fbgemm-gpu-genai-nightly 2025.9.20__cp313-cp313-manylinux_2_28_x86_64.whl → 2025.11.4__cp313-cp313-manylinux_2_28_x86_64.whl

fbgemm_gpu/__init__.py

@@ -5,6 +5,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+import json
 import logging
 import os
 import re
@@ -26,6 +27,19 @@ _fbgemm_torch_compat_table = {
 }
 
 
+def _load_target_info(target: str) -> dict[str, str]:
+    try:
+        filepath = os.path.join(
+            os.path.dirname(__file__), "docs", f"target.{target}.json.py"
+        )
+        with open(filepath, "r") as file:
+            data = json.load(file)
+    except Exception:
+        data = {}
+
+    return data
+
+
 def _load_library(filename: str, version: str, no_throw: bool = False) -> None:
     """Load a shared library from the given filename."""
 
@@ -98,13 +112,16 @@ open_source: bool = True
 # Trigger the manual addition of docstrings to pybind11-generated operators
 import fbgemm_gpu.docs  # noqa: F401, E402
 
+
+__targets_infos__ = {
+    target: _load_target_info(target) for target in ["default", "genai", "hstu"]
+}
+__targets_infos__ = {k: v for (k, v) in __targets_infos__.items() if v}
+
 try:
-    # Export the version string from the version file auto-generated by setup.py
-    from fbgemm_gpu.docs.version import (  # noqa: F401, E402
-        __target__,
-        __variant__,
-        __version__,
-    )
+    __target__, __info__ = next(iter(__targets_infos__.items()))
+    __variant__ = __info__["variant"]
+    __version__ = __info__["version"]
 except Exception:
     __variant__: str = "INTERNAL"
     __version__: str = "INTERNAL"
@@ -145,18 +162,19 @@ libraries_to_load = {
     "genai": fbgemm_genai_libraries,
 }
 
-for library in libraries_to_load.get(__target__, []):
-    # NOTE: In all cases, we want to throw an error if we cannot load the
-    # library.  However, this appears to break the OSS documentation build,
-    # where the Python documentation doesn't show up in the generated docs.
-    #
-    # To work around this problem, we introduce a fake build variant called
-    # `docs` and we only throw a library load error when the variant is not
-    # `docs`.  For more information, see:
-    #
-    #   https://github.com/pytorch/FBGEMM/pull/3477
-    #   https://github.com/pytorch/FBGEMM/pull/3717
-    _load_library(f"{library}.so", __version__, __variant__ == "docs")
+for target, info in __targets_infos__.items():
+    for library in libraries_to_load.get(target, []):
+        # NOTE: In all cases, we want to throw an error if we cannot load the
+        # library.  However, this appears to break the OSS documentation build,
+        # where the Python documentation doesn't show up in the generated docs.
+        #
+        # To work around this problem, we introduce a fake build variant called
+        # `docs` and we only throw a library load error when the variant is not
+        # `docs`.  For more information, see:
+        #
+        #   https://github.com/pytorch/FBGEMM/pull/3477
+        #   https://github.com/pytorch/FBGEMM/pull/3717
+        _load_library(f"{library}.so", info["version"], info["variant"] == "docs")
 
 try:
     # Trigger meta operator registrations

fbgemm_gpu/batched_unary_embeddings_ops.py

@@ -9,7 +9,6 @@
 
 
 from math import sqrt
-from typing import List
 
 import torch
 
@@ -22,7 +21,7 @@ except Exception:
     load_torch_module("//deeplearning/fbgemm/fbgemm_gpu:sparse_ops")
 
 
-def wrap_weight_to_parameter(weights: List[torch.Tensor]) -> List[torch.Tensor]:
+def wrap_weight_to_parameter(weights: list[torch.Tensor]) -> list[torch.Tensor]:
     for i, v in enumerate(weights):
         if not isinstance(v, torch.nn.Parameter):
             weights[i] = torch.nn.Parameter(v)
@@ -31,7 +30,7 @@ def wrap_weight_to_parameter(weights: List[torch.Tensor]) -> List[torch.Tensor]:
 
 class BatchedUnaryEmbeddingBag(torch.nn.Module):
     # pyre-fixme[3]: Return type must be annotated.
-    def __init__(self, num_tasks: int, hash_sizes: List[int], long_index: bool = False):
+    def __init__(self, num_tasks: int, hash_sizes: list[int], long_index: bool = False):
         super().__init__()
         self.num_tasks = num_tasks
         self.hash_sizes = hash_sizes

fbgemm_gpu/config/feature_list.py

@@ -11,7 +11,7 @@ from enum import auto, Enum
 import torch
 
 try:
-    torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu:config_cpp")
+    torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu:config_cpp_torch_op")
 except Exception:
     import fbgemm_gpu  # noqa F401
 

fbgemm_gpu/docs/target.genai.json.py

@@ -0,0 +1,6 @@
+
+{
+    "version": "2025.11.4",
+    "target": "genai",
+    "variant": "cuda"
+}

fbgemm_gpu/enums.py

@@ -8,14 +8,13 @@
 # pyre-strict
 
 import enum
-import typing
-from typing import Any, Callable, List, Tuple
+from typing import Any, Callable
 
 
 # Create enums in given namespace with information from query_op
 def create_enums(
-    namespace: typing.Dict[str, Any],
-    query_op: Callable[[], List[Tuple[str, List[Tuple[str, int]]]]],
+    namespace: dict[str, Any],
+    query_op: Callable[[], list[tuple[str, list[tuple[str, int]]]]],
 ) -> None:
     for enum_name, items in query_op():
         # Create matching python enumeration

fbgemm_gpu/experimental/example/__init__.py

@@ -15,10 +15,6 @@ try:
     # pyre-ignore[21]
     # @manual=//deeplearning/fbgemm/fbgemm_gpu:test_utils
     from fbgemm_gpu import open_source
-
-    # pyre-ignore[21]
-    # @manual=//deeplearning/fbgemm/fbgemm_gpu:test_utils
-    from fbgemm_gpu.docs.version import __version__  # noqa: F401
 except Exception:
     open_source: bool = False
 

fbgemm_gpu/experimental/gemm/triton_gemm/__init__.py

@@ -11,9 +11,5 @@ try:
     # pyre-ignore[21]
     # @manual=//deeplearning/fbgemm/fbgemm_gpu:test_utils
     from fbgemm_gpu import open_source
-
-    # pyre-ignore[21]
-    # @manual=//deeplearning/fbgemm/fbgemm_gpu:test_utils
-    from fbgemm_gpu.docs.version import __version__  # noqa: F401
 except Exception:
     open_source: bool = False