Kernels:

kernels-community
/

mlx-quantization-metal-kernels

Trusted publisher

Kernel card Files Files and versions

xet

Community

kernels-bot commited on 20 days ago

Commit

b41c07c

verified ·

1 Parent(s): c88dadb

Uploaded using `kernel-builder`.

Browse files

Files changed (12) hide show

.gitattributes +3 -0
build/torch210-metal-aarch64-darwin/{_mlx_quantization_metal_kernels_metal_86f75d9.abi3.so → _mlx_quantization_metal_kernels_metal_38ff04e.abi3.so} +2 -2
build/torch210-metal-aarch64-darwin/_ops.py +3 -3
build/torch210-metal-aarch64-darwin/metadata.json +1 -1
build/torch211-metal-aarch64-darwin/{_mlx_quantization_metal_kernels_metal_86f75d9.abi3.so → _mlx_quantization_metal_kernels_metal_38ff04e.abi3.so} +2 -2
build/torch211-metal-aarch64-darwin/_ops.py +3 -3
build/torch211-metal-aarch64-darwin/metadata.json +1 -1
build/torch212-metal-aarch64-darwin/__init__.py +162 -0
build/torch212-metal-aarch64-darwin/_mlx_quantization_metal_kernels_metal_38ff04e.abi3.so +3 -0
build/torch212-metal-aarch64-darwin/_ops.py +9 -0
build/torch212-metal-aarch64-darwin/metadata.json +10 -0
build/torch212-metal-aarch64-darwin/mlx_quantization_metal_kernels/__init__.py +26 -0

.gitattributes CHANGED Viewed

@@ -40,3 +40,6 @@ build/torch210-metal-aarch64-darwin/_mlx_quantization_metal_kernels_metal_0aaded
 build/torch29-metal-aarch64-darwin/_mlx_quantization_metal_kernels_metal_0aaded7.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch210-metal-aarch64-darwin/_mlx_quantization_metal_kernels_metal_86f75d9.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch211-metal-aarch64-darwin/_mlx_quantization_metal_kernels_metal_86f75d9.abi3.so filter=lfs diff=lfs merge=lfs -text

 build/torch29-metal-aarch64-darwin/_mlx_quantization_metal_kernels_metal_0aaded7.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch210-metal-aarch64-darwin/_mlx_quantization_metal_kernels_metal_86f75d9.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch211-metal-aarch64-darwin/_mlx_quantization_metal_kernels_metal_86f75d9.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-metal-aarch64-darwin/_mlx_quantization_metal_kernels_metal_38ff04e.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-metal-aarch64-darwin/_mlx_quantization_metal_kernels_metal_38ff04e.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch212-metal-aarch64-darwin/_mlx_quantization_metal_kernels_metal_38ff04e.abi3.so filter=lfs diff=lfs merge=lfs -text

build/torch210-metal-aarch64-darwin/{_mlx_quantization_metal_kernels_metal_86f75d9.abi3.so → _mlx_quantization_metal_kernels_metal_38ff04e.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cf762ec1255ee6555cb163b30c6bfa6abef2425c04982f908aba65899b4d2f5f
-size 40010744

 version https://git-lfs.github.com/spec/v1
+oid sha256:cdf40b4b1eda395f3b3078db1931397f49c92e35d29017125e6906988c0e635c
+size 40010696

build/torch210-metal-aarch64-darwin/_ops.py CHANGED Viewed

@@ -1,9 +1,9 @@
 import torch
-from . import _mlx_quantization_metal_kernels_metal_86f75d9
-ops = torch.ops._mlx_quantization_metal_kernels_metal_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_mlx_quantization_metal_kernels_metal_86f75d9::{op_name}"

 import torch
+from . import _mlx_quantization_metal_kernels_metal_38ff04e
+ops = torch.ops._mlx_quantization_metal_kernels_metal_38ff04e
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_mlx_quantization_metal_kernels_metal_38ff04e::{op_name}"

build/torch210-metal-aarch64-darwin/metadata.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "mlx-quantization-metal-kernels",
-  "id": "_mlx_quantization_metal_kernels_metal_86f75d9",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

 {
   "name": "mlx-quantization-metal-kernels",
+  "id": "_mlx_quantization_metal_kernels_metal_38ff04e",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

build/torch211-metal-aarch64-darwin/{_mlx_quantization_metal_kernels_metal_86f75d9.abi3.so → _mlx_quantization_metal_kernels_metal_38ff04e.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c0bab012de037b761f8ecac05f87c017c1aa32bf7a09187164c4193e77237222
-size 40010744

 version https://git-lfs.github.com/spec/v1
+oid sha256:b4b007a9dd75308d7337b104b2eb53a98590c2274c112478003714f6b506d963
+size 40010696

build/torch211-metal-aarch64-darwin/_ops.py CHANGED Viewed

@@ -1,9 +1,9 @@
 import torch
-from . import _mlx_quantization_metal_kernels_metal_86f75d9
-ops = torch.ops._mlx_quantization_metal_kernels_metal_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_mlx_quantization_metal_kernels_metal_86f75d9::{op_name}"

 import torch
+from . import _mlx_quantization_metal_kernels_metal_38ff04e
+ops = torch.ops._mlx_quantization_metal_kernels_metal_38ff04e
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_mlx_quantization_metal_kernels_metal_38ff04e::{op_name}"

build/torch211-metal-aarch64-darwin/metadata.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "mlx-quantization-metal-kernels",
-  "id": "_mlx_quantization_metal_kernels_metal_86f75d9",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

 {
   "name": "mlx-quantization-metal-kernels",
+  "id": "_mlx_quantization_metal_kernels_metal_38ff04e",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

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

	@@ -0,0 +1,162 @@

+from typing import Optional
+import torch
+from ._ops import ops
+# =============================================================================
+# FP-quantized (MXFP4) operations
+# =============================================================================
+def mxfp4_qmm_n(
+    x: torch.Tensor,
+    w: torch.Tensor,
+    scales: torch.Tensor,
+    output_features: int,
+) -> torch.Tensor:
+    """Matrix-matrix multiply with MXFP4 quantized non-transposed weight.
+    Computes y = x @ dequantize(w, scales).
+    x: [..., M, K], w: [K_packed, N_packed] (uint32), y: [..., M, output_features]
+    """
+    return ops.mxfp4_qmm_n(x, w, scales, output_features)
+def mxfp4_qmv(
+    x: torch.Tensor,
+    w: torch.Tensor,
+    scales: torch.Tensor,
+    output_features: int,
+) -> torch.Tensor:
+    """Matrix-vector multiply with MXFP4 quantized weight.
+    Computes y = dequantize(w, scales) @ x.
+    x: [..., K], w: [N, K_packed] (uint32), y: [..., output_features]
+    """
+    return ops.mxfp4_qmv(x, w, scales, output_features)
+# =============================================================================
+# Affine quantized operations (scales + biases)
+# =============================================================================
+def affine_qmv(
+    x: torch.Tensor,
+    w: torch.Tensor,
+    scales: torch.Tensor,
+    biases: torch.Tensor,
+    output_features: int,
+    group_size: int = 128,
+    bits: int = 4,
+) -> torch.Tensor:
+    """Matrix-vector multiply with affine quantized weight.
+    x: [..., K], w: [N, K_packed], y: [..., output_features]
+    """
+    return ops.affine_qmv(x, w, scales, biases, group_size, bits, output_features)
+def affine_qmm_t(
+    x: torch.Tensor,
+    w: torch.Tensor,
+    scales: torch.Tensor,
+    biases: torch.Tensor,
+    group_size: int = 128,
+    bits: int = 4,
+) -> torch.Tensor:
+    """Matrix-matrix multiply with affine quantized transposed weight.
+    Computes y = x @ dequantize(w, scales, biases).T
+    x: [..., M, K], w: [N, K_packed], y: [..., M, N]
+    N is inferred from w.size(0).
+    """
+    return ops.affine_qmm_t(x, w, scales, biases, group_size, bits)
+def affine_qmm_n(
+    x: torch.Tensor,
+    w: torch.Tensor,
+    scales: torch.Tensor,
+    biases: torch.Tensor,
+    output_features: int,
+    group_size: int = 128,
+    bits: int = 4,
+) -> torch.Tensor:
+    """Matrix-matrix multiply with affine quantized non-transposed weight.
+    Computes y = x @ dequantize(w, scales, biases)
+    x: [..., M, K], w: [K_packed, N_packed], y: [..., M, output_features]
+    """
+    return ops.affine_qmm_n(x, w, scales, biases, group_size, bits, output_features)
+# =============================================================================
+# Affine quantized NAX operations (MetalPerformancePrimitives accelerated)
+# =============================================================================
+def affine_qmm_t_nax(
+    x: torch.Tensor,
+    w: torch.Tensor,
+    scales: torch.Tensor,
+    biases: torch.Tensor,
+    group_size: int = 128,
+    bits: int = 4,
+) -> torch.Tensor:
+    """NAX-accelerated matrix-matrix multiply with transposed quantized weight.
+    x: [..., M, K], w: [N, K_packed], y: [..., M, N]
+    """
+    return ops.affine_qmm_t_nax(x, w, scales, biases, group_size, bits)
+def affine_qmm_n_nax(
+    x: torch.Tensor,
+    w: torch.Tensor,
+    scales: torch.Tensor,
+    biases: torch.Tensor,
+    output_features: int,
+    group_size: int = 128,
+    bits: int = 4,
+) -> torch.Tensor:
+    """NAX-accelerated matrix-matrix multiply with non-transposed quantized weight.
+    x: [..., M, K], w: [K_packed, N_packed], y: [..., M, output_features]
+    """
+    return ops.affine_qmm_n_nax(x, w, scales, biases, group_size, bits, output_features)
+def affine_gather_qmm_rhs_nax(
+    x: torch.Tensor,
+    w: torch.Tensor,
+    scales: torch.Tensor,
+    biases: torch.Tensor,
+    indices: torch.Tensor,
+    output_features: int,
+    group_size: int = 128,
+    bits: int = 4,
+    transpose: bool = True,
+) -> torch.Tensor:
+    """NAX-accelerated gather + matrix-matrix multiply.
+    Gathers weight rows using indices, then computes matmul.
+    x: [M, K], w: [num_experts, ...], indices: [M], y: [M, output_features]
+    """
+    return ops.affine_gather_qmm_rhs_nax(
+        x, w, scales, biases, indices, group_size, bits, output_features, transpose
+    )
+__all__ = [
+    "mxfp4_qmm_n",
+    "mxfp4_qmv",
+    "affine_qmv",
+    "affine_qmm_t",
+    "affine_qmm_n",
+    "affine_qmm_t_nax",
+    "affine_qmm_n_nax",
+    "affine_gather_qmm_rhs_nax",
+]

build/torch212-metal-aarch64-darwin/_mlx_quantization_metal_kernels_metal_38ff04e.abi3.so ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:2917440f5876e5f88744089b2582892684a1714c48610ac4402bd9c2f02e8f57
+size 40010728

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

	@@ -0,0 +1,9 @@

+import torch
+from . import _mlx_quantization_metal_kernels_metal_38ff04e
+ops = torch.ops._mlx_quantization_metal_kernels_metal_38ff04e
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_mlx_quantization_metal_kernels_metal_38ff04e::{op_name}"

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

	@@ -0,0 +1,10 @@

+{
+  "name": "mlx-quantization-metal-kernels",
+  "id": "_mlx_quantization_metal_kernels_metal_38ff04e",
+  "version": 1,
+  "license": "MIT",
+  "python-depends": [],
+  "backend": {
+    "type": "metal"
+  }
+}

build/torch212-metal-aarch64-darwin/mlx_quantization_metal_kernels/__init__.py ADDED Viewed

	@@ -0,0 +1,26 @@

+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))