Support recent latest Torch versions

build.toml

@@ -1,6 +1,6 @@
 [general]
 name = "quantization"
-universal = false
+backends = ["cuda"]
 
 [torch]
 include = ["."]
@@ -22,6 +22,7 @@ cuda-capabilities = [
     "10.1",
     "12.0",
 ]
+cuda-flags = ["-static-global-template-stub=false"]
 depends = ["torch"]
 include = ["."]
 src = [
@@ -95,6 +96,7 @@ depends = ["torch"]
 include = ["."]
 src = [
     "compressed_tensors/int8_quant_kernels.cu",
+    "cub_helpers.h",
     "dispatch_utils.h",
     "vectorization_utils.cuh",
 ]
@@ -119,6 +121,7 @@ include = ["."]
 src = [
     "fp8/common.cu",
     "fp8/common.cuh",
+    "cub_helpers.h",
     "dispatch_utils.h",
     "utils.cuh",
     "vectorization.cuh",
@@ -229,6 +232,7 @@ cuda-capabilities = [
     "10.1",
     "12.0",
 ]
+cuda-flags = ["-static-global-template-stub=false"]
 depends = ["torch"]
 include = ["."]
 src = [

build/torch210-cxx11-cu126-x86_64-linux/__init__.py

@@ -0,0 +1,48 @@
+from .compressed_tensors import scaled_fp8_quant, scaled_int8_quant
+from .cutlass import (
+    cutlass_scaled_mm_supports_block_fp8,
+    cutlass_scaled_mm_supports_fp8,
+    cutlass_scaled_mm,
+    cutlass_scaled_mm_azp,
+)
+from .marlin import (
+    awq_marlin_repack,
+    gptq_marlin_gemm,
+    gptq_marlin_repack,
+    gptq_marlin_24_gemm,
+    marlin_qqq_gemm,
+    marlin_gemm,
+)
+from .scalar_type import (
+    ScalarType,
+    scalar_types,
+)
+from ._ops import ops
+
+from .utils import marlin_utils
+from .utils import marlin_utils_fp4
+from .utils import marlin_utils_fp8
+from .utils import quant_utils
+
+
+__all__ = [
+    "ScalarType",
+    "awq_marlin_repack",
+    "cutlass_scaled_mm",
+    "cutlass_scaled_mm_azp",
+    "cutlass_scaled_mm_supports_block_fp8",
+    "cutlass_scaled_mm_supports_fp8",
+    "gptq_marlin_24_gemm",
+    "gptq_marlin_gemm",
+    "gptq_marlin_repack",
+    "marlin_gemm",
+    "marlin_qqq_gemm",
+    "marlin_utils",
+    "marlin_utils_fp4",
+    "marlin_utils_fp8",
+    "ops",
+    "quant_utils",
+    "scalar_types",
+    "scaled_fp8_quant",
+    "scaled_int8_quant",
+]

build/{torch28-cxx11-cu126-x86_64-linux/quantization → torch210-cxx11-cu126-x86_64-linux}/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _quantization_eabe7c2
-ops = torch.ops._quantization_eabe7c2
+from . import _quantization_e384bb2
+ops = torch.ops._quantization_e384bb2
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_quantization_eabe7c2::{op_name}"
+    return f"_quantization_e384bb2::{op_name}"

build/{torch28-cxx11-cu126-x86_64-linux/quantization/_quantization_eabe7c2.abi3.so → torch210-cxx11-cu126-x86_64-linux/_quantization_e384bb2.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:72083bbb272942332e2586f6b0068dd7f4347981193db1e682e90f54b0b01c87
-size 160171952
+oid sha256:85e946af129d0f8ad6cd758d0270a5ff336a6cba021626d72b81e40529b55fc9
+size 196115272

build/torch210-cxx11-cu126-x86_64-linux/metadata.json

@@ -0,0 +1,3 @@
+{
+  "python-depends": []
+}

build/torch210-cxx11-cu126-x86_64-linux/quantization/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import sys
+
+import importlib
+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")))

build/torch210-cxx11-cu128-x86_64-linux/__init__.py

@@ -0,0 +1,48 @@
+from .compressed_tensors import scaled_fp8_quant, scaled_int8_quant
+from .cutlass import (
+    cutlass_scaled_mm_supports_block_fp8,
+    cutlass_scaled_mm_supports_fp8,
+    cutlass_scaled_mm,
+    cutlass_scaled_mm_azp,
+)
+from .marlin import (
+    awq_marlin_repack,
+    gptq_marlin_gemm,
+    gptq_marlin_repack,
+    gptq_marlin_24_gemm,
+    marlin_qqq_gemm,
+    marlin_gemm,
+)
+from .scalar_type import (
+    ScalarType,
+    scalar_types,
+)
+from ._ops import ops
+
+from .utils import marlin_utils
+from .utils import marlin_utils_fp4
+from .utils import marlin_utils_fp8
+from .utils import quant_utils
+
+
+__all__ = [
+    "ScalarType",
+    "awq_marlin_repack",
+    "cutlass_scaled_mm",
+    "cutlass_scaled_mm_azp",
+    "cutlass_scaled_mm_supports_block_fp8",
+    "cutlass_scaled_mm_supports_fp8",
+    "gptq_marlin_24_gemm",
+    "gptq_marlin_gemm",
+    "gptq_marlin_repack",
+    "marlin_gemm",
+    "marlin_qqq_gemm",
+    "marlin_utils",
+    "marlin_utils_fp4",
+    "marlin_utils_fp8",
+    "ops",
+    "quant_utils",
+    "scalar_types",
+    "scaled_fp8_quant",
+    "scaled_int8_quant",
+]

build/{torch28-cxx11-cu128-x86_64-linux/quantization → torch210-cxx11-cu128-x86_64-linux}/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _quantization_eabe7c2
-ops = torch.ops._quantization_eabe7c2
+from . import _quantization_e384bb2
+ops = torch.ops._quantization_e384bb2
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_quantization_eabe7c2::{op_name}"
+    return f"_quantization_e384bb2::{op_name}"

build/{torch28-cxx11-cu128-x86_64-linux/quantization/_quantization_eabe7c2.abi3.so → torch210-cxx11-cu128-x86_64-linux/_quantization_e384bb2.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dec4454faaae3f515aac5e2ff702d910c660ea7b23d88d1ef643abddd71c21c2
-size 296976072
+oid sha256:1e996222b90650b5074935088ed5cf27816148197c6dd75023d56a0c7b7f67d1
+size 332877416

build/torch210-cxx11-cu128-x86_64-linux/metadata.json

@@ -0,0 +1,3 @@
+{
+  "python-depends": []
+}

build/torch210-cxx11-cu128-x86_64-linux/quantization/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import sys
+
+import importlib
+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")))

build/torch210-cxx11-cu130-x86_64-linux/__init__.py

@@ -0,0 +1,48 @@
+from .compressed_tensors import scaled_fp8_quant, scaled_int8_quant
+from .cutlass import (
+    cutlass_scaled_mm_supports_block_fp8,
+    cutlass_scaled_mm_supports_fp8,
+    cutlass_scaled_mm,
+    cutlass_scaled_mm_azp,
+)
+from .marlin import (
+    awq_marlin_repack,
+    gptq_marlin_gemm,
+    gptq_marlin_repack,
+    gptq_marlin_24_gemm,
+    marlin_qqq_gemm,
+    marlin_gemm,
+)
+from .scalar_type import (
+    ScalarType,
+    scalar_types,
+)
+from ._ops import ops
+
+from .utils import marlin_utils
+from .utils import marlin_utils_fp4
+from .utils import marlin_utils_fp8
+from .utils import quant_utils
+
+
+__all__ = [
+    "ScalarType",
+    "awq_marlin_repack",
+    "cutlass_scaled_mm",
+    "cutlass_scaled_mm_azp",
+    "cutlass_scaled_mm_supports_block_fp8",
+    "cutlass_scaled_mm_supports_fp8",
+    "gptq_marlin_24_gemm",
+    "gptq_marlin_gemm",
+    "gptq_marlin_repack",
+    "marlin_gemm",
+    "marlin_qqq_gemm",
+    "marlin_utils",
+    "marlin_utils_fp4",
+    "marlin_utils_fp8",
+    "ops",
+    "quant_utils",
+    "scalar_types",
+    "scaled_fp8_quant",
+    "scaled_int8_quant",
+]

build/{torch28-cxx11-cu129-x86_64-linux/quantization → torch210-cxx11-cu130-x86_64-linux}/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _quantization_eabe7c2
-ops = torch.ops._quantization_eabe7c2
+from . import _quantization_e384bb2
+ops = torch.ops._quantization_e384bb2
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_quantization_eabe7c2::{op_name}"
+    return f"_quantization_e384bb2::{op_name}"

build/{torch28-cxx11-cu129-x86_64-linux/quantization/_quantization_eabe7c2.abi3.so → torch210-cxx11-cu130-x86_64-linux/_quantization_e384bb2.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d6cf80d6102e5a72eac6a15260a296193fefaa2d878ed5ca7458e5e54707a85a
-size 304924864
+oid sha256:f94359faa54a316eb410e331755a19c96c17e87d74ffae035ca8187d62f58891
+size 289949016

build/torch210-cxx11-cu130-x86_64-linux/metadata.json

@@ -0,0 +1,3 @@
+{
+  "python-depends": []
+}

build/torch210-cxx11-cu130-x86_64-linux/quantization/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import sys
+
+import importlib
+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")))