Add ufmt linter for pyproject (#47)

Summary:
Add linter to make sure PR is consistent with internal `arc lint`.

To format the code, run the following in the repo directory:

```
ufmt format .
```

We also test 25 operators in our H100 CI. Note that many of the flash_attention backends do not work right now and we need to fix them.

Pull Request resolved: #47

Reviewed By: FindHao

Differential Revision: D65709181

Pulled By: xuzhao9

fbshipit-source-id: 5b013906e7b04c8ee41d74db5756de08eec5b5b2

.ci/tritonbench/test-gpu.sh

@@ -8,4 +8,7 @@ fi
 
 . "${SETUP_SCRIPT}"
 
+# install deps
+pip install psutil tabulate
+
 python -m unittest test.test_gpu.main

.github/workflows/linter.yaml

@@ -0,0 +1,29 @@
+name: Linter
+on:
+  pull_request:
+  push:
+    branches:
+      - main
+  workflow_dispatch:
+
+jobs:
+  pylint:
+    permissions:
+      contents: read
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v3
+        with:
+          path: tritonbench
+      - name: Install deps
+        run: |
+          pip install ruff-api==0.1.0
+      - name: Check Formatting
+        uses: omnilib/ufmt@action-v1
+        with:
+          path: tritonbench
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true

.github/workflows/pr.yaml

@@ -5,6 +5,9 @@ on:
       - .ci/*
       - tritonbench/*
       - .github/workflows/pr.yaml
+  push:
+    branches:
+      - main
 
 jobs:
   h100-pytorch-test:
@@ -27,6 +30,10 @@ jobs:
           sudo nvidia-smi -pm 1
           sudo ldconfig
           nvidia-smi
-      - name: Test Tritonbench operators
+      - name: Test Tritonbench operators on H100 GPU
         run: |
-          bash ./.ci/tritonbench/test-operators.sh
+          bash ./.ci/tritonbench/test-gpu.sh
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
+  cancel-in-progress: true

pyproject.toml

@@ -0,0 +1,7 @@
+[tool.ufmt]
+formatter = "ruff-api"
+excludes = ["submodules/"]
+
+[tool.black]
+line-length = 88
+target-version = ["py312"]

requirements.txt

@@ -1,3 +1,5 @@
 packaging
 pynvml
+psutil
+tabulate
 transformers==4.46.1

test/test_cpu/main.py

@@ -7,7 +7,6 @@
 
 
 class TestTritonbenchCpu(unittest.TestCase):
-
     def _get_test_op(self):
         parser = get_parser(["--device", "cpu", "--op", "test_op"])
         tb_args, extra_args = parser.parse_known_args(

test/test_gpu/main.py

@@ -18,7 +18,11 @@
     fbcode_skip_file_path = "fb/skip_tests_h100_fbcode.yaml"
     SKIP_FILE = importlib.resources.files(__package__).joinpath(fbcode_skip_file_path)
 else:
-    SKIP_FILE = "skip_tests_h100_pytorch.yaml"
+    import os
+
+    SKIP_FILE = os.path.abspath(
+        os.path.join(os.path.dirname(__file__), "skip_tests_h100_pytorch.yaml")
+    )
 
 with open(SKIP_FILE, "r") as f:
     skip_tests = yaml.safe_load(f)
@@ -55,7 +59,7 @@ def _run_one_operator(
 ):
     if tb_args.op in skip_tests:
         # If the op itself is in the skip list, skip all tests
-        if skip_tests[tb_args.op] is None:
+        if not skip_tests[tb_args.op]:
             return
         tb_args.skip = ",".join(skip_tests[tb_args.op])
     Operator = load_opbench_by_name(tb_args.op)

test/test_gpu/skip_tests_h100_pytorch.yaml

@@ -1,11 +1,37 @@
 # Tests we skip in OSS CI
 # This file is regarding to the Triton version bundled with pytorch
-# Use <op-name> to skip an entire operator
-# Use <op-name/impl-name> to skip an impl
-- test_op
-- bf16xint16_gemm/bf16xint16
-- fp8_attention/colfax_fmha
-- fp8_fused_quant_gemm_rowwise
-- fp8_gemm/triton_persistent_fp8_gemm
-- fp8_gemm/triton_tma_persistent_fp8_gemm
-- fp8_gemm_rowwise
+# Use <op-name:> to skip an entire operator
+# Use <op-name:\n  - impl-name> to skip an impl
+bf16xint16_gemm:
+  - bf16xint16
+# TODO: we have many buggy backends for flash_attention
+# Need to fix them in the CI
+flash_attention:
+#   - triton_tutorial_flash_v2_tma
+#   - triton_op_flash_v2
+#   - xformers_splitk
+#   - colfax_cutlass
+#   - tk
+#   - sdpa
+#   - cudnn
+#   - flex_attention
+fp8_attention:
+  - colfax_fmha
+fp8_fused_quant_gemm_rowwise:
+fp8_gemm:
+  - triton_persistent_fp8_gemm
+  - triton_tma_persistent_fp8_gemm
+fp8_gemm_rowwise:
+gemm:
+grouped_gemm:
+int4_gemm:
+jagged_layer_norm:
+jagged_mean:
+jagged_softmax:
+jagged_sum:
+layer_norm:
+low_mem_dropout:
+rms_norm:
+rope:
+template_attention:
+test_op:

tools/cuda_utils.py

@@ -238,7 +238,5 @@ def check_torch_nightly_version(force_date: Optional[str] = None):
     if args.install_torch_nightly:
         install_pytorch_nightly(cuda_version=args.cudaver, env=os.environ)
     if args.check_torch_nightly_version:
-        assert (
-            not args.install_torch_nightly
-        ), "Error: Can't run install torch nightly and check version in the same command."
+        assert not args.install_torch_nightly, "Error: Can't run install torch nightly and check version in the same command."
         check_torch_nightly_version(args.force_date)

tritonbench/components/workers/subprocess_rpc.py

@@ -274,8 +274,10 @@ def write(self, msg: bytes) -> None:
     def get_writer_pid(self) -> int:
         assert (
             self._writer_pid is not None
-        ), "Writer pid is not specified. Maybe calling from child process or input pipe.\
+        ), (
+            "Writer pid is not specified. Maybe calling from child process or input pipe.\
                                               Please report a bug."
+        )
         return self._writer_pid
 
     def set_writer_pid(self, writer_pid: int) -> None:

tritonbench/kernels/triton_fused_attention.py

@@ -37,7 +37,6 @@
 
 
 class TmaAutoTuneHelper:
-
     # duck typing wrapper to implement the same interface as TmaDescKernelParam in Triton PR #4498
     class KernelParamWrapper:
         def __init__(self, desc):
@@ -734,7 +733,6 @@ def _attn_fwd_tma(  # Q, V, desc_k, desc_v, sm_scale, M, Out,  #
     HEAD_DIM: tl.constexpr,  #
     STAGE: tl.constexpr,  #
 ):
-
     tl.static_assert(BLOCK_N <= HEAD_DIM)
     start_m = tl.program_id(0)
     off_hz = tl.program_id(1)
@@ -848,7 +846,14 @@ def _attn_fwd_tma(  # Q, V, desc_k, desc_v, sm_scale, M, Out,  #
 
 @triton.jit
 def _attn_bwd_preprocess(
-    O, DO, Delta, Z, H, N_CTX, BLOCK_M: tl.constexpr, HEAD_DIM: tl.constexpr  #  #  #  #
+    O,
+    DO,
+    Delta,
+    Z,
+    H,
+    N_CTX,
+    BLOCK_M: tl.constexpr,
+    HEAD_DIM: tl.constexpr,  #  #  #  #
 ):
     off_m = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
     off_hz = tl.program_id(1)
@@ -1179,7 +1184,6 @@ def _attn_bwd(
 
 
 class _attention_ws(torch.autograd.Function):
-
     @staticmethod
     def forward(ctx, q, k, v, causal, sm_scale):
         # shape constraints
@@ -1232,7 +1236,7 @@ def forward(ctx, q, k, v, causal, sm_scale):
             N_CTX=q.shape[2],  #
             HEAD_DIM=HEAD_DIM_K,  #
             STAGE=stage,  #
-            **extra_kern_args
+            **extra_kern_args,
         )
 
         ctx.save_for_backward(q, k, v, o, M)
@@ -1304,7 +1308,6 @@ def backward(ctx, do):
 
 
 class _attention(torch.autograd.Function):
-
     @staticmethod
     def forward(ctx, q, k, v, causal, sm_scale):
         # shape constraints
@@ -1355,7 +1358,7 @@ def forward(ctx, q, k, v, causal, sm_scale):
             N_CTX=q.shape[2],  #
             HEAD_DIM=HEAD_DIM_K,  #
             STAGE=stage,  #
-            **extra_kern_args
+            **extra_kern_args,
         )
 
         ctx.save_for_backward(q, k, v, o, M)
@@ -1427,7 +1430,6 @@ def backward(ctx, do):
 
 
 class _attention_tma(torch.autograd.Function):
-
     @staticmethod
     def forward(ctx, q, k, v, causal, sm_scale):
         # shape constraints
@@ -1587,7 +1589,7 @@ def grid_tma(META):
             N_CTX=q.shape[2],  #
             HEAD_DIM=HEAD_DIM_K,  #
             STAGE=stage,  #
-            **extra_kern_args
+            **extra_kern_args,
         )
 
         ctx.save_for_backward(q, k, v, o, M)

tritonbench/operators/gather_gemv/operator.py

@@ -26,7 +26,6 @@
 
 
 class Operator(BenchmarkOperator):
-
     @register_metric()
     def gbps(self, fn_name, example_inputs, metrics: BenchmarkOperatorMetrics):
         arg0_1, arg1_1, arg2_1 = example_inputs

tritonbench/operators/gather_gemv/triton_gather_gemv.py

@@ -82,7 +82,7 @@ def triton_red_fused_mv_0(
     rbase = tl.arange(0, RBLOCK)[None, :].to(tl.int64)
     x0 = xindex
     # x0 // rnumel should have the same value of either 0 or 1
-    tmp0 = tl.load(in_ptr0 + ((x0 // rnumel)), None, eviction_policy="evict_last")
+    tmp0 = tl.load(in_ptr0 + (x0 // rnumel), None, eviction_policy="evict_last")
     _tmp11 = tl.full([XBLOCK, RBLOCK], 0, tl.float32)
     for roffset in range(0, rnumel, RBLOCK):
         rindex = roffset + rbase

tritonbench/operators/jagged_layer_norm/operator.py

@@ -36,7 +36,6 @@ def parse_op_args(args: List[str]):
 
 
 class Operator(BenchmarkOperator):
-
     DEFAULT_METRICS = ["latency", "accuracy"]
     DEFAULT_PRECISION = "fp32"
 
@@ -48,8 +47,8 @@ def __init__(
         self, tb_args: argparse.Namespace, extra_args: Optional[List[str]] = None
     ):
         super().__init__(tb_args, extra_args)
-        self.sizes = list(range(2, 12, 4)) + list(
-            range(12, 23, 3)
+        self.sizes = (
+            list(range(2, 12, 4)) + list(range(12, 23, 3))
         )  # bias towards larger sizes, which are more representative of real-world shapes
 
         args = parse_op_args(self.extra_args)
@@ -105,8 +104,8 @@ def _inner():
             )  # a sum over (1, 2) ensures layer norm, whereas a sum over (1) would be an instance norm
 
             padded_normalized = (
-                padded_values - mean
-            ) * padded_mask_values  # mask elements outside of the ragged dimension size for correct variance calculation
+                (padded_values - mean) * padded_mask_values
+            )  # mask elements outside of the ragged dimension size for correct variance calculation
 
             variance = (
                 torch.sum(

tritonbench/operators/jagged_mean/kernels.py

@@ -53,8 +53,9 @@ def triton_jagged_mean_kernel_simple_fused_sum_then_buffer(
     offsets_m = block_start_m + tl.arange(0, BLOCK_SIZE_M)
     mask_m = offsets_m < M
 
-    ragged_start, ragged_end = tl.load(input_ptr_offsets + pid_b), tl.load(
-        input_ptr_offsets + (pid_b + 1)
+    ragged_start, ragged_end = (
+        tl.load(input_ptr_offsets + pid_b),
+        tl.load(input_ptr_offsets + (pid_b + 1)),
     )  # load start and end offsets for current program, similar to offsets[i] and offsets[i + 1]
     ragged_len = ragged_end - ragged_start
 
@@ -133,8 +134,9 @@ def triton_jagged_mean_kernel_simple_fused_buffer_then_sum(
     offsets_m = block_start_m + tl.arange(0, BLOCK_SIZE_M)
     mask_m = offsets_m < M
 
-    ragged_start, ragged_end = tl.load(input_ptr_offsets + pid_b), tl.load(
-        input_ptr_offsets + (pid_b + 1)
+    ragged_start, ragged_end = (
+        tl.load(input_ptr_offsets + pid_b),
+        tl.load(input_ptr_offsets + (pid_b + 1)),
     )  # load start and end offsets for current program, similar to offsets[i] and offsets[i + 1]
     ragged_len = ragged_end - ragged_start
 
@@ -212,8 +214,9 @@ def triton_jagged_mean_kernel_variable_length_loop_sum_then_buffer(
     offsets_m = block_start_m + tl.arange(0, BLOCK_SIZE_M)
     mask_m = offsets_m < M
 
-    ragged_start, ragged_end = tl.load(input_ptr_offsets + pid_b), tl.load(
-        input_ptr_offsets + (pid_b + 1)
+    ragged_start, ragged_end = (
+        tl.load(input_ptr_offsets + pid_b),
+        tl.load(input_ptr_offsets + (pid_b + 1)),
     )  # load start and end offsets for current program, similar to offsets[i] and offsets[i + 1]
     ragged_len = ragged_end - ragged_start
 
@@ -288,8 +291,9 @@ def triton_jagged_mean_kernel_variable_length_loop_buffer_then_sum(
     offsets_m = block_start_m + tl.arange(0, BLOCK_SIZE_M)
     mask_m = offsets_m < M
 
-    ragged_start, ragged_end = tl.load(input_ptr_offsets + pid_ragged), tl.load(
-        input_ptr_offsets + (pid_ragged + 1)
+    ragged_start, ragged_end = (
+        tl.load(input_ptr_offsets + pid_ragged),
+        tl.load(input_ptr_offsets + (pid_ragged + 1)),
     )  # load start and end offsets for current program, similar to offsets[i] and offsets[i + 1]
     ragged_len = ragged_end - ragged_start