Build (aarch64)

build/torch26-cxx11-cu126-aarch64-linux/quantization/__init__.py

@@ -19,6 +19,11 @@ from .scalar_type import (
 )
 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",
@@ -32,7 +37,11 @@ __all__ = [
     "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/torch26-cxx11-cu126-aarch64-linux/quantization/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _quantization_9035540
-ops = torch.ops._quantization_9035540
+from . import _quantization_82ffd1f
+ops = torch.ops._quantization_82ffd1f
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_quantization_9035540::{op_name}"
+    return f"_quantization_82ffd1f::{op_name}"

build/torch26-cxx11-cu126-aarch64-linux/quantization/{_quantization_9035540.abi3.so → _quantization_82ffd1f.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:aee128710f3a8587386120a226a6caddd5e77cd7a0296a1f7fad51b4028550b1
-size 159934120
+oid sha256:163383785e3ca9a472f18c802591218f18ef3c9cde4bb83fa623575a8adfd085
+size 159999656

build/torch26-cxx11-cu126-aarch64-linux/quantization/compressed_tensors.py

@@ -1,8 +1,10 @@
-from typing import Optional, Tuple
+from typing import Optional, Union
 
 import torch
 
 from ._ops import ops
+from .platforms import current_platform
+
 
 # fp8
 def scaled_fp8_quant(

build/torch26-cxx11-cu126-aarch64-linux/quantization/platforms.py

@@ -27,6 +27,29 @@ class DeviceCapability(NamedTuple):
 class Platform(ABC):
     simple_compile_backend: str = "inductor"
 
+    @classmethod
+    def fp8_dtype(cls) -> torch.dtype:
+        """
+        Returns the preferred FP8 type on the current platform.
+
+        See the documentation for is_fp8_fnuz for details.
+        """
+        return torch.float8_e4m3fn
+
+    @classmethod
+    def is_fp8_fnuz(cls) -> bool:
+        """
+        Returns whether the preferred FP8 type is FNUZ on the current platform.
+
+        There are two representations of FP8, OCP FP8 and FNUZ FP8.
+        The OCP specification can be found at https://tinyurl.com/b7jvwpft.
+        The FNUZ specification can be found at https://tinyurl.com/5n6hwwu5.
+
+        AMD's MI300 and MI325 have native hardware support for FNUZ. All other
+        hardware has converged on the OCP FP8 standard.
+        """
+        return False
+
     @classmethod
     @abstractmethod
     def get_device_name(cls, device_id: int = 0) -> str: ...
@@ -51,6 +74,18 @@ class CudaPlatform(Platform):
 
 
 class RocmPlatform(Platform):
+    @classmethod
+    def fp8_dtype(cls) -> torch.dtype:
+        if cls.is_fp8_fnuz():
+            return torch.float8_e4m3fnuz
+        else:
+            return torch.float8_e4m3fn
+
+    @classmethod
+    def is_fp8_fnuz(cls) -> bool:
+        # only device 0 is checked, this assumes MI300 platforms are homogeneous
+        return "gfx94" in torch.cuda.get_device_properties(0).gcnArchName
+
     @classmethod
     @lru_cache(maxsize=8)
     def get_device_capability(cls, device_id: int = 0) -> DeviceCapability:

build/torch26-cxx98-cu126-aarch64-linux/quantization/__init__.py

@@ -19,6 +19,11 @@ from .scalar_type import (
 )
 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",
@@ -32,7 +37,11 @@ __all__ = [
     "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/torch26-cxx98-cu126-aarch64-linux/quantization/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _quantization_9035540
-ops = torch.ops._quantization_9035540
+from . import _quantization_82ffd1f
+ops = torch.ops._quantization_82ffd1f
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_quantization_9035540::{op_name}"
+    return f"_quantization_82ffd1f::{op_name}"

build/torch26-cxx98-cu126-aarch64-linux/quantization/{_quantization_9035540.abi3.so → _quantization_82ffd1f.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3685a434362226370f1956f59790a58d2f4c8999f9f35acafd25ca9d73bfc5ae
+oid sha256:f2bf0942eeeb2b821331211fc74ce7c37fccad95fc1ac6aa8bbc322a6f8ac249
 size 159991696

build/torch26-cxx98-cu126-aarch64-linux/quantization/compressed_tensors.py

@@ -1,8 +1,10 @@
-from typing import Optional, Tuple
+from typing import Optional, Union
 
 import torch
 
 from ._ops import ops
+from .platforms import current_platform
+
 
 # fp8
 def scaled_fp8_quant(

build/torch26-cxx98-cu126-aarch64-linux/quantization/platforms.py

@@ -27,6 +27,29 @@ class DeviceCapability(NamedTuple):
 class Platform(ABC):
     simple_compile_backend: str = "inductor"
 
+    @classmethod
+    def fp8_dtype(cls) -> torch.dtype:
+        """
+        Returns the preferred FP8 type on the current platform.
+
+        See the documentation for is_fp8_fnuz for details.
+        """
+        return torch.float8_e4m3fn
+
+    @classmethod
+    def is_fp8_fnuz(cls) -> bool:
+        """
+        Returns whether the preferred FP8 type is FNUZ on the current platform.
+
+        There are two representations of FP8, OCP FP8 and FNUZ FP8.
+        The OCP specification can be found at https://tinyurl.com/b7jvwpft.
+        The FNUZ specification can be found at https://tinyurl.com/5n6hwwu5.
+
+        AMD's MI300 and MI325 have native hardware support for FNUZ. All other
+        hardware has converged on the OCP FP8 standard.
+        """
+        return False
+
     @classmethod
     @abstractmethod
     def get_device_name(cls, device_id: int = 0) -> str: ...
@@ -51,6 +74,18 @@ class CudaPlatform(Platform):
 
 
 class RocmPlatform(Platform):
+    @classmethod
+    def fp8_dtype(cls) -> torch.dtype:
+        if cls.is_fp8_fnuz():
+            return torch.float8_e4m3fnuz
+        else:
+            return torch.float8_e4m3fn
+
+    @classmethod
+    def is_fp8_fnuz(cls) -> bool:
+        # only device 0 is checked, this assumes MI300 platforms are homogeneous
+        return "gfx94" in torch.cuda.get_device_properties(0).gcnArchName
+
     @classmethod
     @lru_cache(maxsize=8)
     def get_device_capability(cls, device_id: int = 0) -> DeviceCapability:

build/torch27-cxx11-cu126-aarch64-linux/quantization/__init__.py

@@ -19,6 +19,11 @@ from .scalar_type import (
 )
 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",
@@ -32,7 +37,11 @@ __all__ = [
     "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/torch27-cxx11-cu126-aarch64-linux/quantization/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _quantization_9035540
-ops = torch.ops._quantization_9035540
+from . import _quantization_82ffd1f
+ops = torch.ops._quantization_82ffd1f
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_quantization_9035540::{op_name}"
+    return f"_quantization_82ffd1f::{op_name}"

build/torch27-cxx11-cu126-aarch64-linux/quantization/{_quantization_9035540.abi3.so → _quantization_82ffd1f.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:219fc94b48e46777769dd2cd61785791b4fd00c58824d6de5252defbf48c30e5
-size 159999608
+oid sha256:ef0e68ff25982049ce0b6af570f6546c8f62a49e373397d352f89376c1805de4
+size 159934080

build/torch27-cxx11-cu126-aarch64-linux/quantization/compressed_tensors.py

@@ -1,8 +1,10 @@
-from typing import Optional, Tuple
+from typing import Optional, Union
 
 import torch
 
 from ._ops import ops
+from .platforms import current_platform
+
 
 # fp8
 def scaled_fp8_quant(

build/torch27-cxx11-cu126-aarch64-linux/quantization/platforms.py

@@ -27,6 +27,29 @@ class DeviceCapability(NamedTuple):
 class Platform(ABC):
     simple_compile_backend: str = "inductor"
 
+    @classmethod
+    def fp8_dtype(cls) -> torch.dtype:
+        """
+        Returns the preferred FP8 type on the current platform.
+
+        See the documentation for is_fp8_fnuz for details.
+        """
+        return torch.float8_e4m3fn
+
+    @classmethod
+    def is_fp8_fnuz(cls) -> bool:
+        """
+        Returns whether the preferred FP8 type is FNUZ on the current platform.
+
+        There are two representations of FP8, OCP FP8 and FNUZ FP8.
+        The OCP specification can be found at https://tinyurl.com/b7jvwpft.
+        The FNUZ specification can be found at https://tinyurl.com/5n6hwwu5.
+
+        AMD's MI300 and MI325 have native hardware support for FNUZ. All other
+        hardware has converged on the OCP FP8 standard.
+        """
+        return False
+
     @classmethod
     @abstractmethod
     def get_device_name(cls, device_id: int = 0) -> str: ...
@@ -51,6 +74,18 @@ class CudaPlatform(Platform):
 
 
 class RocmPlatform(Platform):
+    @classmethod
+    def fp8_dtype(cls) -> torch.dtype:
+        if cls.is_fp8_fnuz():
+            return torch.float8_e4m3fnuz
+        else:
+            return torch.float8_e4m3fn
+
+    @classmethod
+    def is_fp8_fnuz(cls) -> bool:
+        # only device 0 is checked, this assumes MI300 platforms are homogeneous
+        return "gfx94" in torch.cuda.get_device_properties(0).gcnArchName
+
     @classmethod
     @lru_cache(maxsize=8)
     def get_device_capability(cls, device_id: int = 0) -> DeviceCapability:

build/torch27-cxx11-cu128-aarch64-linux/quantization/__init__.py

@@ -19,6 +19,11 @@ from .scalar_type import (
 )
 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",
@@ -32,7 +37,11 @@ __all__ = [
     "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/torch27-cxx11-cu128-aarch64-linux/quantization/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _quantization_9035540
-ops = torch.ops._quantization_9035540
+from . import _quantization_82ffd1f
+ops = torch.ops._quantization_82ffd1f
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_quantization_9035540::{op_name}"
+    return f"_quantization_82ffd1f::{op_name}"

build/torch27-cxx11-cu128-aarch64-linux/quantization/{_quantization_9035540.abi3.so → _quantization_82ffd1f.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4d670f7d449a8d177ce46784fb4617dcb0edc30f8d8a62305ed1213310256167
-size 296561248
+oid sha256:7c8d85c6222df8ff6de82adbad94502fdc5c1910dbaa367034c8975c4f85244a
+size 296561256

build/torch27-cxx11-cu128-aarch64-linux/quantization/compressed_tensors.py

@@ -1,8 +1,10 @@
-from typing import Optional, Tuple
+from typing import Optional, Union
 
 import torch
 
 from ._ops import ops
+from .platforms import current_platform
+
 
 # fp8
 def scaled_fp8_quant(

build/torch27-cxx11-cu128-aarch64-linux/quantization/platforms.py

@@ -27,6 +27,29 @@ class DeviceCapability(NamedTuple):
 class Platform(ABC):
     simple_compile_backend: str = "inductor"
 
+    @classmethod
+    def fp8_dtype(cls) -> torch.dtype:
+        """
+        Returns the preferred FP8 type on the current platform.
+
+        See the documentation for is_fp8_fnuz for details.
+        """
+        return torch.float8_e4m3fn
+
+    @classmethod
+    def is_fp8_fnuz(cls) -> bool:
+        """
+        Returns whether the preferred FP8 type is FNUZ on the current platform.
+
+        There are two representations of FP8, OCP FP8 and FNUZ FP8.
+        The OCP specification can be found at https://tinyurl.com/b7jvwpft.
+        The FNUZ specification can be found at https://tinyurl.com/5n6hwwu5.
+
+        AMD's MI300 and MI325 have native hardware support for FNUZ. All other
+        hardware has converged on the OCP FP8 standard.
+        """
+        return False
+
     @classmethod
     @abstractmethod
     def get_device_name(cls, device_id: int = 0) -> str: ...
@@ -51,6 +74,18 @@ class CudaPlatform(Platform):
 
 
 class RocmPlatform(Platform):
+    @classmethod
+    def fp8_dtype(cls) -> torch.dtype:
+        if cls.is_fp8_fnuz():
+            return torch.float8_e4m3fnuz
+        else:
+            return torch.float8_e4m3fn
+
+    @classmethod
+    def is_fp8_fnuz(cls) -> bool:
+        # only device 0 is checked, this assumes MI300 platforms are homogeneous
+        return "gfx94" in torch.cuda.get_device_properties(0).gcnArchName
+
     @classmethod
     @lru_cache(maxsize=8)
     def get_device_capability(cls, device_id: int = 0) -> DeviceCapability: