fbgemm-gpu-hstu-nightly 2025.6.15__cp313-cp313-manylinux_2_28_x86_64.whl → 2025.6.17__cp313-cp313-manylinux_2_28_x86_64.whl

fbgemm_gpu/docs/version.py

@@ -6,6 +6,6 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-__version__: str = "2025.6.15"
+__version__: str = "2025.6.17"
 __target__: str = "hstu"
 __variant__: str = "cuda"

fbgemm_gpu/split_table_batched_embeddings_ops_training.py

@@ -1947,6 +1947,7 @@ class SplitTableBatchedEmbeddingBagsCodegen(nn.Module):
             per_sample_weights,
             batch_size_per_feature_per_rank,
             force_cast_input_types=True,
+            prefetch_pipeline=False,
         )
 
         # Print input stats if enable (for debugging purpose only)
@@ -2478,6 +2479,7 @@ class SplitTableBatchedEmbeddingBagsCodegen(nn.Module):
             per_sample_weights=None,
             batch_size_per_feature_per_rank=batch_size_per_feature_per_rank,
             force_cast_input_types=False,
+            prefetch_pipeline=self.prefetch_pipeline,
         )
 
         with self._recording_to_timer(
@@ -3543,6 +3545,7 @@ class SplitTableBatchedEmbeddingBagsCodegen(nn.Module):
         per_sample_weights: Optional[Tensor] = None,
         batch_size_per_feature_per_rank: Optional[List[List[int]]] = None,
         force_cast_input_types: bool = True,
+        prefetch_pipeline: bool = False,
     ) -> Tuple[Tensor, Tensor, Optional[Tensor], invokers.lookup_args.VBEMetadata]:
         """
         Prepare TBE inputs as follows:
@@ -3613,9 +3616,17 @@ class SplitTableBatchedEmbeddingBagsCodegen(nn.Module):
                 per_sample_weights = per_sample_weights.float()
 
         if self.bounds_check_mode_int != BoundsCheckMode.NONE.value:
+            # Override the bounds check version based on prefetch_pipeline
+            use_bounds_check_v2 = self.bounds_check_version == 2 or prefetch_pipeline
+            bounds_check_version = (
+                2 if use_bounds_check_v2 else self.bounds_check_version
+            )
+
+            vbe = vbe_metadata.B_offsets is not None
+
             # Compute B info and VBE metadata for bounds_check_indices only if
             # VBE and bounds check indices v2 are used
-            if vbe and self.bounds_check_version == 2:
+            if vbe and use_bounds_check_v2:
                 B_offsets = vbe_metadata.B_offsets
                 B_offsets_rank_per_feature = vbe_metadata.B_offsets_rank_per_feature
                 output_offsets_feature_rank = vbe_metadata.output_offsets_feature_rank
@@ -3653,7 +3664,8 @@ class SplitTableBatchedEmbeddingBagsCodegen(nn.Module):
                 b_t_map=b_t_map,
                 info_B_num_bits=self.info_B_num_bits,
                 info_B_mask=self.info_B_mask,
-                bounds_check_version=self.bounds_check_version,
+                bounds_check_version=bounds_check_version,
+                prefetch_pipeline=prefetch_pipeline,
             )
 
         return indices, offsets, per_sample_weights, vbe_metadata

fbgemm_gpu/utils/filestore.py

@@ -155,4 +155,53 @@ class FileStore:
             True if file exists, False otherwise.
         """
         filepath = f"{self.bucket}/{path}"
-        return os.path.isfile(filepath)
+        return os.path.exists(filepath)
+
+    def create_directory(self, path: str) -> "FileStore":
+        """
+        Creates a directory in the file store.
+
+        Args:
+            path (str): The path of the node or symlink to a directory (relative
+            to `self.bucket`) to be created.
+
+        Returns:
+            self.  This allows for method-chaining.
+        """
+        filepath = f"{self.bucket}/{path}"
+        event = f"creating directory {filepath}"
+        logger.info(f"FileStore: {event}")
+
+        try:
+            if not os.path.exists(filepath):
+                os.makedirs(filepath, exist_ok=True)
+        except Exception as e:
+            logger.error(f"FileStore: exception occurred when {event}: {e}")
+            raise e
+
+        return self
+
+    def remove_directory(self, path: str) -> "FileStore":
+        """
+        Removes a directory from the file store.
+
+        Args:
+            path (str): The path of the node or symlink to a directory (relative
+            to `self.bucket`) to be removed.
+
+        Returns:
+            self.  This allows for method-chaining.
+        """
+        filepath = f"{self.bucket}/{path}"
+        event = f"deleting {filepath}"
+        logger.info(f"FileStore: {event}")
+
+        try:
+            if os.path.isdir(filepath):
+                os.rmdir(filepath)
+
+        except Exception as e:
+            logger.error(f"Manifold: exception occurred when {event}: {e}")
+            raise e
+
+        return self

{fbgemm_gpu_hstu_nightly-2025.6.15.dist-info → fbgemm_gpu_hstu_nightly-2025.6.17.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fbgemm_gpu_hstu_nightly
-Version: 2025.6.15
+Version: 2025.6.17
 Home-page: https://github.com/pytorch/fbgemm
 Author: FBGEMM Team
 Author-email: packages@pytorch.org
@@ -40,9 +40,6 @@ PyTorch GPU operator libraries for training and inference.  The library provides
 efficient table batched embedding bag, data layout transformation, and
 quantization supports.
 
-FBGEMM_GPU is currently tested with CUDA 12.4 and 11.8 in CI, and with PyTorch
-packages (2.1+) that are built against those CUDA versions.
-
 See the full [Documentation](https://pytorch.org/FBGEMM) for more information
 on building, installing, and developing with FBGEMM_GPU, as well as the most
 up-to-date support matrix for this library.

{fbgemm_gpu_hstu_nightly-2025.6.15.dist-info → fbgemm_gpu_hstu_nightly-2025.6.17.dist-info}/RECORD

@@ -1,8 +1,8 @@
 fbgemm_gpu/__init__.py,sha256=BrIitwvFsRtKEk1ZBHFUi9j6ZUgoA5K7CvepoBez0u4,3419
-fbgemm_gpu/asmjit.so,sha256=N7cRqpxFdbuDfCa4auIqhOARZL6mtifGbfY0OyaEuh8,501728
+fbgemm_gpu/asmjit.so,sha256=1mgsQhqX1yiUdU9p2w3e7XNhDxhMprHy8qkFKYM01Ww,488288
 fbgemm_gpu/batched_unary_embeddings_ops.py,sha256=u7LfseNeM5gGFQGLAMVO7h2QkFWEOL3ezV5RuhbZn4M,2928
 fbgemm_gpu/enums.py,sha256=GVuzF5cFTLzttkvlH1SdcGrxrppMhDSbQj_Vm_4zmEo,789
-fbgemm_gpu/fbgemm.so,sha256=l2CnLwTqGYFdS-vdhZ56w6EPfMDjeh0vIa1uQycYObU,5408824
+fbgemm_gpu/fbgemm.so,sha256=2giLGFkDpN5f6NtML_Din2J98LCdwJ0kgL_U3sbGoc0,5634864
 fbgemm_gpu/metrics.py,sha256=TsurFLJf0nJvPDN7urWb4LMQlf5RgdWPTTTDO7S4wtI,5663
 fbgemm_gpu/permute_pooled_embedding_modules.py,sha256=kjWuWmQY8e2kMRwIPTzjGjyjV4syKPrphtHdsQTAjWM,5136
 fbgemm_gpu/permute_pooled_embedding_modules_split.py,sha256=cUrEbRIvLFW_3Zmh07QkN4S1Cfvvge6TYO1VXBFCpz8,2752
@@ -17,7 +17,7 @@ fbgemm_gpu/split_embedding_utils.py,sha256=Gb40ZKeATxIKEKI3aVQMgDDBanNpKMc53Z43m
 fbgemm_gpu/split_table_batched_embeddings_ops.py,sha256=_MIp6uHYHLn4GxGdrGsfddfSsZ2Z9mjsYIrih3ncI1I,2339
 fbgemm_gpu/split_table_batched_embeddings_ops_common.py,sha256=qglNRKKuHkrKiTw90ACjZpMzcjHKXKV7ME3a8QHfQt4,8237
 fbgemm_gpu/split_table_batched_embeddings_ops_inference.py,sha256=bUDWa6IR0vGLDThgB3nmD1yfYa8_HD34B0dtLnd7thw,81692
-fbgemm_gpu/split_table_batched_embeddings_ops_training.py,sha256=YCLPSW9CXrRwMN5KEU6x0ESbutdhzKTaNOO8oN5kX7I,163875
+fbgemm_gpu/split_table_batched_embeddings_ops_training.py,sha256=GY3Wm5X2utjIXcjIcHwByzI38hLa9NlShtHg58wIOOU,164383
 fbgemm_gpu/split_table_batched_embeddings_ops_training_common.py,sha256=ktC10-nakOBpcmJNCOGQsxuBCP8XTwXJ2WeEgIg91tc,5455
 fbgemm_gpu/ssd_split_table_batched_embeddings_ops.py,sha256=7qGkO8FARku38mFYl4Bc4qL8dS1wrfyorS9l1m5ZAVA,718
 fbgemm_gpu/tbe_input_multiplexer.py,sha256=DjU7dPHgAT1avXGvgi8SFfw2Pq7yT8S_7IH8qCXoptA,3069
@@ -32,10 +32,10 @@ fbgemm_gpu/docs/merge_pooled_embedding_ops.py,sha256=oJLgSgZQmhsyGLbTmZTxNgQrk65
 fbgemm_gpu/docs/permute_pooled_embedding_ops.py,sha256=tZUqLVXlk5O6VAKKDA-OEMx2fCu5QPOOeoAPZA9_nLY,4454
 fbgemm_gpu/docs/quantize_ops.py,sha256=xTtOaVK1P02ymreE_i21YiyYDZCqhoZY9eWp_mEIRlo,1297
 fbgemm_gpu/docs/sparse_ops.py,sha256=NTcTm0q9h8W2B8PKPoic2fHsAaCbCYunSa_EYK0LtHQ,21382
-fbgemm_gpu/docs/version.py,sha256=Wfdofi1dDfFLcvCYRMDlfYaAgz8J7SIMZPIBXTmCxcA,315
+fbgemm_gpu/docs/version.py,sha256=mYY8Au7MNpxkPW4e6-KmEl1zCys0U_I92tLFTvoN8Oc,315
 fbgemm_gpu/experimental/hstu/__init__.py,sha256=KNisP6qDMwgjgxkGlqUZRNjJ_8o8R-cTmm3HxF7pSqI,1564
 fbgemm_gpu/experimental/hstu/cuda_hstu_attention.py,sha256=5425GRjJuzpXQC-TowgQOCFjZmOwv_EK0lKbURhHBTQ,9920
-fbgemm_gpu/experimental/hstu/fbgemm_gpu_experimental_hstu.so,sha256=ICHJaRmXwXoi_gfj281yCbMnahYSMEjtn4t6rr_blaw,352696288
+fbgemm_gpu/experimental/hstu/fbgemm_gpu_experimental_hstu.so,sha256=Y-D0fGB76-GmSS0nU9dqRf3NkXvCmq86ucb8pPhtU0s,352287576
 fbgemm_gpu/quantize/__init__.py,sha256=pftciXHE7csekDFkl7Ui1AWglVMMnSrOO04mREnUdb0,921
 fbgemm_gpu/quantize/quantize_ops.py,sha256=25AIOv9n2UoxamMUaI6EK1Ur4gSHxbZIReHBtgOjjCs,2228
 fbgemm_gpu/sll/__init__.py,sha256=rgXh35-OFUE54E9gGBq3NGxouGvgMv2ccY2bWUTxONY,4191
@@ -90,10 +90,10 @@ fbgemm_gpu/triton/quantize_ref.py,sha256=q4RBmFaqPVPELU52lbSgB0n26Aun7apeK7bRF2M
 fbgemm_gpu/triton/jagged/__init__.py,sha256=om0yhjuzKuE1UQakFMWHsXN4WNb8mvNkZtYofQ8hdn4,246
 fbgemm_gpu/triton/jagged/triton_jagged_tensor_ops.py,sha256=AIC1G6_QBQtMVTyOyEV4ZKJyDzu36UI_9HDgWmZIRaA,29884
 fbgemm_gpu/utils/__init__.py,sha256=JQQNdcTTaEU6ptK-OW-ZQBwTFxEZZpWOtBXWwEZm39o,354
-fbgemm_gpu/utils/filestore.py,sha256=Zshw1dA03m9aHMMAtETdq4bgOLocyLhzlkAUoG8VkdM,4743
+fbgemm_gpu/utils/filestore.py,sha256=ijoJYDqHUQlv0OrEtLrgSjkNCreCvy5ZXHfd8atwewc,6186
 fbgemm_gpu/utils/loader.py,sha256=1hCEhNvkflniH46fGcrguLeP1z-6uyOu2QFwqKU5CIM,990
 fbgemm_gpu/utils/torch_library.py,sha256=dQcHv1qgpu5QYlJjxjd6oeHjtxnmmXzx3PL6vjCmxL4,4199
-fbgemm_gpu_hstu_nightly-2025.6.15.dist-info/METADATA,sha256=SxliCSqrubFUeLb4Kc97w7fTVCFw0mTOAK4PtWGAw6U,2794
-fbgemm_gpu_hstu_nightly-2025.6.15.dist-info/WHEEL,sha256=Nkv8TSWVt7XcnRf1cdq5HOzycTl6Pjzlmn7gPSv4NiQ,108
-fbgemm_gpu_hstu_nightly-2025.6.15.dist-info/top_level.txt,sha256=2tlbTWLkPjhqvLF_6BbqKzkcPluSE-oPRVjI8axK76I,11
-fbgemm_gpu_hstu_nightly-2025.6.15.dist-info/RECORD,,
+fbgemm_gpu_hstu_nightly-2025.6.17.dist-info/METADATA,sha256=JW6tQAPy8jF9_SLsAZ8BY4c4YnTy3uVQaEc2ac8VyA4,2654
+fbgemm_gpu_hstu_nightly-2025.6.17.dist-info/WHEEL,sha256=Nkv8TSWVt7XcnRf1cdq5HOzycTl6Pjzlmn7gPSv4NiQ,108
+fbgemm_gpu_hstu_nightly-2025.6.17.dist-info/top_level.txt,sha256=2tlbTWLkPjhqvLF_6BbqKzkcPluSE-oPRVjI8axK76I,11
+fbgemm_gpu_hstu_nightly-2025.6.17.dist-info/RECORD,,