[fbgemm_gpu] Gate OSS CI workflows

- Add conditionals to prevent pytorch-infra-based OSS CI from running when repo is hosted in a non-pytorch organization

.github/workflows/build_wheels_linux_aarch64.yml

@@ -24,6 +24,7 @@ permissions:
 
 jobs:
   generate-matrix:
+    if: ${{ github.repository_owner == 'pytorch' }}
     uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@main
     with:
       package-type: wheel
@@ -32,6 +33,7 @@ jobs:
       test-infra-ref: main
       with-cuda: disable
   build:
+    if: ${{ github.repository_owner == 'pytorch' }}
     needs: generate-matrix
     strategy:
       fail-fast: false

.github/workflows/build_wheels_linux_x86.yml

@@ -24,6 +24,7 @@ permissions:
 
 jobs:
   generate-matrix:
+    if: ${{ github.repository_owner == 'pytorch' }}
     uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@main
     with:
       package-type: wheel
@@ -34,6 +35,7 @@ jobs:
       with-rocm: enable
       with-cpu: enable
   build:
+    if: ${{ github.repository_owner == 'pytorch' }}
     needs: generate-matrix
     name: pytorch/FBGEMM
     uses: pytorch/test-infra/.github/workflows/build_wheels_linux.yml@main

.github/workflows/fbgemm_ci.yml

@@ -20,7 +20,7 @@ concurrency:
 
 jobs:
   build-linux:
-    runs-on: ${{ matrix.host-machine.instance }}
+    runs-on: ${{ github.repository_owner == 'pytorch' && matrix.host-machine.instance || 'ubuntu-latest' }}
     container:
       image: amazonlinux:2023
       options: --user root
@@ -105,7 +105,7 @@ jobs:
 
 
   build-bazel:
-    runs-on: linux.12xlarge
+    runs-on: ${{ github.repository_owner == 'pytorch' && matrix.host-machine.instance || 'ubuntu-latest' }}
     container:
       image: amazonlinux:2023
       options: --user root

.github/workflows/fbgemm_gpu_ci_cpu.yml

@@ -47,6 +47,7 @@ concurrency:
 jobs:
   # Build on CPU hosts, run tests, and upload to GHA
   build_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: amazonlinux:2023
@@ -118,6 +119,7 @@ jobs:
 
   # Download the built artifact from GHA, test on GPU, and push to PyPI
   test_and_publish_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: amazonlinux:2023

.github/workflows/fbgemm_gpu_ci_cuda.yml

@@ -46,6 +46,7 @@ concurrency:
 jobs:
   # Build on CPU hosts and upload to GHA
   build_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: amazonlinux:2023
@@ -127,6 +128,7 @@ jobs:
 
   # Download the built artifact from GHA, test on GPU, and push to PyPI
   test_and_publish_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     # runs-on: linux.4xlarge.nvidia.gpu
     # Use available instance types - https://github.com/pytorch/test-infra/blob/main/.github/scale-config.yml
     runs-on: ${{ matrix.host-machine.instance }}

.github/workflows/fbgemm_gpu_ci_genai.yml

@@ -46,6 +46,7 @@ concurrency:
 jobs:
   # Build on CPU hosts and upload to GHA
   build_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: amazonlinux:2023
@@ -127,6 +128,7 @@ jobs:
 
   # Download the built artifact from GHA, test on GPU, and push to PyPI
   test_and_publish_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     # Use available instance types - https://github.com/pytorch/test-infra/blob/main/.github/scale-config.yml
     runs-on: ${{ matrix.host-machine.instance }}
     defaults:

.github/workflows/fbgemm_gpu_ci_genai_generic_infra.yml

@@ -32,6 +32,7 @@ concurrency:
 jobs:
   # Build on CPU hosts and upload to GHA
   build_artifact:
+    if: ${{ github.repository_owner != 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: amazonlinux:2023
@@ -116,6 +117,7 @@ jobs:
 
   # Download the built artifact from GHA, test on GPU, and push to PyPI
   test_artifact:
+    if: ${{ github.repository_owner != 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: amazonlinux:2023

.github/workflows/fbgemm_gpu_ci_rocm.yml

@@ -46,6 +46,7 @@ concurrency:
 jobs:
   # Build on CPU hosts and upload to GHA
   build_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: ${{ matrix.container-image }}
@@ -125,6 +126,7 @@ jobs:
 
   # Download the built artifact from GHA, test on GPU, and push to PyPI
   test_and_publish_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: "rocm/dev-ubuntu-20.04:${{ matrix.rocm-version }}-complete"

.github/workflows/fbgemm_gpu_docs.yml

@@ -24,6 +24,7 @@ on:
 
 jobs:
   build-docs:
+    if: ${{ github.repository_owner == 'pytorch' }}
     permissions:
       # Grant write permission here so that the generated docs can be pushed to `gh-pages` branch
       contents: write

.github/workflows/fbgemm_gpu_lint.yml

@@ -35,7 +35,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: [ "3.11" ]
+        python-version: [ "3.12" ]
 
     steps:
     - name: Checkout the Repository

.github/workflows/fbgemm_gpu_pip.yml

@@ -45,7 +45,7 @@ on:
 
 jobs:
   test_pypi_install_cpu:
-    if: ${{ github.event_name == 'schedule' || (github.event_name == 'workflow_dispatch' && github.event.inputs.fbgemm_gpu_variant_type == 'cpu') }}
+    if: ${{ github.repository_owner == 'pytorch' && (github.event_name == 'schedule' || (github.event_name == 'workflow_dispatch' && github.event.inputs.fbgemm_gpu_variant_type == 'cpu')) }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: amazonlinux:2023
@@ -104,7 +104,7 @@ jobs:
 
 
   test_pypi_install_cuda:
-    if: ${{ github.event_name == 'schedule' || (github.event_name == 'workflow_dispatch' && github.event.inputs.fbgemm_gpu_variant_type == 'cuda') }}
+    if: ${{ github.repository_owner == 'pytorch' && (github.event_name == 'schedule' || (github.event_name == 'workflow_dispatch' && github.event.inputs.fbgemm_gpu_variant_type == 'cuda') }}
     runs-on: ${{ matrix.host-machine.instance }}
     defaults:
       run:
@@ -165,7 +165,7 @@ jobs:
 
 
   test_pypi_install_rocm:
-    if: ${{ github.event_name == 'schedule' || (github.event_name == 'workflow_dispatch' && github.event.inputs.fbgemm_gpu_variant_type == 'rocm') }}
+    if: ${{ github.repository_owner == 'pytorch' && (github.event_name == 'schedule' || (github.event_name == 'workflow_dispatch' && github.event.inputs.fbgemm_gpu_variant_type == 'rocm') }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: "rocm/dev-ubuntu-20.04:${{ matrix.rocm-version }}-complete"

.github/workflows/fbgemm_gpu_release_cpu.yml

@@ -44,6 +44,7 @@ concurrency:
 jobs:
   # Build on CPU hosts, run tests, and upload to GHA
   build_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: amazonlinux:2023
@@ -114,6 +115,7 @@ jobs:
 
   # Download the built artifact from GHA, test on GPU, and push to PyPI
   test_and_publish_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: amazonlinux:2023

.github/workflows/fbgemm_gpu_release_cuda.yml

@@ -50,6 +50,7 @@ concurrency:
 jobs:
   # Build on CPU hosts and upload to GHA
   build_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: amazonlinux:2023
@@ -126,6 +127,7 @@ jobs:
 
   # Download the built artifact from GHA, test on GPU, and push to PyPI
   test_and_publish_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     defaults:
       run:

.github/workflows/fbgemm_gpu_release_genai.yml

@@ -50,6 +50,7 @@ concurrency:
 jobs:
   # Build on CPU hosts and upload to GHA
   build_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     container:
       image: amazonlinux:2023
@@ -126,6 +127,7 @@ jobs:
 
   # Download the built artifact from GHA, test on GPU, and push to PyPI
   test_and_publish_artifact:
+    if: ${{ github.repository_owner == 'pytorch' }}
     runs-on: ${{ matrix.host-machine.instance }}
     defaults:
       run:

fbgemm_gpu/bench/quantize_ops_benchmark.py

@@ -469,7 +469,7 @@ def mixdim(
     )  # output is FP32
 
     print(
-        f"Input tensor batch_size: {batch_size}, num_tables: {num_tables}, tensor_size: {input_data.numel() / (1 << 30)} GB, average table dimension: {sum(table_dims) * 1.0/num_tables}."
+        f"Input tensor batch_size: {batch_size}, num_tables: {num_tables}, tensor_size: {input_data.numel() / (1 << 30)} GB, average table dimension: {sum(table_dims) * 1.0 / num_tables}."
     )
     print(
         f"Mixed dim dequantize average time per iter FP32: {average_time_mixed_dim_fp32} s, bandwidth : {input_data.numel() / (1 << 30) / average_time_mixed_dim_fp32} GB/s."

fbgemm_gpu/bench/split_table_batched_embeddings_benchmark.py

@@ -835,7 +835,7 @@ def cache(  # noqa C901
     param_size_multiplier = weights_precision.bit_rate() / 8.0
     logging.info(
         f"Embedding tables: {E * T} rows, {nparams / 1.0e9: .2f} GParam, "
-        f"{nparams * param_size_multiplier  / 1.0e9: .2f} GB"
+        f"{nparams * param_size_multiplier / 1.0e9: .2f} GB"
     )
     logging.info(
         f"Accessed weights per batch: {B * T * L} rows, "
@@ -889,11 +889,11 @@ def cache(  # noqa C901
         cache_misses.append((emb.lxu_cache_locations_list[0] == NOT_FOUND).sum().item())
         emb.forward(indices.long(), offsets.long())
     logging.info(
-        f"Exchanged cache lines -- mean: {sum(exchanged_cache_lines)/len(requests): .2f}, "
+        f"Exchanged cache lines -- mean: {sum(exchanged_cache_lines) / len(requests): .2f}, "
         f"max: {max(exchanged_cache_lines)}, min: {min(exchanged_cache_lines)}"
     )
     logging.info(
-        f"Cache miss -- mean: {sum(cache_misses)/len(requests)}, "
+        f"Cache miss -- mean: {sum(cache_misses) / len(requests)}, "
         f"max: {max(cache_misses)}, min: {min(cache_misses)}"
     )
 
@@ -2386,24 +2386,24 @@ def nbit_cache(  # noqa C901
         input_indices.append(len(indices))
 
     logging.info(
-        f"Exchanged cache lines -- mean: {sum(exchanged_cache_lines)/len(requests): .2f}, "
+        f"Exchanged cache lines -- mean: {sum(exchanged_cache_lines) / len(requests): .2f}, "
         f"max: {max(exchanged_cache_lines)}, min: {min(exchanged_cache_lines)}"
     )
     logging.info(
-        f"Cache miss -- mean: {sum(cache_misses)/len(requests)}, "
+        f"Cache miss -- mean: {sum(cache_misses) / len(requests)}, "
         f"max: {max(cache_misses)}, min: {min(cache_misses)}"
     )
     logging.info(
-        f"input_indices -- mean: {sum(input_indices)/len(requests)}, "
+        f"input_indices -- mean: {sum(input_indices) / len(requests)}, "
         f"max: {max(input_indices)}, min: {min(input_indices)}"
     )
     logging.info(
-        f"unique_indices -- mean: {sum(unique_indices)/len(requests)}, "
+        f"unique_indices -- mean: {sum(unique_indices) / len(requests)}, "
         f"max: {max(unique_indices)}, min: {min(unique_indices)}"
     )
     unique_miss_rate = [a / b for (a, b) in zip(exchanged_cache_lines, unique_indices)]
     logging.info(
-        f"unique_miss_rate -- mean: {sum(unique_miss_rate)/len(requests)}, "
+        f"unique_miss_rate -- mean: {sum(unique_miss_rate) / len(requests)}, "
         f"max: {max(unique_miss_rate)}, min: {min(unique_miss_rate)}"
     )
     if record_cache_miss_counter or record_tablewise_cache_miss:

fbgemm_gpu/bench/ssd_table_batched_embeddings_benchmark.py

@@ -149,7 +149,7 @@ def benchmark_read_write(
         gibps_wr = byte_seconds_per_ns / (write_lat_ns * 2**30)
         gibps_tot = 2 * byte_seconds_per_ns / ((read_lat_ns + write_lat_ns) * 2**30)
         logging.info(
-            f"Total bytes: {total_bytes/1e9:0.2f} GB, "
+            f"Total bytes: {total_bytes / 1e9:0.2f} GB, "
             f"Read_us: {read_lat_ns / 1000:8.0f}, "
             f"Write_us: {write_lat_ns / 1000:8.0f}, "
             f"Total_us: {(read_lat_ns + write_lat_ns) / 1000:8.0f}, "
@@ -389,7 +389,7 @@ def gen_split_tbe_generator(
             + param_size_multiplier * B * sum(Ds) * L
         )
 
-    logging.info(f"Batch read write bytes: {read_write_bytes/1.0e9: .2f} GB")
+    logging.info(f"Batch read write bytes: {read_write_bytes / 1.0e9: .2f} GB")
 
     # Compute width of test name and bandwidth widths to improve report
     # readability

fbgemm_gpu/fbgemm_gpu/split_table_batched_embeddings_ops_inference.py

@@ -531,10 +531,10 @@ def print_cache_miss_counter(self) -> None:
             f"Miss counter value [3] - # of total requested indices : {self.cache_miss_counter[3]}, "
         )
         logging.info(
-            f"unique_miss_rate using counter : {self.cache_miss_counter[1]/self.cache_miss_counter[2]}, \n"
+            f"unique_miss_rate using counter : {self.cache_miss_counter[1] / self.cache_miss_counter[2]}, \n"
         )
         logging.info(
-            f"total_miss_rate using counter : {self.cache_miss_counter[1]/self.cache_miss_counter[3]}, \n"
+            f"total_miss_rate using counter : {self.cache_miss_counter[1] / self.cache_miss_counter[3]}, \n"
         )
 
     def get_uvm_cache_stats(self) -> Tensor:
@@ -558,8 +558,8 @@ def print_uvm_cache_stats(self) -> None:
         )
         if uvm_cache_stats[1]:
             logging.info(
-                f"unique indices / requested indices: {uvm_cache_stats[2]/uvm_cache_stats[1]}\n"
-                f"unique misses / requested indices: {uvm_cache_stats[3]/uvm_cache_stats[1]}\n"
+                f"unique indices / requested indices: {uvm_cache_stats[2] / uvm_cache_stats[1]}\n"
+                f"unique misses / requested indices: {uvm_cache_stats[3] / uvm_cache_stats[1]}\n"
             )
 
     @torch.jit.export

fbgemm_gpu/test/quantize/fp8_rowwise_test.py

@@ -225,9 +225,9 @@ def test_quantize_and_dequantize_op_padded_fp8_rowwise(
             logging.info(f"qref {torch.gather(qref, dim=1, index=idx)}")
             logging.info(f"dqcat {torch.gather(dqcat, dim=1, index=idx)}")
             logging.info(
-                f"relative error: max: {errors.abs().max()*100:.1f}%, "
-                f"median: {errors.abs().median()*100:.1f}%, "
-                f"mean: {errors.abs().mean()*100:.1f}%"
+                f"relative error: max: {errors.abs().max() * 100:.1f}%, "
+                f"median: {errors.abs().median() * 100:.1f}%, "
+                f"mean: {errors.abs().mean() * 100:.1f}%"
             )
 
         torch.testing.assert_allclose(dqcat, qref, rtol=0.1, atol=0.05)