Update Optional[x] -> x | None and Union[x, y] to x | y (#26633)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

benchmarks/kernels/bench_per_token_quant_fp8.py

@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import itertools
-from typing import Callable
+from collections.abc import Callable
 from unittest.mock import patch
 
 import pandas as pd

benchmarks/kernels/benchmark_device_communicators.py

@@ -22,8 +22,8 @@ Example:
 import json
 import os
 import time
+from collections.abc import Callable
 from contextlib import nullcontext
-from typing import Callable, Optional
 
 import torch
 import torch.distributed as dist
@@ -264,12 +264,12 @@ class CommunicatorBenchmark:
     def benchmark_allreduce_single(
         self,
         sequence_length: int,
-        allreduce_fn: Callable[[torch.Tensor], Optional[torch.Tensor]],
+        allreduce_fn: Callable[[torch.Tensor], torch.Tensor | None],
         should_use_fn: Callable[[torch.Tensor], bool],
         context,
         num_warmup: int,
         num_trials: int,
-    ) -> Optional[float]:
+    ) -> float | None:
         """Benchmark method with CUDA graph optimization."""
         try:
             # Create test tensor (2D: sequence_length x hidden_size)

benchmarks/kernels/benchmark_lora.py

@@ -6,11 +6,12 @@ import copy
 import json
 import pickle
 import time
+from collections.abc import Callable
 from dataclasses import dataclass
 from enum import Enum, auto
 from itertools import product
 from pathlib import Path
-from typing import Any, Callable, Optional
+from typing import Any
 
 import torch
 import torch.utils.benchmark as TBenchmark
@@ -158,7 +159,7 @@ def ref_group_gemm(
     seq_lens_cpu: torch.Tensor,
     prompt_lora_mapping_cpu: torch.Tensor,
     scaling: float,
-    add_inputs: Optional[bool],
+    add_inputs: bool | None,
 ):
     """
     Torch group gemm reference implementation to test correctness of
@@ -316,8 +317,8 @@ class BenchmarkContext:
     lora_rank: int
     sort_by_lora_id: bool
     dtype: torch.dtype
-    seq_length: Optional[int] = None
-    num_slices: Optional[int] = None  # num_slices for slice based ops
+    seq_length: int | None = None
+    num_slices: int | None = None  # num_slices for slice based ops
 
     def with_seq_length(self, seq_length: int) -> "BenchmarkContext":
         ctx = copy.copy(self)
@@ -561,7 +562,7 @@ class BenchmarkTensors:
         }
 
     def bench_fn_kwargs(
-        self, op_type: OpType, add_inputs: Optional[bool] = None
+        self, op_type: OpType, add_inputs: bool | None = None
     ) -> dict[str, Any]:
         if op_type.is_shrink_fn():
             assert add_inputs is None
@@ -575,7 +576,7 @@ class BenchmarkTensors:
         raise ValueError(f"Unrecognized optype {self}")
 
     def test_correctness(
-        self, op_type: OpType, expand_fn_add_inputs: Optional[bool]
+        self, op_type: OpType, expand_fn_add_inputs: bool | None
     ) -> bool:
         """
         Test correctness of op_type implementation against a grouped gemm
@@ -611,8 +612,8 @@ def bench_optype(
     ctx: BenchmarkContext,
     arg_pool_size: int,
     op_type: OpType,
-    cuda_graph_nops: Optional[int] = None,
-    expand_fn_add_inputs: Optional[bool] = None,
+    cuda_graph_nops: int | None = None,
+    expand_fn_add_inputs: bool | None = None,
     test_correctness: bool = False,
 ) -> TMeasurement:
     assert arg_pool_size >= 1
@@ -679,7 +680,7 @@ def bench_torch_mm(
     ctx: BenchmarkContext,
     arg_pool_size: int,
     op_type: OpType,
-    cuda_graph_nops: Optional[int] = None,
+    cuda_graph_nops: int | None = None,
 ) -> TMeasurement:
     """
     Benchmark basic torch.mm as a roofline.
@@ -744,7 +745,7 @@ def use_cuda_graph_recommendation() -> str:
             """
 
 
-def print_timers(timers: list[TMeasurement], args: Optional[argparse.Namespace] = None):
+def print_timers(timers: list[TMeasurement], args: argparse.Namespace | None = None):
     compare = TBenchmark.Compare(timers)
     compare.print()
 

benchmarks/kernels/benchmark_machete.py

@@ -8,10 +8,9 @@ import math
 import os
 import pickle as pkl
 import time
-from collections.abc import Iterable
+from collections.abc import Callable, Iterable
 from dataclasses import dataclass
 from itertools import product
-from typing import Callable, Optional
 
 import pandas as pd
 import torch
@@ -63,23 +62,23 @@ class BenchmarkTensors:
     a: torch.Tensor
 
     w_q: torch.Tensor
-    group_size: Optional[int]
+    group_size: int | None
     wtype: ScalarType
     w_g_s: torch.Tensor
-    w_g_zp: Optional[torch.Tensor]
-    w_ch_s: Optional[torch.Tensor]
-    w_tok_s: Optional[torch.Tensor]
+    w_g_zp: torch.Tensor | None
+    w_ch_s: torch.Tensor | None
+    w_tok_s: torch.Tensor | None
 
 
 @dataclass
 class TypeConfig:
     act_type: torch.dtype
     weight_type: ScalarType
-    output_type: Optional[torch.dtype]
-    group_scale_type: Optional[torch.dtype]
-    group_zero_type: Optional[torch.dtype]
-    channel_scale_type: Optional[torch.dtype]
-    token_scale_type: Optional[torch.dtype]
+    output_type: torch.dtype | None
+    group_scale_type: torch.dtype | None
+    group_zero_type: torch.dtype | None
+    channel_scale_type: torch.dtype | None
+    token_scale_type: torch.dtype | None
 
 
 def rand_data(shape, dtype=torch.float16, scale=1):
@@ -93,8 +92,8 @@ def quantize_and_pack(
     atype: torch.dtype,
     w: torch.Tensor,
     wtype: ScalarType,
-    stype: Optional[torch.dtype],
-    group_size: Optional[int],
+    stype: torch.dtype | None,
+    group_size: int | None,
     zero_points: bool = False,
 ):
     assert wtype.is_integer(), "TODO: support floating point weights"
@@ -113,7 +112,7 @@ def quantize_and_pack(
 
 
 def create_bench_tensors(
-    shape: tuple[int, int, int], types: TypeConfig, group_size: Optional[int]
+    shape: tuple[int, int, int], types: TypeConfig, group_size: int | None
 ) -> list[BenchmarkTensors]:
     m, n, k = shape
 
@@ -331,8 +330,8 @@ def bench_fns(label: str, sub_label: str, description: str, fns: list[Callable])
     return res
 
 
-_SWEEP_SCHEDULES_RESULTS: Optional[pd.DataFrame] = None
-_SWEEP_SCHEDULES_RESULTS_CSV: Optional[str] = None
+_SWEEP_SCHEDULES_RESULTS: pd.DataFrame | None = None
+_SWEEP_SCHEDULES_RESULTS_CSV: str | None = None
 
 
 def bench(

benchmarks/kernels/benchmark_paged_attention.py

@@ -3,7 +3,6 @@
 
 import random
 import time
-from typing import Optional
 
 import torch
 
@@ -37,7 +36,7 @@ def main(
     seed: int,
     do_profile: bool,
     device: str = "cuda",
-    kv_cache_dtype: Optional[str] = None,
+    kv_cache_dtype: str | None = None,
 ) -> None:
     current_platform.seed_everything(seed)
 

benchmarks/kernels/benchmark_per_token_group_quant.py

@@ -3,8 +3,8 @@
 
 import argparse
 import math
+from collections.abc import Callable
 from contextlib import contextmanager
-from typing import Callable
 from unittest.mock import patch
 
 import torch

benchmarks/kernels/benchmark_reshape_and_cache.py

@@ -1,7 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from __future__ import annotations
-
 import random
 import time
 

benchmarks/kernels/benchmark_reshape_and_cache_flash.py

@@ -1,7 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from __future__ import annotations
-
 import random
 import time
 

benchmarks/kernels/benchmark_rmsnorm.py

@@ -2,7 +2,6 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import itertools
-from typing import Optional, Union
 
 import torch
 from flashinfer.norm import fused_add_rmsnorm, rmsnorm
@@ -21,8 +20,8 @@ class HuggingFaceRMSNorm(nn.Module):
     def forward(
         self,
         x: torch.Tensor,
-        residual: Optional[torch.Tensor] = None,
-    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
+        residual: torch.Tensor | None = None,
+    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
         orig_dtype = x.dtype
         x = x.to(torch.float32)
         if residual is not None:
@@ -41,7 +40,7 @@ class HuggingFaceRMSNorm(nn.Module):
 def rmsnorm_naive(
     x: torch.Tensor,
     weight: torch.Tensor,
-    residual: Optional[torch.Tensor] = None,
+    residual: torch.Tensor | None = None,
     eps: float = 1e-6,
 ):
     naive_norm = HuggingFaceRMSNorm(x.shape[-1], eps=eps)
@@ -65,7 +64,7 @@ def rmsnorm_naive(
 def rmsnorm_flashinfer(
     x: torch.Tensor,
     weight: torch.Tensor,
-    residual: Optional[torch.Tensor] = None,
+    residual: torch.Tensor | None = None,
     eps: float = 1e-6,
 ):
     orig_shape = x.shape
@@ -89,7 +88,7 @@ def rmsnorm_flashinfer(
 def rmsnorm_vllm(
     x: torch.Tensor,
     weight: torch.Tensor,
-    residual: Optional[torch.Tensor] = None,
+    residual: torch.Tensor | None = None,
     eps: float = 1e-6,
 ):
     orig_shape = x.shape

benchmarks/kernels/benchmark_rope.py

@@ -2,7 +2,6 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from itertools import accumulate
-from typing import Optional
 
 import nvtx
 import torch
@@ -18,7 +17,7 @@ def benchmark_rope_kernels_multi_lora(
     seq_len: int,
     num_heads: int,
     head_size: int,
-    rotary_dim: Optional[int],
+    rotary_dim: int | None,
     dtype: torch.dtype,
     seed: int,
     device: str,

benchmarks/kernels/benchmark_trtllm_decode_attention.py

@@ -4,7 +4,6 @@
 import csv
 import os
 from datetime import datetime
-from typing import Optional
 
 import flashinfer
 import torch
@@ -28,9 +27,7 @@ def to_float8(x, dtype=torch.float8_e4m3fn):
 @torch.no_grad()
 def benchmark_decode(
     dtype: torch.dtype,
-    quant_dtypes: tuple[
-        Optional[torch.dtype], Optional[torch.dtype], Optional[torch.dtype]
-    ],
+    quant_dtypes: tuple[torch.dtype | None, torch.dtype | None, torch.dtype | None],
     batch_size: int,
     max_seq_len: int,
     num_heads: tuple[int, int] = (64, 8),

benchmarks/kernels/benchmark_trtllm_prefill_attention.py

@@ -4,7 +4,6 @@
 import csv
 import os
 from datetime import datetime
-from typing import Optional
 
 import flashinfer
 import torch
@@ -28,9 +27,7 @@ def to_float8(x, dtype=torch.float8_e4m3fn):
 @torch.no_grad()
 def benchmark_prefill(
     dtype: torch.dtype,
-    quant_dtypes: tuple[
-        Optional[torch.dtype], Optional[torch.dtype], Optional[torch.dtype]
-    ],
+    quant_dtypes: tuple[torch.dtype | None, torch.dtype | None, torch.dtype | None],
     batch_size: int,
     max_seq_len: int,
     num_heads: tuple[int, int] = (64, 8),

benchmarks/kernels/utils.py

@@ -2,8 +2,8 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import dataclasses
-from collections.abc import Iterable
-from typing import Any, Callable, Optional
+from collections.abc import Callable, Iterable
+from typing import Any
 
 import torch
 import torch.utils.benchmark as TBenchmark
@@ -55,7 +55,7 @@ class Bench:
 
     def __init__(
         self,
-        cuda_graph_params: Optional[CudaGraphBenchParams],
+        cuda_graph_params: CudaGraphBenchParams | None,
         label: str,
         sub_label: str,
         description: str,