fbgemm-gpu-genai-nightly 2025.10.11__cp39-cp39-manylinux_2_28_x86_64.whl → 2025.10.12__cp39-cp39-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 = "2025.10.11"
+__version__: str = "2025.10.12"
 __target__: str = "genai"
 __variant__: str = "cuda"

fbgemm_gpu/tbe/ssd/training.py

@@ -3971,8 +3971,8 @@ class SSDTableBatchedEmbeddingBags(nn.Module):
             self.step, stats_reporter.report_interval  # pyre-ignore
         )
 
-        if len(dram_kv_perf_stats) != 24:
-            logging.error("dram cache perf stats should have 24 elements")
+        if len(dram_kv_perf_stats) != 36:
+            logging.error("dram cache perf stats should have 36 elements")
             return
 
         dram_read_duration = dram_kv_perf_stats[0]
@@ -4002,6 +4002,19 @@ class SSDTableBatchedEmbeddingBags(nn.Module):
         dram_kv_actual_used_chunk_bytes = dram_kv_perf_stats[21]
         dram_kv_num_rows = dram_kv_perf_stats[22]
         dram_kv_read_counts = dram_kv_perf_stats[23]
+        dram_metadata_write_sharding_total_duration = dram_kv_perf_stats[24]
+        dram_metadata_write_total_duration = dram_kv_perf_stats[25]
+        dram_metadata_write_allocate_avg_duration = dram_kv_perf_stats[26]
+        dram_metadata_write_lookup_cache_avg_duration = dram_kv_perf_stats[27]
+        dram_metadata_write_acquire_lock_avg_duration = dram_kv_perf_stats[28]
+        dram_metadata_write_cache_miss_avg_count = dram_kv_perf_stats[29]
+
+        dram_read_metadata_total_duration = dram_kv_perf_stats[30]
+        dram_read_metadata_sharding_total_duration = dram_kv_perf_stats[31]
+        dram_read_metadata_cache_hit_copy_avg_duration = dram_kv_perf_stats[32]
+        dram_read_metadata_lookup_cache_total_avg_duration = dram_kv_perf_stats[33]
+        dram_read_metadata_acquire_lock_avg_duration = dram_kv_perf_stats[34]
+        dram_read_read_metadata_load_size = dram_kv_perf_stats[35]
 
         stats_reporter.report_duration(
             iteration_step=self.step,
@@ -4168,6 +4181,88 @@ class SSDTableBatchedEmbeddingBags(nn.Module):
             data_bytes=dram_kv_num_rows,
             enable_tb_metrics=True,
         )
+        stats_reporter.report_duration(
+            iteration_step=self.step,
+            event_name="dram_kv.perf.set.dram_eviction_score_write_sharding_total_duration_us",
+            duration_ms=dram_metadata_write_sharding_total_duration,
+            enable_tb_metrics=True,
+            time_unit="us",
+        )
+        stats_reporter.report_duration(
+            iteration_step=self.step,
+            event_name="dram_kv.perf.set.dram_eviction_score_write_total_duration_us",
+            duration_ms=dram_metadata_write_total_duration,
+            enable_tb_metrics=True,
+            time_unit="us",
+        )
+        stats_reporter.report_duration(
+            iteration_step=self.step,
+            event_name="dram_kv.perf.set.dram_eviction_score_write_allocate_avg_duration_us",
+            duration_ms=dram_metadata_write_allocate_avg_duration,
+            enable_tb_metrics=True,
+            time_unit="us",
+        )
+        stats_reporter.report_duration(
+            iteration_step=self.step,
+            event_name="dram_kv.perf.set.dram_eviction_score_write_lookup_cache_avg_duration_us",
+            duration_ms=dram_metadata_write_lookup_cache_avg_duration,
+            enable_tb_metrics=True,
+            time_unit="us",
+        )
+        stats_reporter.report_duration(
+            iteration_step=self.step,
+            event_name="dram_kv.perf.set.dram_eviction_score_write_acquire_lock_avg_duration_us",
+            duration_ms=dram_metadata_write_acquire_lock_avg_duration,
+            enable_tb_metrics=True,
+            time_unit="us",
+        )
+        stats_reporter.report_data_amount(
+            iteration_step=self.step,
+            event_name="dram_kv.perf.set.dram_eviction_score_write_cache_miss_avg_count",
+            data_bytes=dram_metadata_write_cache_miss_avg_count,
+            enable_tb_metrics=True,
+        )
+        stats_reporter.report_duration(
+            iteration_step=self.step,
+            event_name="dram_kv.perf.get.dram_eviction_score_read_total_duration_us",
+            duration_ms=dram_read_metadata_total_duration,
+            enable_tb_metrics=True,
+            time_unit="us",
+        )
+        stats_reporter.report_duration(
+            iteration_step=self.step,
+            event_name="dram_kv.perf.get.dram_eviction_score_read_sharding_total_duration_us",
+            duration_ms=dram_read_metadata_sharding_total_duration,
+            enable_tb_metrics=True,
+            time_unit="us",
+        )
+        stats_reporter.report_duration(
+            iteration_step=self.step,
+            event_name="dram_kv.perf.get.dram_eviction_score_read_cache_hit_copy_avg_duration_us",
+            duration_ms=dram_read_metadata_cache_hit_copy_avg_duration,
+            enable_tb_metrics=True,
+            time_unit="us",
+        )
+        stats_reporter.report_duration(
+            iteration_step=self.step,
+            event_name="dram_kv.perf.get.dram_eviction_score_read_lookup_cache_total_avg_duration_us",
+            duration_ms=dram_read_metadata_lookup_cache_total_avg_duration,
+            enable_tb_metrics=True,
+            time_unit="us",
+        )
+        stats_reporter.report_duration(
+            iteration_step=self.step,
+            event_name="dram_kv.perf.get.dram_eviction_score_read_acquire_lock_avg_duration_us",
+            duration_ms=dram_read_metadata_acquire_lock_avg_duration,
+            enable_tb_metrics=True,
+            time_unit="us",
+        )
+        stats_reporter.report_data_amount(
+            iteration_step=self.step,
+            event_name="dram_kv.perf.get.dram_eviction_score_read_load_size",
+            data_bytes=dram_read_read_metadata_load_size,
+            enable_tb_metrics=True,
+        )
 
     def _recording_to_timer(
         self, timer: Optional[AsyncSeriesTimer], **kwargs: Any

{fbgemm_gpu_genai_nightly-2025.10.11.dist-info → fbgemm_gpu_genai_nightly-2025.10.12.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fbgemm_gpu_genai_nightly
-Version: 2025.10.11
+Version: 2025.10.12
 Home-page: https://github.com/pytorch/fbgemm
 Author: FBGEMM Team
 Author-email: packages@pytorch.org

{fbgemm_gpu_genai_nightly-2025.10.11.dist-info → fbgemm_gpu_genai_nightly-2025.10.12.dist-info}/RECORD

@@ -1,8 +1,8 @@
 fbgemm_gpu/__init__.py,sha256=FdQCmpvETH80tlIPP6W8MrOmzLaX9eoGY-fuHtVPbj0,5747
-fbgemm_gpu/asmjit.so,sha256=JDql0ZOBynsM53nSLqGG070z1oYke3r68K90AeBP0hY,484232
+fbgemm_gpu/asmjit.so,sha256=RMRQVgIB3mf-bUr2BOMzrgzIv3H9arLfYnJ8CTGZ8RQ,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/version.py,sha256=37Mm_y7ZEW2PuqA4FzAUhP66mtP19dXqT6kcG1kkcHs,317
+fbgemm_gpu/docs/version.py,sha256=tx4meXyr0j01--CHHgF4Lv7ZZK5ONOKhmiHp_Ggngnw,317
 fbgemm_gpu/experimental/example/__init__.py,sha256=V_XrGMq2oNVMpzwe1srlaTaHeIcZJw5oAGbo3seM_Ks,870
-fbgemm_gpu/experimental/example/fbgemm_gpu_experimental_example_py.so,sha256=eife1eWK6QFDUtDL-EcAQNvgJyfR86tAGpvDNHsSlpA,232480
+fbgemm_gpu/experimental/example/fbgemm_gpu_experimental_example_py.so,sha256=owFe-ilNrqChzBPf7R9BfcY3lpFA6kMuXZn_bKPb2l8,239808
 fbgemm_gpu/experimental/example/utils.py,sha256=Je__VkMlBMLOhh7NXOocOdvaa2gz9kl9Dkqeu25tpFA,562
 fbgemm_gpu/experimental/gemm/triton_gemm/__init__.py,sha256=AqHefiOaN_SjP5ew7RYGuKFuSlhedOJL_6f97TtLv7c,566
 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=qwfuF5E5K4oDiH7RJkpC7zth3kAsG7wv_glCl2A_G2A,1860
-fbgemm_gpu/experimental/gen_ai/fbgemm_gpu_experimental_gen_ai.so,sha256=twpjHd8M2YkcqqcLaOT7CS7wHII4Gs6ij2n0XvhLBfg,72323352
+fbgemm_gpu/experimental/gen_ai/fbgemm_gpu_experimental_gen_ai.so,sha256=zsAN7i8yprDxNNsVPIGEouGZTjGR9Y4_hyYjynvkaEI,73016984
 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=oExepXpjMOwM43gARZARY0UtR-EX2zqRnSrOaQPy448,1044
 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=iepmavrK6cSXrqWg3TqVXL5gVmSftfHJycesp7I0Nw4,194911
+fbgemm_gpu/tbe/ssd/training.py,sha256=X-u33ZUpxsVUXdQ8UgVtorywAAse5YLrH-ElAG6RnJs,199517
 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-2025.10.11.dist-info/METADATA,sha256=6M2BPNwybS1j-b2XjITa24q7t8uq-mIq-MEafDcYh70,2656
-fbgemm_gpu_genai_nightly-2025.10.11.dist-info/WHEEL,sha256=LLKFkWvXQZlw1kR7spdgzB2PptbzNM_HUUPM3Q0J_XE,106
-fbgemm_gpu_genai_nightly-2025.10.11.dist-info/top_level.txt,sha256=_2s1Aa08r_eDn0JP4FjOhzK09Q8bVlEI7q8pMep51UY,25
-fbgemm_gpu_genai_nightly-2025.10.11.dist-info/RECORD,,
+fbgemm_gpu_genai_nightly-2025.10.12.dist-info/METADATA,sha256=mVnsemyIjnAIiA0h1-0U1yChU5jXX7DhKtbyEnX1nUc,2656
+fbgemm_gpu_genai_nightly-2025.10.12.dist-info/WHEEL,sha256=LLKFkWvXQZlw1kR7spdgzB2PptbzNM_HUUPM3Q0J_XE,106
+fbgemm_gpu_genai_nightly-2025.10.12.dist-info/top_level.txt,sha256=_2s1Aa08r_eDn0JP4FjOhzK09Q8bVlEI7q8pMep51UY,25
+fbgemm_gpu_genai_nightly-2025.10.12.dist-info/RECORD,,