Uploaded using `kernel-builder`.

Browse files

Files changed (8) hide show

build/torch212-metal-aarch64-darwin/__init__.py +17 -1
build/torch212-metal-aarch64-darwin/{_gguf_kernels_metal_b4f7ed2.abi3.so → _ggml_quantization_metal_09da59d.abi3.so} +2 -2
build/torch212-metal-aarch64-darwin/_ops.py +3 -3
build/torch212-metal-aarch64-darwin/metadata.json +6 -6
build/torch213-metal-aarch64-darwin/__init__.py +17 -1
build/torch213-metal-aarch64-darwin/{_gguf_kernels_metal_b4f7ed2.abi3.so → _ggml_quantization_metal_09da59d.abi3.so} +2 -2
build/torch213-metal-aarch64-darwin/_ops.py +3 -3
build/torch213-metal-aarch64-darwin/metadata.json +6 -6

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

@@ -14,7 +14,7 @@ import torch
 from ._ops import add_op_namespace_prefix, ops
-__all__ = ["GEMV_TYPES", "MAX_GEMV_ROWS", "dequantize", "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 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("dequantize"))
 def _dequantize_fake(blocks, ggml_type, rows, cols, dtype):
     return blocks.new_empty((rows, cols), dtype=dtype)

 from ._ops import add_op_namespace_prefix, ops
+__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.
     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:
 # 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_b4f7ed2.abi3.so → _ggml_quantization_metal_09da59d.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d6f5cda7143d4352cb1671e340f7495aed1c0005ddef924f687e74117c15775b
-size 814816

 version https://git-lfs.github.com/spec/v1
+oid sha256:f3ec2c0bd762a1cb5855006a816cdb22f18c11e969006d17131460e35e577ebb
+size 817976

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

@@ -1,9 +1,9 @@
 import torch
-from . import _gguf_kernels_metal_b4f7ed2
-ops = torch.ops._gguf_kernels_metal_b4f7ed2
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_gguf_kernels_metal_b4f7ed2::{op_name}"

 import torch
+from . import _ggml_quantization_metal_09da59d
+ops = torch.ops._ggml_quantization_metal_09da59d
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_ggml_quantization_metal_09da59d::{op_name}"

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

@@ -1,6 +1,6 @@
 {
-  "name": "gguf-kernels",
-  "id": "_gguf_kernels_metal_b4f7ed2",
   "version": 1,
   "license": "MIT",
   "python-depends": [],
@@ -10,9 +10,9 @@
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "__init__.py": "GoFcHJbbJ9C8b40XNznMxsTcC+CV9kaFE8a1DP763kU=",
-      "_gguf_kernels_metal_b4f7ed2.abi3.so": "1vXNpxQ9Q1LLFnHjQPdJWu0cAAXd75JPaH50EXwVd1s=",
-      "_ops.py": "nK7J3dCzRE3uvmtyRDGFaZJ7dOcht804Bgq4RVGex0E="
     }
   },
   "provenance": {
@@ -22,7 +22,7 @@
       "dirty": false
     },
     "kernel": {
-      "sha": "b4f7ed2480df20d04cd0e47b5a18518bfd712b6e",
       "dirty": false
     }
   }

 {
+  "name": "ggml-quantization",
+  "id": "_ggml_quantization_metal_09da59d",
   "version": 1,
   "license": "MIT",
   "python-depends": [],
   "digest": {
     "algorithm": "sha256",
     "files": {
+      "__init__.py": "68gDnPDp55W5ZSVOVBPJqML9O/YHq9mzcdCvNs4rnEg=",
+      "_ggml_quantization_metal_09da59d.abi3.so": "8+wsC9dioctYVQBqgWzbIvGMEelpAG0XExRg415Xfrs=",
+      "_ops.py": "KPScFwQZRvXLLvmzyYkY4+0hIJB+X5ZJ9opUxecMnZs="
     }
   },
   "provenance": {
       "dirty": false
     },
     "kernel": {
+      "sha": "09da59d51d3f73baa39a941eb07bf6c386223f82",
       "dirty": false
     }
   }

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

@@ -14,7 +14,7 @@ import torch
 from ._ops import add_op_namespace_prefix, ops
-__all__ = ["GEMV_TYPES", "MAX_GEMV_ROWS", "dequantize", "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 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("dequantize"))
 def _dequantize_fake(blocks, ggml_type, rows, cols, dtype):
     return blocks.new_empty((rows, cols), dtype=dtype)

 from ._ops import add_op_namespace_prefix, ops
+__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.
     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:
 # 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_b4f7ed2.abi3.so → _ggml_quantization_metal_09da59d.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6a58f765ed13fc4403ab86e82143dabadba2a63d0ebeba855787516f881189d8
-size 814816

 version https://git-lfs.github.com/spec/v1
+oid sha256:391156ef4b8ec8f557c1ecb59a89d716a289e41e1c3853c6f0566330aafe196a
+size 817976

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

@@ -1,9 +1,9 @@
 import torch
-from . import _gguf_kernels_metal_b4f7ed2
-ops = torch.ops._gguf_kernels_metal_b4f7ed2
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_gguf_kernels_metal_b4f7ed2::{op_name}"

 import torch
+from . import _ggml_quantization_metal_09da59d
+ops = torch.ops._ggml_quantization_metal_09da59d
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_ggml_quantization_metal_09da59d::{op_name}"

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

@@ -1,6 +1,6 @@
 {
-  "name": "gguf-kernels",
-  "id": "_gguf_kernels_metal_b4f7ed2",
   "version": 1,
   "license": "MIT",
   "python-depends": [],
@@ -10,9 +10,9 @@
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "__init__.py": "GoFcHJbbJ9C8b40XNznMxsTcC+CV9kaFE8a1DP763kU=",
-      "_gguf_kernels_metal_b4f7ed2.abi3.so": "alj3Ze0T/EQDq4boIUPautuipj0OvrqFV4dRb4gRidg=",
-      "_ops.py": "nK7J3dCzRE3uvmtyRDGFaZJ7dOcht804Bgq4RVGex0E="
     }
   },
   "provenance": {
@@ -22,7 +22,7 @@
       "dirty": false
     },
     "kernel": {
-      "sha": "b4f7ed2480df20d04cd0e47b5a18518bfd712b6e",
       "dirty": false
     }
   }

 {
+  "name": "ggml-quantization",
+  "id": "_ggml_quantization_metal_09da59d",
   "version": 1,
   "license": "MIT",
   "python-depends": [],
   "digest": {
     "algorithm": "sha256",
     "files": {
+      "__init__.py": "68gDnPDp55W5ZSVOVBPJqML9O/YHq9mzcdCvNs4rnEg=",
+      "_ggml_quantization_metal_09da59d.abi3.so": "ORFW70uOyPVXwey1monXFqKJ5B4cOFPG8FZjMKr+GWo=",
+      "_ops.py": "KPScFwQZRvXLLvmzyYkY4+0hIJB+X5ZJ9opUxecMnZs="
     }
   },
   "provenance": {
       "dirty": false
     },
     "kernel": {
+      "sha": "09da59d51d3f73baa39a941eb07bf6c386223f82",
       "dirty": false
     }
   }