fbgemm-gpu-genai-nightly 2025.10.20__cp311-cp311-manylinux_2_28_x86_64.whl → 2025.10.26__cp311-cp311-manylinux_2_28_x86_64.whl

fbgemm_gpu/__init__.py

@@ -15,7 +15,6 @@ import torch
 # Based on the FBGEMM-PyTorch compatibility table at
 # https://docs.pytorch.org/FBGEMM/general/Releases.html#fbgemm-releases-compatibility
 _fbgemm_torch_compat_table = {
-    "1.4": "2.9",
     "1.3": "2.8",
     "1.2": "2.7",
     "1.1": "2.6",

fbgemm_gpu/docs/target.genai.json.py

@@ -1,6 +1,6 @@
 
 {
-    "version": "2025.10.20",
+    "version": "2025.10.26",
     "target": "genai",
     "variant": "cuda"
 }

fbgemm_gpu/experimental/gemm/triton_gemm/fp8_gemm.py

@@ -3840,6 +3840,10 @@ _MATMUL_CONFIG_TUPLES_PINGPONG_4K_8K_16K = [
     (256, 128, 128, 1, 1, 2, 16, 1, 8, 2),
     (128, 256, 128, 2, 1, 2, 16, 2, 4, 1),
     (256, 128, 64, 2, 1, 2, 16, 1, 4, 2),
+    (128, 128, 256, 2, 1, 0, 16, 2, 8, 2),
+    (128, 64, 128, 2, 1, 2, 16, 2, 4, 2),
+    (128, 128, 64, 2, 1, 0, 16, 1, 4, 2),
+    (128, 128, 128, 1, 1, 2, 16, 1, 4, 2),
 ]
 
 

fbgemm_gpu/experimental/gen_ai/attention/cutlass_blackwell_fmha/cutlass_blackwell_fmha_interface.py

@@ -61,6 +61,8 @@ def _cutlass_blackwell_fmha_forward(
     softmax_scale: float | None = None,
     causal: bool = False,
     seqlen_kv: torch.Tensor | None = None,
+    page_table: torch.Tensor | None = None,
+    seqlen_k: int | None = None,
     window_left: int = -1,
     window_right: int = -1,
     bottom_right: bool = True,
@@ -79,6 +81,8 @@ def _cutlass_blackwell_fmha_forward(
         softmax_scale=softmax_scale,
         causal=causal,
         seqlen_kv=seqlen_kv,
+        page_table=page_table,
+        seqlen_k=seqlen_k,
         window_size_left=window_left,
         window_size_right=window_right,
         bottom_right=bottom_right,
@@ -171,6 +175,8 @@ class CutlassBlackwellFmhaFunc(torch.autograd.Function):
         max_seq_len_q: Optional[int] = None,
         max_seq_len_k: Optional[int] = None,
         seqlen_kv: Optional[torch.Tensor] = None,
+        page_table: Optional[torch.Tensor] = None,
+        seqlen_k: Optional[int] = None,
         window_size: tuple[int, int] = (-1, -1),
         bottom_right: bool = True,
         deterministic: bool = False,
@@ -220,6 +226,8 @@ class CutlassBlackwellFmhaFunc(torch.autograd.Function):
                 softmax_scale,
                 causal,
                 seqlen_kv,
+                page_table,
+                seqlen_k,
                 window_left,
                 window_right,
                 bottom_right,
@@ -252,6 +260,8 @@ class CutlassBlackwellFmhaFunc(torch.autograd.Function):
         None,
         None,
         None,
+        None,
+        None,
     ]:
         if ctx.is_gen:
             # For gen case, no backward pass is needed (generation is inference only)
@@ -279,7 +289,23 @@ class CutlassBlackwellFmhaFunc(torch.autograd.Function):
             bottom_right=ctx.bottom_right,
             deterministic=ctx.deterministic,
         )
-        return dq, dk, dv, None, None, None, None, None, None, None, None, None, None
+        return (
+            dq,
+            dk,
+            dv,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+            None,
+        )
 
 
 def cutlass_blackwell_fmha_func(
@@ -293,6 +319,8 @@ def cutlass_blackwell_fmha_func(
     max_seq_len_q: int | None = None,
     max_seq_len_k: int | None = None,
     seqlen_kv: torch.Tensor | None = None,
+    page_table: torch.Tensor | None = None,
+    seqlen_k: int | None = None,
     window_size: tuple[int, int] | None = (-1, -1),
     bottom_right: bool = True,
     deterministic: bool = False,
@@ -308,6 +336,8 @@ def cutlass_blackwell_fmha_func(
         max_seq_len_q,
         max_seq_len_k,
         seqlen_kv,
+        page_table,
+        seqlen_k,
         window_size,
         bottom_right,
         deterministic,

fbgemm_gpu/experimental/gen_ai/bench/quantize_ops.py

@@ -175,7 +175,7 @@ class QuantizeOpBase(metaclass=abc.ABCMeta):
             if use_cuda_graph:
                 with torch.cuda.stream(torch.cuda.Stream()):
                     t = triton.testing.do_bench_cudagraph(
-                        lambda: self.quantize_and_compute(*args, **kwargs)
+                        lambda: self.quantize_and_compute(*args, **kwargs), rep=200
                     )
             else:
                 t = triton.testing.do_bench(
@@ -188,7 +188,7 @@ class QuantizeOpBase(metaclass=abc.ABCMeta):
                 if use_cuda_graph:
                     with torch.cuda.stream(torch.cuda.Stream()):
                         t = triton.testing.do_bench_cudagraph(
-                            lambda: self.compute(*args, **kwargs)
+                            lambda: self.compute(*args, **kwargs), rep=200
                         )
                 else:
                     t = triton.testing.do_bench(lambda: self.compute(*args, **kwargs))

fbgemm_gpu/split_table_batched_embeddings_ops_training.py

@@ -1714,10 +1714,129 @@ class SplitTableBatchedEmbeddingBagsCodegen(nn.Module):
             tbe_id=self.uuid,
         )
 
-    @torch.jit.ignore
-    def _report_tbe_mem_usage(
+    def _get_tensor_memory(self, tensor_name: str) -> int:
+        """Get memory usage of a tensor in bytes."""
+        if not hasattr(self, tensor_name):
+            self.log(f"Tensor '{tensor_name}' not found, using 0 bytes")
+            return 0
+        tensor = getattr(self, tensor_name)
+        return tensor.numel() * tensor.element_size()
+
+    def _categorize_memory_by_location(
+        self, tensor_names: list[str]
+    ) -> tuple[int, int]:
+        """Categorize memory into HBM and UVM for given tensors.
+
+        Returns:
+            (hbm_bytes, uvm_bytes)
+        """
+        uvm_set = set(self._uvm_tensors_log)
+        hbm_bytes = 0
+        uvm_bytes = 0
+
+        for name in tensor_names:
+            size = self._get_tensor_memory(name)
+            if name in uvm_set:
+                uvm_bytes += size
+            else:
+                hbm_bytes += size
+
+        return hbm_bytes, uvm_bytes
+
+    def _report_hbm_breakdown(
+        self,
+        stats_reporter: TBEStatsReporter,
+        embeddings: int,
+        optimizer_states: int,
+        cache: int,
+        total_static_sparse: int,
+        ephemeral: int,
+    ) -> None:
+        """Report HBM memory breakdown to stats reporter."""
+        stats_reporter.report_data_amount(
+            iteration_step=self.step,
+            event_name="tbe.hbm.embeddings",
+            data_bytes=embeddings,
+            embedding_id=self.logging_table_name,
+            tbe_id=self.uuid,
+        )
+        stats_reporter.report_data_amount(
+            iteration_step=self.step,
+            event_name="tbe.hbm.optimizer_states",
+            data_bytes=optimizer_states,
+            embedding_id=self.logging_table_name,
+            tbe_id=self.uuid,
+        )
+        stats_reporter.report_data_amount(
+            iteration_step=self.step,
+            event_name="tbe.hbm.cache",
+            data_bytes=cache,
+            embedding_id=self.logging_table_name,
+            tbe_id=self.uuid,
+        )
+        stats_reporter.report_data_amount(
+            iteration_step=self.step,
+            event_name="tbe.hbm.total_static_sparse",
+            data_bytes=total_static_sparse,
+            embedding_id=self.logging_table_name,
+            tbe_id=self.uuid,
+        )
+        stats_reporter.report_data_amount(
+            iteration_step=self.step,
+            event_name="tbe.hbm.ephemeral",
+            data_bytes=ephemeral,
+            embedding_id=self.logging_table_name,
+            tbe_id=self.uuid,
+        )
+
+    def _report_uvm_breakdown(
         self,
+        stats_reporter: TBEStatsReporter,
+        embeddings: int,
+        optimizer_states: int,
+        cache: int,
+        total_static_sparse: int,
+        ephemeral: int,
     ) -> None:
+        """Report UVM memory breakdown to stats reporter."""
+        stats_reporter.report_data_amount(
+            iteration_step=self.step,
+            event_name="tbe.uvm.embeddings",
+            data_bytes=embeddings,
+            embedding_id=self.logging_table_name,
+            tbe_id=self.uuid,
+        )
+        stats_reporter.report_data_amount(
+            iteration_step=self.step,
+            event_name="tbe.uvm.optimizer_states",
+            data_bytes=optimizer_states,
+            embedding_id=self.logging_table_name,
+            tbe_id=self.uuid,
+        )
+        stats_reporter.report_data_amount(
+            iteration_step=self.step,
+            event_name="tbe.uvm.cache",
+            data_bytes=cache,
+            embedding_id=self.logging_table_name,
+            tbe_id=self.uuid,
+        )
+        stats_reporter.report_data_amount(
+            iteration_step=self.step,
+            event_name="tbe.uvm.total_static_sparse",
+            data_bytes=total_static_sparse,
+            embedding_id=self.logging_table_name,
+            tbe_id=self.uuid,
+        )
+        stats_reporter.report_data_amount(
+            iteration_step=self.step,
+            event_name="tbe.uvm.ephemeral",
+            data_bytes=ephemeral,
+            embedding_id=self.logging_table_name,
+            tbe_id=self.uuid,
+        )
+
+    @torch.jit.ignore
+    def _report_tbe_mem_usage(self) -> None:
         if self.stats_reporter is None:
             return
 
@@ -1725,22 +1844,24 @@ class SplitTableBatchedEmbeddingBagsCodegen(nn.Module):
         if not stats_reporter.should_report(self.step):
             return
 
+        # Calculate total memory from all parameters and buffers (always needed)
         total_mem_usage = sum(
-            param.numel() * param.element_size() for param in self.parameters()
-        ) + sum(buffer.numel() * buffer.element_size() for buffer in self.buffers())
+            p.numel() * p.element_size() for p in self.parameters()
+        ) + sum(b.numel() * b.element_size() for b in self.buffers())
+
+        # Calculate total HBM and UVM usage (always needed)
         if self.use_cpu:
             total_hbm_usage = 0
             total_uvm_usage = total_mem_usage
         else:
-            # hbm usage is total usage minus uvm usage
             total_uvm_usage = sum(
-                getattr(self, tensor_name).numel()
-                * getattr(self, tensor_name).element_size()
-                for tensor_name in self._uvm_tensors_log
-                if hasattr(self, tensor_name)
+                self._get_tensor_memory(name)
+                for name in self._uvm_tensors_log
+                if hasattr(self, name)
             )
             total_hbm_usage = total_mem_usage - total_uvm_usage
 
+        # Report total memory usage metrics (always reported for backward compatibility)
         stats_reporter.report_data_amount(
             iteration_step=self.step,
             event_name="tbe.total_hbm_usage",
@@ -1756,6 +1877,76 @@ class SplitTableBatchedEmbeddingBagsCodegen(nn.Module):
             tbe_id=self.uuid,
         )
 
+        # Check if detailed memory breakdown is enabled via environment variable
+        # Set FBGEMM_TBE_MEM_BREAKDOWN=1 to enable expensive detailed breakdown
+        enable_detailed_breakdown = (
+            int(os.environ.get("FBGEMM_TBE_MEM_BREAKDOWN", "0")) == 1
+        )
+
+        if not enable_detailed_breakdown:
+            return
+
+        # Tensor groups for sparse memory categorization
+        weight_tensors = ["weights_dev", "weights_host", "weights_uvm"]
+        optimizer_tensors = [
+            "momentum1_dev",
+            "momentum1_host",
+            "momentum1_uvm",
+            "momentum2_dev",
+            "momentum2_host",
+            "momentum2_uvm",
+        ]
+        cache_tensors = [
+            "lxu_cache_weights",
+            "lxu_cache_state",
+            "lxu_state",
+            "cache_hash_size_cumsum",
+            "cache_index_table_map",
+            "cache_miss_counter",
+            "lxu_cache_locking_counter",
+        ]
+
+        # Calculate total memory for each component
+        weights_total = sum(self._get_tensor_memory(t) for t in weight_tensors)
+        optimizer_total = sum(self._get_tensor_memory(t) for t in optimizer_tensors)
+        cache_total = sum(self._get_tensor_memory(t) for t in cache_tensors)
+
+        # Categorize memory by location (HBM vs UVM)
+        if self.use_cpu:
+            weights_hbm, weights_uvm = 0, weights_total
+            opt_hbm, opt_uvm = 0, optimizer_total
+            cache_hbm, cache_uvm = 0, cache_total
+        else:
+            weights_hbm, weights_uvm = self._categorize_memory_by_location(
+                weight_tensors
+            )
+            opt_hbm, opt_uvm = self._categorize_memory_by_location(optimizer_tensors)
+            cache_hbm, cache_uvm = self._categorize_memory_by_location(cache_tensors)
+
+        # Calculate ephemeral memory split between HBM and UVM
+        static_sparse_hbm = weights_hbm + opt_hbm + cache_hbm
+        static_sparse_uvm = weights_uvm + opt_uvm + cache_uvm
+        ephemeral_hbm = total_hbm_usage - static_sparse_hbm
+        ephemeral_uvm = total_uvm_usage - static_sparse_uvm
+
+        # Report granular memory breakdowns
+        self._report_hbm_breakdown(
+            stats_reporter,
+            weights_hbm,
+            opt_hbm,
+            cache_hbm,
+            static_sparse_hbm,
+            ephemeral_hbm,
+        )
+        self._report_uvm_breakdown(
+            stats_reporter,
+            weights_uvm,
+            opt_uvm,
+            cache_uvm,
+            static_sparse_uvm,
+            ephemeral_uvm,
+        )
+
     @torch.jit.ignore
     def _report_io_size_count(self, event: str, data: Tensor) -> Tensor:
         if self.stats_reporter is None:

fbgemm_gpu/tbe/cache/kv_embedding_ops_inference.py

@@ -76,6 +76,7 @@ class KVEmbeddingInference(IntNBitTableBatchedEmbeddingBagsCodegen):
         reverse_qparam: bool = False,  # True to load qparams at end of each row; False to load qparam at begnning of each row.
         feature_names_per_table: Optional[list[list[str]]] = None,
         indices_dtype: torch.dtype = torch.int32,  # Used for construction of the remap_indices tensors.  Should match the dtype of the indices passed in the forward() call (INT32 or INT64).
+        embedding_cache_mode: bool = False,  # True for zero initialization, False for randomized initialization
     ) -> None:  # noqa C901  # tuple of (rows, dims,)
         super(KVEmbeddingInference, self).__init__(
             embedding_specs=embedding_specs,
@@ -114,9 +115,13 @@ class KVEmbeddingInference(IntNBitTableBatchedEmbeddingBagsCodegen):
         num_shards = 32
         uniform_init_lower: float = -0.01
         uniform_init_upper: float = 0.01
+
         # pyre-fixme[4]: Attribute must be annotated.
         self.kv_embedding_cache = torch.classes.fbgemm.DramKVEmbeddingInferenceWrapper(
-            num_shards, uniform_init_lower, uniform_init_upper
+            num_shards,
+            uniform_init_lower,
+            uniform_init_upper,
+            embedding_cache_mode,  # in embedding_cache_mode, we disable random init
         )
 
         self.specs: list[tuple[int, int, int]] = [

{fbgemm_gpu_genai_nightly-2025.10.20.dist-info → fbgemm_gpu_genai_nightly-2025.10.26.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fbgemm_gpu_genai_nightly
-Version: 2025.10.20
+Version: 2025.10.26
 Home-page: https://github.com/pytorch/fbgemm
 Author: FBGEMM Team
 Author-email: packages@pytorch.org

{fbgemm_gpu_genai_nightly-2025.10.20.dist-info → fbgemm_gpu_genai_nightly-2025.10.26.dist-info}/RECORD

@@ -1,8 +1,8 @@
-fbgemm_gpu/__init__.py,sha256=BxqlqUln-q_ljJpo_Cg3R2RYZqxCbZ0UjvdBe6DzNZk,6301
-fbgemm_gpu/asmjit.so,sha256=6jIoG1FsEL7UVNPKBvLusNQxe5VjorDnlUmfjuLv-MQ,485392
+fbgemm_gpu/__init__.py,sha256=A3DuseilQ-sEtBpeZsG0LOqN5Cl3e5DHI_YgCZEMhnE,6283
+fbgemm_gpu/asmjit.so,sha256=RxTYI8zY4PpIBRpSKT_-U7bRIVeTRohdtRFUmLNU1tQ,501728
 fbgemm_gpu/batched_unary_embeddings_ops.py,sha256=GYeJ9pg-Wc9FokXVci_npDsL6UV18-pJXID2xzrJ9O8,2904
 fbgemm_gpu/enums.py,sha256=37ewGSfO1x7sO31ZkRiqV1yKuklfHXT5qZIxzeeGogo,755
-fbgemm_gpu/fbgemm.so,sha256=jzEQ0d5jEsVw3X1VLInzOaJ-e-3ecs4hbVzUFgS0ifE,5798280
+fbgemm_gpu/fbgemm.so,sha256=B9y6MDLC6Ou7Bw_pT07Xfw5g5Q4j2yn9Xsp96QVpgEU,5646712
 fbgemm_gpu/metrics.py,sha256=TsurFLJf0nJvPDN7urWb4LMQlf5RgdWPTTTDO7S4wtI,5663
 fbgemm_gpu/permute_pooled_embedding_modules.py,sha256=vOXMYclaGnwSt0St_SOAlAe18kz6WjMyTeHnC9jLhcE,5130
 fbgemm_gpu/permute_pooled_embedding_modules_split.py,sha256=f3VJvH_kw9Ltd_DXtaf_PJPHmlmEWrQgzQ7MDkhh5Nw,2746
@@ -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=76ME0692CC691xpjiOsY3Xxy-LD_XKs8w9vq1gcm9tM,16440
 fbgemm_gpu/split_table_batched_embeddings_ops_inference.py,sha256=dGC85xjQiRUrequBibSf9oMAVHT5Q49zsVo2zW4n_88,81679
-fbgemm_gpu/split_table_batched_embeddings_ops_training.py,sha256=lF9eP6GDTyqbEJgl-SO6gNYUk2dv2YE2bMEtzGkY21c,173757
+fbgemm_gpu/split_table_batched_embeddings_ops_training.py,sha256=f0sXfvkE0Wx0Rd3qTT4XmCbBK0wYgWGzhPncZEv-p48,180420
 fbgemm_gpu/split_table_batched_embeddings_ops_training_common.py,sha256=e3O9ElaWBGvG7TdT3Ok_8cB06jhskXuyCQ0t40dzsEY,5449
 fbgemm_gpu/ssd_split_table_batched_embeddings_ops.py,sha256=7qGkO8FARku38mFYl4Bc4qL8dS1wrfyorS9l1m5ZAVA,718
 fbgemm_gpu/tbe_input_multiplexer.py,sha256=TQjwkJ2JkOaQsMYuRdk9RbNa9759EPEtx8bYclChtZY,3063
@@ -32,27 +32,27 @@ 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=gSLUFdnu8lle_6gLewFkM20wL3ek2jKLvDGMKR6POaY,27292
-fbgemm_gpu/docs/target.genai.json.py,sha256=EffeoYnTPp4BLew_sFOpBinWQgXup1DReXuDroDTnh8,79
+fbgemm_gpu/docs/target.genai.json.py,sha256=ajhntGyjTTpz2gTBZ0AIGOreSmrqXYTCsLcujrtNSHk,79
 fbgemm_gpu/experimental/example/__init__.py,sha256=OvJHZgWnycL1gWKyCXFJCTKuys3KAqx4iadjx3R-tBQ,723
-fbgemm_gpu/experimental/example/fbgemm_gpu_experimental_example_py.so,sha256=rGppi6HTFubDghRJjz7WlBOpZv9mdEzF6NfHETnnRw0,399600
+fbgemm_gpu/experimental/example/fbgemm_gpu_experimental_example_py.so,sha256=CFzw20PsjS83INRnHwrOYcWybr1SaUXcA3jod77r-LM,243904
 fbgemm_gpu/experimental/example/utils.py,sha256=Je__VkMlBMLOhh7NXOocOdvaa2gz9kl9Dkqeu25tpFA,562
 fbgemm_gpu/experimental/gemm/triton_gemm/__init__.py,sha256=1CqUfzlYyXTvU-BNaUq4RZpLV-2lKAVCAHeJzSIZFWw,419
 fbgemm_gpu/experimental/gemm/triton_gemm/fp4_quantize.py,sha256=2RjIDSzUXtoFoC2ryp-C-j5H83mbSjPwvsvTrThfrqE,215658
-fbgemm_gpu/experimental/gemm/triton_gemm/fp8_gemm.py,sha256=5m4SdgUsf2rM_Vul8czgRn_5oVnyi-52TmeidXh05hg,152754
+fbgemm_gpu/experimental/gemm/triton_gemm/fp8_gemm.py,sha256=q1o0FfGcUAQjkxKlJjjqKVSaPd3HaBSs6L9qVHY7qKI,152924
 fbgemm_gpu/experimental/gemm/triton_gemm/grouped_gemm.py,sha256=rbjxTMefjQWgJrWK_bYFtBklJigFwv4awPeVexkkiIA,44511
 fbgemm_gpu/experimental/gemm/triton_gemm/matmul_perf_model.py,sha256=SltbY_dsit5e7B8lDIB_VYPrEq0t9kckthj9mQaVNfA,7571
 fbgemm_gpu/experimental/gemm/triton_gemm/utils.py,sha256=rULXIpVaaRS3GKUZ1RHcWUrUyy0xMVREwS1SFShGgcw,4302
 fbgemm_gpu/experimental/gen_ai/__init__.py,sha256=r3NlNCXuIh0pfKwKU5v14y6AZkpoIkKWbtzxSprgeKA,1713
-fbgemm_gpu/experimental/gen_ai/fbgemm_gpu_experimental_gen_ai.so,sha256=sVB4x2xinMV-3rG15pisuG19zHLEEcTlkxRUSYKEEA4,286901408
+fbgemm_gpu/experimental/gen_ai/fbgemm_gpu_experimental_gen_ai.so,sha256=z38Aw4HDVn4BRpK6a9g1b0PdbClb_10LiqWzn22C0hM,74888976
 fbgemm_gpu/experimental/gen_ai/quantize.py,sha256=KAljWSdN-1_c5DWfT-3MDxWLMULK49Yu36t6TmQI9Tw,12599
 fbgemm_gpu/experimental/gen_ai/attention/cutlass_blackwell_fmha/__init__.py,sha256=ntFgFs0foi6NQx8eqs5I3fCjzKSI0spXfEWiMhlcT00,897
 fbgemm_gpu/experimental/gen_ai/attention/cutlass_blackwell_fmha/cutlass_blackwell_fmha_custom_op.py,sha256=FADVTYzS2u8fA-3iChS5CbtWd0mWF8F3lnXcwr_7vDw,7821
-fbgemm_gpu/experimental/gen_ai/attention/cutlass_blackwell_fmha/cutlass_blackwell_fmha_interface.py,sha256=7ydkrZ6qqyiah1dlJX6EuEXXw6WwOqCj7D48PWNJcUw,9259
+fbgemm_gpu/experimental/gen_ai/attention/cutlass_blackwell_fmha/cutlass_blackwell_fmha_interface.py,sha256=K9cPXGOF4E9VHzuVtJjDPoTC7JjhEqS1RmmWSehQrKU,9887
 fbgemm_gpu/experimental/gen_ai/bench/__init__.py,sha256=XpAK_eyqDSKeFC5J9KpnKtbZG07mrDh9d2j1LFKzr-8,404
 fbgemm_gpu/experimental/gen_ai/bench/comm_bench.py,sha256=ApEyJOf_rdIo8V_EgvhZXBGNov8ITC_dnB95v8szulI,8515
 fbgemm_gpu/experimental/gen_ai/bench/gather_scatter_bench.py,sha256=K9Nib6D7xJbw1QwEVuCJrVyI1qs988moo3cieVKYuFY,12057
 fbgemm_gpu/experimental/gen_ai/bench/quantize_bench.py,sha256=BWl6t-4acbuRSEX2aVNDlFrSWZkqMWK2sI3VONaMd3Q,24047
-fbgemm_gpu/experimental/gen_ai/bench/quantize_ops.py,sha256=H6AchejyZs76_snM_ae5vV0cPr_Q0h35OQ8qED0r1N4,104915
+fbgemm_gpu/experimental/gen_ai/bench/quantize_ops.py,sha256=cDZS2rCb1W2IEQYxsnGjauhlUhg2PFZ-9LqJ_SEdbiQ,104933
 fbgemm_gpu/experimental/gen_ai/moe/README.md,sha256=z9ybHmv4KFJ1drj5OByuFaOY0tRQwwiIW3Q22TB_2-k,904
 fbgemm_gpu/experimental/gen_ai/moe/__init__.py,sha256=lwSvff07yEav024B1XyfgW8r8hwNe--aEDywcO7rnbM,1905
 fbgemm_gpu/experimental/gen_ai/moe/activation.py,sha256=NiXhWyCNagI3P9N3N89iSX7xKuShdkq9DxEUAzoV6y0,7892
@@ -94,7 +94,7 @@ fbgemm_gpu/tbe/bench/tbe_data_config_loader.py,sha256=MNddYzoRlu0mNhnsVVG57JN7pB
 fbgemm_gpu/tbe/bench/tbe_data_config_param_models.py,sha256=sptdqcNE9JlgyIJ17neZaMxagKG469_ynX0mVx_JKBY,6090
 fbgemm_gpu/tbe/bench/utils.py,sha256=cq_6FJHlgZ5femAK6XKpj7nJ9jc03qXI16N1ht1CcLg,1721
 fbgemm_gpu/tbe/cache/__init__.py,sha256=lrYwhvqX2eWN0vAPe89HYgMW_O1vccoOcoFHJ9cyM-s,398
-fbgemm_gpu/tbe/cache/kv_embedding_ops_inference.py,sha256=m8rCF8bc_5vBrg9677TDZTQXqRdFt6YPUVVKv85up5s,14380
+fbgemm_gpu/tbe/cache/kv_embedding_ops_inference.py,sha256=VmG9EennGcq2By8Tj8VkFsJG0oOCGw8EhlPo8-t--Fk,14604
 fbgemm_gpu/tbe/cache/split_embeddings_cache_ops.py,sha256=vZHj7KIe1DoJDy5eft29XtGg6I-tRx60tjKOcTHRAYI,1321
 fbgemm_gpu/tbe/ssd/__init__.py,sha256=wzfMT10cp_dqK2lrebC449hOdexBnizcf_98lA1NyHs,483
 fbgemm_gpu/tbe/ssd/common.py,sha256=1J8K7sTQswgCYWaVwF-ZdCJj7mNN6O9GI70AaZWzJGE,1044
@@ -121,7 +121,7 @@ fbgemm_gpu/utils/loader.py,sha256=1hCEhNvkflniH46fGcrguLeP1z-6uyOu2QFwqKU5CIM,99
 fbgemm_gpu/utils/torch_library.py,sha256=ywsAHjbuwesj50LjEu99WkAH17FlaVgePZ9OmFg6YE4,4193
 list_versions/__init__.py,sha256=UmTeqCk-UJWFtlZQWvZao3xvui2w9E3X_JdOXVjRaNw,315
 list_versions/cli_run.py,sha256=CChZoXQ-tiKaWboXAYlPVJ5w8K5zAKiKcncA087I1sc,4508
-fbgemm_gpu_genai_nightly-2025.10.20.dist-info/METADATA,sha256=YSlW54hUiRgcqmN2NJaqU8mF-KLXXKCF_MpwTI2USC0,2656
-fbgemm_gpu_genai_nightly-2025.10.20.dist-info/WHEEL,sha256=V2Q6mQKbouIadCxoRjt9FQ9oKfi45-uZUcoc77zzs0M,108
-fbgemm_gpu_genai_nightly-2025.10.20.dist-info/top_level.txt,sha256=_2s1Aa08r_eDn0JP4FjOhzK09Q8bVlEI7q8pMep51UY,25
-fbgemm_gpu_genai_nightly-2025.10.20.dist-info/RECORD,,
+fbgemm_gpu_genai_nightly-2025.10.26.dist-info/METADATA,sha256=26cyhltFwGI-kyhkYZONi8FZb30P1Y99MRu-nh9dD8g,2656
+fbgemm_gpu_genai_nightly-2025.10.26.dist-info/WHEEL,sha256=V2Q6mQKbouIadCxoRjt9FQ9oKfi45-uZUcoc77zzs0M,108
+fbgemm_gpu_genai_nightly-2025.10.26.dist-info/top_level.txt,sha256=_2s1Aa08r_eDn0JP4FjOhzK09Q8bVlEI7q8pMep51UY,25
+fbgemm_gpu_genai_nightly-2025.10.26.dist-info/RECORD,,