2026-08-29 nightly release (5e56ddb)

Author: pytorchbot

cmake/modules/CppLibrary.cmake

@@ -90,7 +90,35 @@ function(fbgemm_get_warning_flags)
     # Accepted and diagnosing on BOTH g++ 11.5 and clang 22 (probed), so it
     # goes here rather than behind the clang guard. Subplan 04 called this out
     # correctly.
-    -Waddress-of-packed-member)
+    -Waddress-of-packed-member
+    # ---- Phase B2: shadowing ----------------------------------------------
+    # Portable: probed accepted AND diagnosing on both g++ 11.5 and clang 22
+    # (the bare flag; -Wshadow-all, -Wshadow-uncaptured-local and
+    # -Wshadow-field-in-constructor are clang-only and are NOT added here).
+    #
+    # This is enabled DEMOTED -- see -Wno-error=shadow in the shared
+    # suppressions. The B0 census measured 299 first-party diagnostics across
+    # 22 files, so enabling it at full strength would break the build outright.
+    # Demoting makes those 299 visible in CI and keeps the census reproducible
+    # instead of being a one-off local probe, while the fixups land
+    # incrementally. The escape comes out when the count reaches zero.
+    #
+    # -Wshadow is the right first Phase B flag on two measurements: it is the
+    # most concentrated (79% of diagnostics in 5 files) and it is the only one
+    # of the three with ZERO third-party exposure, so unlike
+    # -Wshorten-64-to-32 and -Wzero-as-null-pointer-constant it is not blocked
+    # on giving PyTorch/ATen/c10 headers -isystem treatment.
+    -Wshadow
+    # ---- Phase B3: zero as null pointer -----------------------------------
+    # Portable: probed accepted AND diagnosing on g++ 11.5 and clang 22.
+    # Enabled DEMOTED -- see -Wno-error=zero-as-null-pointer-constant.
+    #
+    # B0 measured 449 first-party diagnostics across 130 files. That is the
+    # LARGEST job of the three B flags despite having ~260 fewer diagnostics
+    # than -Wshorten-64-to-32, because it is diffuse: only 19% sit in the top
+    # five files, versus 79% for -Wshadow. Rank by file count and
+    # concentration, not by raw count.
+    -Wzero-as-null-pointer-constant)
 
   # Clang-only warning flags. These are appended to `_cc` ONLY when the host
   # compiler is clang (see the guarded append below), because the OSS CI matrix
@@ -228,7 +256,25 @@ function(fbgemm_get_warning_flags)
     -Wno-vla
     -Wno-error=unused-parameter
     -Wno-error=unknown-pragmas
-    -Wno-error=attributes)
+    -Wno-error=attributes
+    # Phase B2 is landing demoted while its 299 first-party diagnostics are
+    # fixed; see -Wshadow in _cc_common. Portable escape -- probed demoting to
+    # a warning on both g++ and clang, unlike the A2.1 escape which g++
+    # hard-errors on and which therefore had to be clang-guarded.
+    # TODO(T169200065): delete once the -Wshadow count reaches zero. This is
+    # the whole point of enabling it demoted rather than silently.
+    -Wno-error=shadow
+    # Phase B3, landing demoted while its 449 diagnostics across 130 files
+    # are fixed; see -Wzero-as-null-pointer-constant in _cc_common. Portable
+    # escape -- probed demoting to a warning on both g++ and clang.
+    #
+    # This escape also covers the 48 diagnostics coming from PyTorch / ATen /
+    # c10 / folly / thrift headers, which are NOT FBGEMM's to fix. Those need
+    # -isystem treatment (unowned work that B0 newly identified) before this
+    # flag can go to hard error, independently of the first-party backlog.
+    # TODO(T169200065): delete once the count reaches zero AND the external
+    # headers are -isystem.
+    -Wno-error=zero-as-null-pointer-constant)
 
   # Clang suppressions, split by the clang version that made them necessary.
   # The CXX path applies these conditionally on the HOST clang version (below);
@@ -260,8 +306,7 @@ function(fbgemm_get_warning_flags)
 
   set(_cc_suppressions_clang_gt17
     -Wno-vla-cxx-extension
-    -Wno-error=global-constructors
-    -Wno-error=shadow)
+    -Wno-error=global-constructors)
 
   # Full clang-shaped suppression set, assembled unconditionally so it is
   # available even when the HOST compiler is GCC. Used only for the hipcc list.

external/composable_kernel

@@ -16,17 +16,13 @@
 import fbgemm_gpu.batched_unary_embeddings_ops as batched_unary_embeddings_ops
 import numpy as np
 import torch
+from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
 from torch.profiler import profile
 
 # pyre-fixme[16]: Module `fbgemm_gpu` has no attribute `open_source`.
 open_source: bool = getattr(fbgemm_gpu, "open_source", False)
 
-if open_source:
-    # pyre-ignore[21]
-    from bench_utils import benchmark_torch_function
-else:
-    from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
-
+if not open_source:
     torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu:sparse_ops")
 
 

fbgemm_gpu/bench/batched_unary_embeddings_benchmark.py

@@ -39,6 +39,7 @@
 
 import click
 import torch
+from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
 from torch import Tensor
 
 _HbcFn: TypeAlias = Callable[[Tensor], tuple[Tensor, Tensor]]
@@ -49,18 +50,9 @@
 try:
     # pyre-ignore[21]
     from fbgemm_gpu import open_source  # noqa: F401
-
-    open_source_fbgemm: bool = True
 except Exception:
-    open_source_fbgemm: bool = False
     torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu:sparse_ops")
 
-if open_source_fbgemm:
-    # pyre-ignore[21]
-    from bench_utils import benchmark_torch_function
-else:
-    from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
-
 
 # ──────────────────────────────────────────────────────────────────────
 # Constants (matching tritonbench port)

fbgemm_gpu/bench/histogram_binning_calibration_benchmark.py

@@ -19,6 +19,7 @@
 import click
 import fbgemm_gpu
 import torch
+from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
 from torch.profiler import profile, schedule
 
 logger: logging.Logger = logging.getLogger()
@@ -27,12 +28,7 @@
 # pyre-fixme[16]: Module `fbgemm_gpu` has no attribute `open_source`.
 open_source: bool = getattr(fbgemm_gpu, "open_source", False)
 
-if open_source:
-    # pyre-ignore[21]
-    from bench_utils import benchmark_torch_function
-else:
-    from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
-
+if not open_source:
     torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu:sparse_ops")
     torch.ops.load_library(
         "//deeplearning/fbgemm/fbgemm_gpu:permute_pooled_embedding_ops_cpu"

fbgemm_gpu/bench/jagged_tensor_benchmark.py

@@ -17,6 +17,7 @@
 import numpy as np
 import tabulate
 import torch
+from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
 from fbgemm_gpu.split_embedding_configs import SparseType
 from fbgemm_gpu.split_table_batched_embeddings_ops_inference import (
     IntNBitTableBatchedEmbeddingBagsCodegen,
@@ -33,12 +34,7 @@
 # pyre-fixme[16]: Module `fbgemm_gpu` has no attribute `open_source`.
 open_source: bool = getattr(fbgemm_gpu, "open_source", False)
 
-if open_source:
-    # pyre-ignore[21]
-    from bench_utils import benchmark_torch_function
-else:
-    from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
-
+if not open_source:
     torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu:merge_pooled_embeddings")
 
 
@@ -251,7 +247,7 @@ def print_p2p_bandwidth(
         for j in range(num_gpus):
             with torch.cuda.device(i):
                 t, _ = benchmark_torch_function(
-                    lambda: (
+                    lambda i=i, j=j: (
                         pooled_ad_embeddings[i].copy_(pooled_ad_embeddings[j])
                         if i != j
                         else pooled_ad_embeddings[i].clone()

fbgemm_gpu/bench/merge_embeddings_benchmark.py

@@ -16,6 +16,7 @@
 import fbgemm_gpu
 import hypothesis.strategies as st
 import torch
+from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
 from fbgemm_gpu.quantize_utils import fp32_to_mx4, mx4_to_fp32
 from hypothesis import given, settings
 
@@ -28,12 +29,7 @@
 # pyre-fixme[16]: Module `fbgemm_gpu` has no attribute `open_source`.
 open_source: bool = getattr(fbgemm_gpu, "open_source", False)
 
-if open_source:
-    # pyre-ignore[21]
-    from bench_utils import benchmark_torch_function
-else:
-    from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
-
+if not open_source:
     torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu:sparse_ops")
 
 
@@ -497,7 +493,11 @@ def mixdim(
         .to(torch.int)
         .cuda()
     )
-    input_refs = [torch.randn((batch_size, d)).cuda() for d in table_dims]
+    # Citrine C3: create benchmark tensors directly on the accelerator.
+    input_refs = [
+        torch.randn((batch_size, d), device=torch.accelerator.current_accelerator())
+        for d in table_dims
+    ]
     input_refs_int8 = [
         torch.ops.fbgemm.FloatToFused8BitRowwiseQuantized(t) for t in input_refs
     ]

fbgemm_gpu/bench/quantize_ops_benchmark.py

@@ -82,19 +82,15 @@
 import click
 import fbgemm_gpu
 import torch
+from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
 
 logger: logging.Logger = logging.getLogger()
 logger.setLevel(logging.INFO)
 
 # pyre-fixme[16]: Module `fbgemm_gpu` has no attribute `open_source`.
 open_source: bool = getattr(fbgemm_gpu, "open_source", False)
 
-if open_source:
-    # pyre-ignore[21]
-    from bench_utils import benchmark_torch_function
-else:
-    from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
-
+if not open_source:
     torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu:sparse_ops")
 
 

fbgemm_gpu/bench/repeat_arange_benchmark.py

@@ -20,6 +20,7 @@
 import fbgemm_gpu
 import numpy as np
 import torch
+from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
 from torch.profiler import profile, schedule
 
 logger: logging.Logger = logging.getLogger()
@@ -28,12 +29,7 @@
 # pyre-fixme[16]: Module `fbgemm_gpu` has no attribute `open_source`.
 open_source: bool = getattr(fbgemm_gpu, "open_source", False)
 
-if open_source:
-    # pyre-ignore[21]
-    from bench_utils import benchmark_torch_function
-else:
-    from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
-
+if not open_source:
     torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu:sparse_ops")
     torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu:sparse_ops_cpu")
     torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops")
@@ -945,15 +941,13 @@ def reorder_batched_ad_indices_bench(
     assert itype == "int" or itype == "long", "Only int and long are supported"
     index_type = torch.int64 if itype == "long" else torch.int32
 
+    # Citrine C3: create index tensors directly on the selected device.
     if broadcast_indices:
         cat_ad_indices = (
             torch.randint(
-                low=0,
-                high=100,
-                size=(batch_size * table_size * length,),
+                low=0, high=100, size=(batch_size * table_size * length,), device=device
             )
             .int()
-            .to(device)
             .to(data_type)
         )
         cat_ad_lengths = (
@@ -973,9 +967,9 @@ def reorder_batched_ad_indices_bench(
                 low=0,
                 high=100,
                 size=(batch_size * table_size * num_ads * length,),
+                device=device,
             )
             .int()
-            .to(device)
             .to(data_type)
         )
         cat_ad_lengths = (
@@ -990,11 +984,10 @@ def reorder_batched_ad_indices_bench(
             .to(device)
         )
 
-    batch_offsets = (
-        torch.tensor([num_ads * b for b in range(batch_size + 1)]).int()
-    ).to(
-        device
-    )  # Fixed: removed unconditional .cuda() call
+    # Citrine C3: create offsets directly on the selected device.
+    batch_offsets = torch.tensor(
+        [num_ads * b for b in range(batch_size + 1)], device=device
+    ).int()  # Fixed: removed unconditional .cuda() call
     num_ads_in_batch = batch_size * num_ads
     reordered_cat_ad_lengths = torch.ops.fbgemm.reorder_batched_ad_lengths(
         cat_ad_lengths, batch_offsets, num_ads_in_batch, broadcast_indices
@@ -1118,9 +1111,10 @@ def reorder_batched_ad_lengths_bench(
 
     # Fixed: use .to(device) directly instead of .int().cuda().to(device)
     # which unconditionally moved to CUDA before moving to the target device
-    batch_offsets = (
-        torch.tensor([num_ads * b for b in range(batch_size + 1)]).int()
-    ).to(device)
+    # Citrine C3: create offsets directly on the selected device.
+    batch_offsets = torch.tensor(
+        [num_ads * b for b in range(batch_size + 1)], device=device
+    ).int()
     num_ads_in_batch = batch_size * num_ads
 
     def _kineto_trace_handler(p: profile) -> None:
@@ -1200,11 +1194,13 @@ def reorder_batched_sequence_embeddings_bench(
         f"T={table_size}, A={num_items}, L={length}, D={dim})."
     )
 
+    # Citrine C3: create embeddings directly on the selected device.
     cat_sequence_embeddings = torch.rand(
         batch_size * table_size * num_items * length,
         dim,
         dtype=data_type,
-    ).to(device)
+        device=device,
+    )
     cat_sequence_embeddings_lengths = (
         torch.cat(
             [
@@ -1217,10 +1213,11 @@ def reorder_batched_sequence_embeddings_bench(
         .to(device)
     )
 
-    batch_offsets = (
-        torch.tensor([num_items * b for b in range(batch_size + 1)])
-        .to(index_type if device == "cpu" else torch.int32)
-        .to(device)
+    # Citrine C3: create offsets directly on the selected device.
+    batch_offsets = torch.tensor(
+        [num_items * b for b in range(batch_size + 1)],
+        dtype=index_type if device == "cpu" else torch.int32,
+        device=device,
     )
     num_items_in_batch = batch_size * num_items
     reordered_cat_sequence_embeddings_lengths = (
@@ -1321,9 +1318,14 @@ def index_select_bench(
 
     # Add optimizer to perform zero grad in order to reset gradients
     # before the accumulation phase
-    optim_index: torch.optim.Optimizer = torch.optim.SGD(inputs, lr=0.1)
-    optim_batch: torch.optim.Optimizer = torch.optim.SGD([concat_inputs], lr=0.1)
-    optim_group: torch.optim.Optimizer = torch.optim.SGD(gis_inputs, lr=0.1)
+    # Citrine C2: use the multi-tensor optimizer implementation.
+    optim_index: torch.optim.Optimizer = torch.optim.SGD(inputs, lr=0.1, foreach=True)
+    optim_batch: torch.optim.Optimizer = torch.optim.SGD(
+        [concat_inputs], lr=0.1, foreach=True
+    )
+    optim_group: torch.optim.Optimizer = torch.optim.SGD(
+        gis_inputs, lr=0.1, foreach=True
+    )
 
     def index_select_fwd_ref(
         inputs: list[torch.Tensor], indices: list[torch.Tensor]

fbgemm_gpu/bench/sparse_ops_benchmark.py

@@ -19,6 +19,7 @@
 import fbgemm_gpu
 import numpy as np
 import torch
+from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
 from fbgemm_gpu.split_embedding_configs import EmbOptimType as OptimType, SparseType
 from fbgemm_gpu.split_table_batched_embeddings_ops_inference import (
     IntNBitTableBatchedEmbeddingBagsCodegen,
@@ -43,12 +44,6 @@
 # pyre-fixme[16]: Module `fbgemm_gpu` has no attribute `open_source`.
 open_source: bool = getattr(fbgemm_gpu, "open_source", False)
 
-if open_source:
-    # pyre-ignore[21]
-    from bench_utils import benchmark_torch_function
-else:
-    from fbgemm_gpu.bench.bench_utils import benchmark_torch_function
-
 logging.basicConfig(level=logging.DEBUG)
 
 
@@ -376,8 +371,9 @@ def bounds_check_indices(  # noqa C901
             oob_positions = torch.randperm(num_indices)[:num_oob]
             req.indices[oob_positions] = E
 
-    warning = torch.tensor([0]).long().to(get_device())
-    rows_per_table = torch.tensor([E for _ in range(T)]).long().to(get_device())
+    # Citrine C3: create bounds-check tensors directly on the target device.
+    warning = torch.tensor([0], device=get_device()).long()
+    rows_per_table = torch.tensor([E for _ in range(T)], device=get_device()).long()
 
     bc_mode = BoundsCheckMode(bounds_check_mode)
     bounds_check_version = 1