fbgemm-gpu-genai-nightly 2025.10.26__cp313-cp313-manylinux_2_28_x86_64.whl → 2025.10.28__cp313-cp313-manylinux_2_28_x86_64.whl

fbgemm_gpu/docs/target.genai.json.py

@@ -1,6 +1,6 @@
 
 {
-    "version": "2025.10.26",
+    "version": "2025.10.28",
     "target": "genai",
     "variant": "cuda"
 }

fbgemm_gpu/tbe/ssd/training.py

@@ -1984,12 +1984,13 @@ class SSDTableBatchedEmbeddingBags(nn.Module):
                     # Store info for evicting the previous iteration's
                     # scratch pad after the corresponding backward pass is
                     # done
-                    self.ssd_location_update_data.append(
-                        (
-                            sp_curr_prev_map_gpu,
-                            inserted_rows,
+                    if self.training:
+                        self.ssd_location_update_data.append(
+                            (
+                                sp_curr_prev_map_gpu,
+                                inserted_rows,
+                            )
                         )
-                    )
 
             # Ensure the previous iterations eviction is complete
             current_stream.wait_event(self.ssd_event_sp_evict)
@@ -2173,7 +2174,7 @@ class SSDTableBatchedEmbeddingBags(nn.Module):
 
             # Store scratch pad info for post backward eviction only for training
             # for eval job, no backward pass, so no need to store this info
-            if self.training and not self._embedding_cache_mode:
+            if self.training:
                 self.ssd_scratch_pad_eviction_data.append(
                     (
                         inserted_rows,
@@ -4548,6 +4549,12 @@ class SSDTableBatchedEmbeddingBags(nn.Module):
                 if len(self.ssd_scratch_pad_eviction_data) > 0:
                     self.ssd_scratch_pad_eviction_data.pop(0)
                     if len(self.ssd_scratch_pad_eviction_data) > 0:
+                        # Wait for any pending backend reads to the next scratch pad
+                        # to complete before we write to it. Otherwise, stale backend data
+                        # will overwrite our direct_write updates.
+                        # The ssd_event_get marks completion of backend fetch operations.
+                        current_stream.wait_event(self.ssd_event_get)
+
                         # if scratch pad exists, write to next batch scratch pad
                         sp = self.ssd_scratch_pad_eviction_data[0][0]
                         sp_idx = self.ssd_scratch_pad_eviction_data[0][1].to(

{fbgemm_gpu_genai_nightly-2025.10.26.dist-info → fbgemm_gpu_genai_nightly-2025.10.28.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fbgemm_gpu_genai_nightly
-Version: 2025.10.26
+Version: 2025.10.28
 Home-page: https://github.com/pytorch/fbgemm
 Author: FBGEMM Team
 Author-email: packages@pytorch.org

{fbgemm_gpu_genai_nightly-2025.10.26.dist-info → fbgemm_gpu_genai_nightly-2025.10.28.dist-info}/RECORD

@@ -1,8 +1,8 @@
 fbgemm_gpu/__init__.py,sha256=A3DuseilQ-sEtBpeZsG0LOqN5Cl3e5DHI_YgCZEMhnE,6283
-fbgemm_gpu/asmjit.so,sha256=PFeEgzpuz45ai1N1fj0C87yHOw9OZFoW6N9VZHyxxHI,484232
+fbgemm_gpu/asmjit.so,sha256=UxnhHlu9LgmoRXa8fZwSX56b5QKffBxfAOs0AZLxRfk,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=DEQ1gRPSFaPAUbupt2byc0F7ZpHYo1yagdZ1Se9Gq1A,5807256
+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
@@ -32,9 +32,9 @@ 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=ajhntGyjTTpz2gTBZ0AIGOreSmrqXYTCsLcujrtNSHk,79
+fbgemm_gpu/docs/target.genai.json.py,sha256=8qUPmreVJ1_8egLGQELc4MkLwuD8aXPlYC2OfOCPJ3g,79
 fbgemm_gpu/experimental/example/__init__.py,sha256=OvJHZgWnycL1gWKyCXFJCTKuys3KAqx4iadjx3R-tBQ,723
-fbgemm_gpu/experimental/example/fbgemm_gpu_experimental_example_py.so,sha256=Sybrpqn3720uHdnjnrC43SfSqjv05ISYUjwQjVfs4cE,232488
+fbgemm_gpu/experimental/example/fbgemm_gpu_experimental_example_py.so,sha256=jDR2ZPfNbqWytICmnGGIuR7IGDNOCzkXLSDPmAroFos,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
@@ -43,7 +43,7 @@ fbgemm_gpu/experimental/gemm/triton_gemm/grouped_gemm.py,sha256=rbjxTMefjQWgJrWK
 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=TXSnyk3ygjW9aR2AbPvquwn67_gp7HD0XTOX5VM6910,74306608
+fbgemm_gpu/experimental/gen_ai/fbgemm_gpu_experimental_gen_ai.so,sha256=klO-9Aat9MhLF1IBnFhDrK1MBrSYeT_3q01srPCe-iI,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
@@ -99,7 +99,7 @@ fbgemm_gpu/tbe/cache/split_embeddings_cache_ops.py,sha256=vZHj7KIe1DoJDy5eft29Xt
 fbgemm_gpu/tbe/ssd/__init__.py,sha256=wzfMT10cp_dqK2lrebC449hOdexBnizcf_98lA1NyHs,483
 fbgemm_gpu/tbe/ssd/common.py,sha256=1J8K7sTQswgCYWaVwF-ZdCJj7mNN6O9GI70AaZWzJGE,1044
 fbgemm_gpu/tbe/ssd/inference.py,sha256=B_uX66ajGA9YKGlFa5TmGWs7b-b1RFigzwxmENZ9Oio,22816
-fbgemm_gpu/tbe/ssd/training.py,sha256=X-u33ZUpxsVUXdQ8UgVtorywAAse5YLrH-ElAG6RnJs,199517
+fbgemm_gpu/tbe/ssd/training.py,sha256=t_8AfKlajVNFAnn3hPST2e82P0dPxzXAsioxMBJ8jrY,199948
 fbgemm_gpu/tbe/ssd/utils/__init__.py,sha256=5DgmR2HA6NtmYh2ddkUgpDsZ6a7hF0DPedA1gMpdh18,250
 fbgemm_gpu/tbe/ssd/utils/partially_materialized_tensor.py,sha256=SFg2-29b-i49LWm-FlaWUkTz2XzXbicYi_AzVj4jKNE,7601
 fbgemm_gpu/tbe/stats/__init__.py,sha256=on29iDtq7cVNh90JR9aeFNG-K9DDoYq0JryzoplL49I,322
@@ -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-2025.10.26.dist-info/METADATA,sha256=26cyhltFwGI-kyhkYZONi8FZb30P1Y99MRu-nh9dD8g,2656
-fbgemm_gpu_genai_nightly-2025.10.26.dist-info/WHEEL,sha256=Nkv8TSWVt7XcnRf1cdq5HOzycTl6Pjzlmn7gPSv4NiQ,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,,
+fbgemm_gpu_genai_nightly-2025.10.28.dist-info/METADATA,sha256=06TbYlUoKnFZX6E1ZpX2rcQaq4tSOj0kJxNOSQ2yO8M,2656
+fbgemm_gpu_genai_nightly-2025.10.28.dist-info/WHEEL,sha256=Nkv8TSWVt7XcnRf1cdq5HOzycTl6Pjzlmn7gPSv4NiQ,108
+fbgemm_gpu_genai_nightly-2025.10.28.dist-info/top_level.txt,sha256=_2s1Aa08r_eDn0JP4FjOhzK09Q8bVlEI7q8pMep51UY,25
+fbgemm_gpu_genai_nightly-2025.10.28.dist-info/RECORD,,