5_0/Q5_1/Q8_0, the K quants, and the IQ quants.
-# `dequantize` covers more, so check this one before choosing the fused path.
-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})
+# ggml type ids `mul_mat_vec` implements, as this build reports them -- it is backend-specific, and
+# a type routed into a gemv that has no kernel for it is a fault, not a fallback. CUDA covers
+# Q4_0/Q4_1/Q5_0/Q5_1/Q8_0, the K quants, the IQ quants, MXFP4/NVFP4 and Q1_0/Q2_0; Metal is the
+# same minus NVFP4/Q1_0/Q2_0, which its metallib has no kernels for. `dequantize` covers more, so
+# check this before choosing the fused path.
+try:
+    GEMV_TYPES = frozenset(ops.gemv_types())
+except AttributeError:
+    # A build published before `gemv_types` existed; those are CUDA-only, so its list is theirs.
+    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 dequantize(

build/torch213-metal-aarch64-darwin/{_gguf_kernels_metal_49187c6.abi3.so → _gguf_kernels_metal_0becb0a.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fa7e74342da59f75783f065ad4557ca3fd197ed37c759763017e01121833d6ec
-size 7378688
+oid sha256:59d0c2717f44308bc965b82e520cce45b0a6f7e496bc9e95dfc2367dbc75f441
+size 7403104

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

@@ -1,9 +1,9 @@
 import torch
-from . import _gguf_kernels_metal_49187c6
-ops = torch.ops._gguf_kernels_metal_49187c6
+from . import _gguf_kernels_metal_0becb0a
+ops = torch.ops._gguf_kernels_metal_0becb0a
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_gguf_kernels_metal_49187c6::{op_name}"
+    return f"_gguf_kernels_metal_0becb0a::{op_name}"

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

@@ -1,6 +1,6 @@
 {
   "name": "gguf-kernels",
-  "id": "_gguf_kernels_metal_49187c6",
+  "id": "_gguf_kernels_metal_0becb0a",
   "version": 1,
   "license": "MIT",
   "python-depends": [],
@@ -10,9 +10,9 @@
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "__init__.py": "Xl1/3Qvp7VYrhYXje6GLBwPa9H7ICY5jHKLc7frZvlo=",
-      "_gguf_kernels_metal_49187c6.abi3.so": "+n50NC2ln3V4PwZa1FV8o/0ZftN8dZdjAX4BEhgz1uw=",
-      "_ops.py": "3/oAViDUMhdbtPnSJwwu0q4n/n9y0JN9OdT5E9FDY5o="
+      "__init__.py": "GoFcHJbbJ9C8b40XNznMxsTcC+CV9kaFE8a1DP763kU=",
+      "_gguf_kernels_metal_0becb0a.abi3.so": "WdDCcX9EMIvJZbguUgzORbCm9+SWvJ6V38I2fbx19EE=",
+      "_ops.py": "3htLDPFBO8zVroZkTqPmryb9pDPjZ5OuMItEBg94Jp4="
     }
   },
   "provenance": {
@@ -22,7 +22,7 @@
       "dirty": false
     },
     "kernel": {
-      "sha": "49187c66b4f3baba2041bf2df9258c3bf8473110",
+      "sha": "0becb0aafe12b27da05fa08f4b6cc7c72b288fbb",
       "dirty": false
     }
   }