triton-windows 3.3.1.post21__cp311-cp311-win_amd64.whl → 3.4.0.post21__cp311-cp311-win_amd64.whl

triton/language/standard.py

@@ -9,7 +9,7 @@ from . import math
 
 def _log2(i: core.constexpr):
     log2 = 0
-    n = i.value
+    n = core.constexpr(i).value
     while n > 1:
         n >>= 1
         log2 += 1
@@ -50,10 +50,14 @@ def sigmoid(x):
 @core._tensor_member_fn
 @jit
 @math._add_math_1arg_docstr("softmax")
-def softmax(x, ieee_rounding=False):
-    z = x - max(x, 0)
+def softmax(x, dim=None, keep_dims=False, ieee_rounding=False):
+    if dim is None:
+        _dim: core.constexpr = 0
+    else:
+        _dim: core.constexpr = dim
+    z = x - max(x, _dim, keep_dims=keep_dims)
     num = math.exp(z)
-    den = sum(num, 0)
+    den = sum(num, _dim, keep_dims=keep_dims)
     return math.fdiv(num, den, ieee_rounding)
 
 
@@ -302,15 +306,37 @@ def xor_sum(input, axis=None, keep_dims=False):
     return core.reduce(input, axis, _xor_combine, keep_dims=keep_dims)
 
 
+# or reduction
+
+
+@jit
+def _or_combine(x, y):
+    return x | y
+
+
+@core._tensor_member_fn
+@jit
+@core._add_reduction_docstr("reduce_of")
+def reduce_or(input, axis, keep_dims=False):
+    core.static_assert(input.type.scalar.is_int(), "reduce_of only supported for integers")
+    return core.reduce(input, axis, _or_combine, keep_dims=keep_dims)
+
+
 # cumsum
 
 
 @core._tensor_member_fn
 @jit
-@core._add_scan_docstr("cumsum")
-def cumsum(input, axis=0, reverse=False):
+@core._add_scan_docstr("cumsum", dtype_arg="dtype")
+def cumsum(input, axis=0, reverse=False, dtype: core.constexpr = None):
     # todo rename this to a generic function name
+
     input = core._promote_bfloat16_to_float32(input)
+    out_dtype: core.constexpr = _pick_sum_dtype(input.dtype, dtype)
+
+    if out_dtype is not None:
+        input = input.to(out_dtype)
+
     return core.associative_scan(input, axis, _sum_combine, reverse)
 
 
@@ -335,53 +361,63 @@ def cumprod(input, axis=0, reverse=False):
 
 
 @jit
-def _compare_and_swap(x, flip, i: core.constexpr, n_dims: core.constexpr):
-    n_outer: core.constexpr = x.numel >> n_dims
-    shape: core.constexpr = [n_outer * 2**i, 2, 2**(n_dims - i - 1)]
-    y = core.reshape(x, shape)
-    # slice left/right with 'stride' 2**(n_dims - i - 1)
-    mask = core.arange(0, 2)[None, :, None]
-    left = core.broadcast_to(sum(y * (1 - mask), 1)[:, None, :], shape).to(y.dtype)
-    right = core.broadcast_to(sum(y * mask, 1)[:, None, :], shape).to(y.dtype)
-    left = core.reshape(left, x.shape)
-    right = core.reshape(right, x.shape)
-    # actual compare-and-swap
+def _indicator(n_dims: core.constexpr, j: core.constexpr):
+    ar = core.arange(0, 2)
+    ar = core.reshape(ar, [1] * (n_dims - j - 1) + [2] + [1] * j)
+    return ar
+
+
+@jit
+def _compare_and_swap(x, flip, i: core.constexpr):
+    # compare-and-swap on the ith *innermost* dimension
+    n_dims: core.constexpr = _log2(x.numel)
+
+    # flip along middle dimension (the bitwise XORs will be optimised away):
     idtype = core.get_int_dtype(bitwidth=x.dtype.primitive_bitwidth, signed=True)
-    ileft = left.to(idtype, bitcast=True)
-    iright = right.to(idtype, bitcast=True)
     ix = x.to(idtype, bitcast=True)
-    ret = ix ^ core.where((left > right) != flip, ileft ^ iright, zeros_like(ix))
-    return ret.to(x.dtype, bitcast=True)
+    iy = ix ^ xor_sum(ix, n_dims - 1 - i, True)
+    y = iy.to(x.dtype, bitcast=True)
+
+    # determines whether we are in the right (rather than left) position along the axis:
+    is_right = _indicator(n_dims, i)
+
+    # conditional swap:
+    ret = core.where((x > y) != (flip ^ is_right), y, x)
+    return ret
 
 
 @jit
-def _bitonic_merge(x, stage: core.constexpr, order: core.constexpr, n_dims: core.constexpr):
+def _bitonic_merge_hypercube(x, stage: core.constexpr, order: core.constexpr):
     '''
     order_type 0 == ascending
     order_type 1 == descending
     order_type 2 == alternating
     '''
-    n_outer: core.constexpr = x.numel >> n_dims
-    core.static_assert(stage <= n_dims)
     # flip denotes whether to re-arrange sub-sequences of elements in ascending or
     # descending order.
     # if flip = 00000000... then all elements will be re-arranged ascendingly at this stage
     # if flip = 00110011... then all the elements will be re-arranged alternatingly (with
     # a stride of 2) at this stage
     if order == 2:
-        shape: core.constexpr = [n_outer * 2**(n_dims - 1 - stage), 2, 2**stage]
-        flip = core.reshape(core.broadcast_to(core.arange(0, 2)[None, :, None], shape), x.shape)
+        flip = _indicator(_log2(x.numel), stage)
     else:
         flip = order
     # perform `stage` rounds of `compare-and-swap`
     for i in core.static_range(stage):
-        x = _compare_and_swap(x, flip, i + (n_dims - stage), n_dims)
+        x = _compare_and_swap(x, flip, stage - 1 - i)
     return x
 
 
-@core._tensor_member_fn
 @jit
-def sort(x, dim: core.constexpr = None, descending: core.constexpr = core.CONSTEXPR_0):
+def _bitonic_merge(x, stage: core.constexpr, order: core.constexpr, n_dims: core.constexpr):
+    h = core.reshape(x, [2] * _log2(x.numel))
+    h = _bitonic_merge_hypercube(h, stage, order)
+    x = core.reshape(h, x.shape)
+    return x
+
+
+@jit
+def sort_impl(x, k: core.constexpr = None, dim: core.constexpr = None, descending: core.constexpr = core.CONSTEXPR_0):
     """
     Sorts a tensor along a specified dimension.
 
@@ -389,20 +425,55 @@ def sort(x, dim: core.constexpr = None, descending: core.constexpr = core.CONSTE
     :type x: Tensor
     :param dim: The dimension along which to sort the tensor. If None, the tensor is sorted along the last dimension. Currently, only sorting along the last dimension is supported.
     :type dim: int, optional
+    :param k: the number of top elements to select. If none, assume k = x.shape[dim]
+    :type k: int, optional
     :param descending: If set to True, the tensor is sorted in descending order. If set to False, the tensor is sorted in ascending order.
     :type descending: bool, optional
     """
     # handle default dimension or check that it is the most minor dim
     _dim: core.constexpr = len(x.shape) - 1 if dim is None else dim
     core.static_assert(_dim == len(x.shape) - 1, "only minor dimension is currently supported")
-    # iteratively run bitonic merge-sort steps
-    n_dims: core.constexpr = _log2(x.shape[_dim])
-    for i in core.static_range(1, n_dims + 1):
-        x = _bitonic_merge(x, i, 2 if i < n_dims else descending, n_dims)
+
+    log_n: core.constexpr = _log2(x.shape[_dim])
+    log_k: core.constexpr = log_n if k is None else _log2(k)
+
+    n_dims: core.constexpr = _log2(x.numel)
+
+    # reshape to hypercube:
+    h = core.reshape(x, [2] * n_dims)
+
+    # run first log_k bitonic sort iterations:
+    for i in core.static_range(1, log_k + 1):
+        h = _bitonic_merge_hypercube(h, i, 2 if i < log_n else descending)
+
+    # select top k elements using bitonic top-k
+    # https://www.doc.ic.ac.uk/~hlgr/pdfs/MassivelyParallelTopK.pdf
+    for i in core.static_range(log_k + 1, log_n + 1):
+        h = max(h, axis=(_log2(h.numel) - 1 - log_k)) if descending else min(h, axis=(_log2(h.numel) - 1 - log_k))
+        h = _bitonic_merge_hypercube(h, log_k, 2 if i < log_n else descending)
+
+    # reshape back:
+    x = core.reshape(h, x.shape[:-1] + [2**log_k])
     return x
 
 
-# flip
+@jit
+def sort(x, dim: core.constexpr = None, descending: core.constexpr = core.CONSTEXPR_0):
+    return sort_impl(x, dim=dim, descending=descending)
+
+
+@jit
+def topk(x, k: core.constexpr, dim: core.constexpr = None):
+    return sort_impl(x, k=k, dim=dim, descending=True)
+
+
+@jit
+def bitonic_merge(x, dim: core.constexpr = None, descending: core.constexpr = core.CONSTEXPR_0):
+    # handle default dimension or check that it is the most minor dim
+    _dim: core.constexpr = len(x.shape) - 1 if dim is None else dim
+    core.static_assert(_dim == len(x.shape) - 1, "only minor dimension is currently supported")
+    n_dims: core.constexpr = _log2(x.shape[-1])
+    return _bitonic_merge(x, n_dims, descending, n_dims)
 
 
 def _get_flip_dim(dim, shape):
@@ -410,7 +481,8 @@ def _get_flip_dim(dim, shape):
     shape = core._unwrap_if_constexpr(shape)
     if dim is None:
         dim = len(shape) - 1
-    assert dim == len(shape) - 1, "Currently only support flipping the last dimension"
+    if dim < 0:  # flip doesn't work if dim < 0 because the xor-swap for loop will start/end at the wrong index
+        dim += len(shape)
     return core.constexpr(dim)
 
 
@@ -422,26 +494,19 @@ def flip(x, dim=None):
 
     :param x: the first input tensor
     :type x: Block
-    :param dim: the dimension to flip along (currently only final dimension supported)
+    :param dim: the dimension to flip along
     :type dim: int
     """
-    core.static_assert(_is_power_of_two(x.shape[_get_flip_dim(dim, x.shape)]))
-    core.static_assert(_is_power_of_two(x.numel))
-    # reshape the tensor to have all dimensions be 2.
-    # TODO: We shouldn't have to change the dimensions not sorted.
-    steps: core.constexpr = _log2(x.numel)
-    start: core.constexpr = _log2(x.numel) - _log2(x.shape[_get_flip_dim(dim, x.shape)])
+    core.static_assert(-len(x.shape) <= dim and dim < len(x.shape))
+    _dim: core.constexpr = _get_flip_dim(dim, x.shape)
+    core.static_assert(_is_power_of_two(x.shape[_dim]))
+    steps: core.constexpr = _log2(x.shape[_dim])
 
+    # reshape the swap dimension to (2, 2, ..., 2)
     idtype = core.get_int_dtype(bitwidth=x.dtype.primitive_bitwidth, signed=True)
-    y = core.reshape(x.to(idtype, bitcast=True), [2] * steps)
-    y = core.expand_dims(y, start)
-    flip = (core.arange(0, 2)[:, None] == 1 - core.arange(0, 2))
-    for i in core.static_range(start, steps):
-        flip2 = flip
-        for j in core.static_range(0, steps + 1):
-            if j != i and j != i + 1:
-                flip2 = core.expand_dims(flip2, j)
-        y = sum(y * flip2, i + 1, keep_dims=True, dtype=y.dtype)
+    y = core.reshape(x.to(idtype, bitcast=True), x.shape[:_dim] + [2] * steps + x.shape[_dim + 1:])
+    for i in core.static_range(steps):
+        y = y ^ xor_sum(y, _dim + i, True)
     x = core.reshape(y, x.shape).to(x.dtype, bitcast=True)
     return x
 

triton/runtime/autotuner.py

@@ -1,11 +1,14 @@
 from __future__ import annotations
 
 import builtins
-import os
 import time
 import inspect
+import hashlib
+import json
+from functools import cached_property
 from typing import Dict, Tuple, List, Optional
 
+from .. import knobs
 from .jit import KernelInterface
 from .errors import OutOfResources, PTXASError
 from .driver import driver
@@ -13,22 +16,9 @@ from .driver import driver
 
 class Autotuner(KernelInterface):
 
-    def __init__(
-        self,
-        fn,
-        arg_names,
-        configs,
-        key,
-        reset_to_zero,
-        restore_value,
-        pre_hook=None,
-        post_hook=None,
-        prune_configs_by: Optional[Dict] = None,
-        warmup=None,
-        rep=None,
-        use_cuda_graph=False,
-        do_bench=None,
-    ):
+    def __init__(self, fn, arg_names, configs, key, reset_to_zero, restore_value, pre_hook=None, post_hook=None,
+                 prune_configs_by: Optional[Dict] = None, warmup=None, rep=None, use_cuda_graph=False, do_bench=None,
+                 cache_results=False):
         """
         :param prune_configs_by: a dict of functions that are used to prune configs, fields:
             'perf_model': performance model used to predicate running time with different configs, returns running time
@@ -36,15 +26,13 @@ class Autotuner(KernelInterface):
             'prune_num_stages_by'(optional): a function used to prune num_stages. It takes configs:List[Config] as its input, and returns pruned configs.
         """
         if not configs:
-            self.configs = [
-                Config({}, num_warps=4, num_stages=3, num_ctas=1, num_buffers_warp_spec=0, num_consumer_groups=0,
-                       reg_dec_producer=0, reg_inc_consumer=0)
-            ]
+            self.configs = [Config({}, num_warps=4, num_stages=3, num_ctas=1)]
         else:
             self.configs = configs
         self.keys = key
         self.cache: Dict[Tuple, Config] = {}
         self.arg_names = arg_names
+        self.cache_results = cache_results or (knobs.autotuning.cache and not knobs.runtime.interpret)
 
         # Reset to zero or restore values
         self.reset_to_zero = []
@@ -97,6 +85,7 @@ class Autotuner(KernelInterface):
         while not inspect.isfunction(self.base_fn):
             self.base_fn = self.base_fn.fn
 
+        self._do_bench = do_bench
         self.num_warmups = warmup
         self.num_reps = rep
         self.use_cuda_graph = use_cuda_graph
@@ -110,7 +99,7 @@ class Autotuner(KernelInterface):
                           stacklevel=1)
             if use_cuda_graph:
                 from ..testing import do_bench_cudagraph
-                self.do_bench = lambda kernel_call, quantiles: do_bench_cudagraph(
+                self._do_bench = lambda kernel_call, quantiles: do_bench_cudagraph(
                     kernel_call,
                     rep=rep if rep is not None else 100,
                     quantiles=quantiles,
@@ -118,7 +107,7 @@ class Autotuner(KernelInterface):
                 return
 
             import triton.testing
-            self.do_bench = lambda kernel_call, quantiles: triton.testing.do_bench(
+            self._do_bench = lambda kernel_call, quantiles: triton.testing.do_bench(
                 kernel_call,
                 warmup=warmup if warmup is not None else 25,
                 rep=rep if rep is not None else 100,
@@ -126,15 +115,16 @@ class Autotuner(KernelInterface):
             )
             return
 
-        if do_bench is None:
-            self.do_bench = driver.active.get_benchmarker()
-        else:
-            self.do_bench = do_bench
+    @cached_property
+    def do_bench(self):
+        if self._do_bench is None:
+            return driver.active.get_benchmarker()
+        return self._do_bench
 
     def _bench(self, *args, config, **meta):
         from ..compiler.errors import CompileTimeAssertionFailure
 
-        verbose = os.environ.get("TRITON_PRINT_AUTOTUNING", None) == "1"
+        verbose = knobs.autotuning.print
         if verbose:
             print(f"Autotuning kernel {self.base_fn.__name__} with config {config}")
 
@@ -173,6 +163,51 @@ class Autotuner(KernelInterface):
                 print(f"Autotuning failed with {e}")
             return [float("inf"), float("inf"), float("inf")]
 
+    def check_disk_cache(self, tuning_key, configs, bench_fn):
+        # We can't serialize prehooks, so just give up and run the benchmarks.
+        if not tuning_key or any(cfg.pre_hook for cfg in configs):
+            bench_fn()
+            return False
+
+        from triton._C.libtriton import get_cache_invalidating_env_vars
+        from triton.compiler.compiler import make_backend, triton_key
+        from triton.runtime.cache import get_cache_manager
+        from triton.runtime.jit import JITFunction
+
+        fn = self.fn
+        while not isinstance(fn, JITFunction):
+            fn = fn.fn
+
+        env_vars = get_cache_invalidating_env_vars()
+        cache_key = [
+            triton_key(),
+            make_backend(driver.active.get_current_target()).hash(),
+            fn.cache_key,
+            str(sorted(env_vars.items())),
+            str(tuning_key),
+        ] + [str(c) for c in configs]
+        cache_key = hashlib.sha256("-".join(cache_key).encode("utf-8")).hexdigest()
+        cache = get_cache_manager(cache_key)
+        file_name = f"{fn.__name__[:150]}.autotune.json"
+        path = cache.get_file(file_name)
+        if path:
+            with open(path, "r") as cached_configs:
+                timings = json.load(cached_configs)["configs_timings"]
+                timings = {Config(**config): timing for config, timing in timings}
+                self.cache[tuning_key] = builtins.min(timings, key=timings.get)
+                self.configs_timings = timings
+            return True
+
+        bench_fn()
+        cache.put(
+            json.dumps({
+                "key":
+                tuning_key,
+                "configs_timings":
+                [(config.__dict__, timings) for config, timings in self.configs_timings.items() if not config.pre_hook],
+            }), file_name, binary=False)
+        return False
+
     def run(self, *args, **kwargs):
         self.nargs = dict(zip(self.arg_names, args))
         used_cached_result = True
@@ -185,24 +220,31 @@ class Autotuner(KernelInterface):
                     key.append(str(arg.dtype))
             key = tuple(key)
             if key not in self.cache:
-                # prune configs
                 used_cached_result = False
                 pruned_configs = self.prune_configs(kwargs)
-                bench_start = time.perf_counter()
-                timings = {config: self._bench(*args, config=config, **kwargs) for config in pruned_configs}
-                bench_end = time.perf_counter()
-                self.bench_time = bench_end - bench_start
-                self.cache[key] = builtins.min(timings, key=timings.get)
-                full_nargs = {**self.nargs, **kwargs, **self.cache[key].all_kwargs()}
-                self.pre_hook(full_nargs, reset_only=True)
-                self.configs_timings = timings
+
+                def benchmark():
+                    bench_start = time.perf_counter()
+                    timings = {config: self._bench(*args, config=config, **kwargs) for config in pruned_configs}
+                    bench_end = time.perf_counter()
+                    self.bench_time = bench_end - bench_start
+                    self.cache[key] = builtins.min(timings, key=timings.get)
+                    full_nargs = {**self.nargs, **kwargs, **self.cache[key].all_kwargs()}
+                    self.pre_hook(full_nargs, reset_only=True)
+                    self.configs_timings = timings
+
+                if self.cache_results:
+                    used_cached_result = self.check_disk_cache(key, pruned_configs, benchmark)
+                else:
+                    benchmark()
+
             config = self.cache[key]
         else:
             config = self.configs[0]
         self.best_config = config
-        if os.getenv("TRITON_PRINT_AUTOTUNING", None) == "1" and not used_cached_result:
-            print(f"Triton autotuning for function {self.base_fn.__name__} finished after "
-                  f"{self.bench_time:.2f}s; best config selected: {self.best_config};")
+        if knobs.autotuning.print and not used_cached_result:
+            print(f"Triton autotuning for function {self.base_fn.__name__},\nwith key as {key},\n"
+                  f"finished after {self.bench_time:.2f}s,\nbest config selected: {self.best_config};")
         if config.pre_hook is not None:
             full_nargs = {**self.nargs, **kwargs, **config.all_kwargs()}
             config.pre_hook(full_nargs)
@@ -241,11 +283,11 @@ class Autotuner(KernelInterface):
     def warmup(self, *args, **kwargs):
         self.nargs = dict(zip(self.arg_names, args))
         ret = []
-        for config in self.prune_configs(kwargs):
+        for autotune_config in self.prune_configs(kwargs):
             ret.append(self.fn.warmup(
                 *args,
                 **kwargs,
-                **config.all_kwargs(),
+                **autotune_config.all_kwargs(),
             ))
         self.nargs = None
         return ret
@@ -263,27 +305,34 @@ class Config:
     :type num_warps: int
     :ivar num_stages: the number of stages that the compiler should use when software-pipelining loops.
                        Mostly useful for matrix multiplication workloads on SM80+ GPUs.
-    :type num_ctas: int
+    :type num_stages: int
     :ivar num_ctas: number of blocks in a block cluster. SM90+ only.
+    :type num_ctas: int
     :type maxnreg: Optional[int]
     :ivar maxnreg: maximum number of registers one thread can use.  Corresponds
                        to ptx .maxnreg directive.  Not supported on all platforms.
     :ivar pre_hook: a function that will be called before the kernel is called. Parameters of this
                     function are args.
+    :ivar ir_override: filename of a user-defined IR (*.{ttgir|llir|ptx|amdgcn}).
     """
 
-    def __init__(self, kwargs, num_warps=4, num_stages=3, num_ctas=1, num_buffers_warp_spec=0, num_consumer_groups=0,
-                 reg_dec_producer=0, reg_inc_consumer=0, maxnreg=None, pre_hook=None):
+    def __init__(self, kwargs, num_warps=4, num_stages=3, num_ctas=1, maxnreg=None, pre_hook=None, ir_override=None):
         self.kwargs = kwargs
         self.num_warps = num_warps
         self.num_ctas = num_ctas
         self.num_stages = num_stages
-        self.num_buffers_warp_spec = num_buffers_warp_spec
-        self.num_consumer_groups = num_consumer_groups
-        self.reg_dec_producer = reg_dec_producer
-        self.reg_inc_consumer = reg_inc_consumer
         self.maxnreg = maxnreg
         self.pre_hook = pre_hook
+        self.ir_override = ir_override
+
+    def __setstate__(self, state):
+        self.kwargs = state.get("kwargs", {})
+        self.num_warps = state.get("num_warps", 4)
+        self.num_stages = state.get("num_stages", 3)
+        self.num_ctas = state.get("num_ctas", 1)
+        self.maxnreg = state.get("maxnreg", None)
+        self.pre_hook = state.get("pre_hook", None)
+        self.ir_override = state.get("ir_override", None)
 
     def all_kwargs(self):
         return {
@@ -293,11 +342,8 @@ class Config:
                     ("num_warps", self.num_warps),
                     ("num_ctas", self.num_ctas),
                     ("num_stages", self.num_stages),
-                    ("num_buffers_warp_spec", self.num_buffers_warp_spec),
-                    ("num_consumer_groups", self.num_consumer_groups),
-                    ("reg_dec_producer", self.reg_dec_producer),
-                    ("reg_inc_consumer", self.reg_inc_consumer),
                     ("maxnreg", self.maxnreg),
+                    ("ir_override", self.ir_override),
                 ) if v is not None
             }
         }
@@ -309,16 +355,26 @@ class Config:
         res.append(f"num_warps: {self.num_warps}")
         res.append(f"num_ctas: {self.num_ctas}")
         res.append(f"num_stages: {self.num_stages}")
-        res.append(f"num_buffers_warp_spec: {self.num_buffers_warp_spec}")
-        res.append(f"num_consumer_groups: {self.num_consumer_groups}")
-        res.append(f"reg_dec_producer: {self.reg_dec_producer}")
-        res.append(f"reg_inc_consumer: {self.reg_inc_consumer}")
         res.append(f"maxnreg: {self.maxnreg}")
         return ", ".join(res)
 
+    def __hash__(self):
+        return hash((*self.all_kwargs().items(), self.pre_hook))
+
+    def __eq__(self, other):
+        self_tuple = tuple((
+            *self.all_kwargs().items(),
+            self.pre_hook,
+        ))
+        other_tuple = tuple((
+            *other.all_kwargs().items(),
+            other.pre_hook,
+        ))
+        return self_tuple == other_tuple
+
 
 def autotune(configs, key, prune_configs_by=None, reset_to_zero=None, restore_value=None, pre_hook=None, post_hook=None,
-             warmup=None, rep=None, use_cuda_graph=False, do_bench=None):
+             warmup=None, rep=None, use_cuda_graph=False, do_bench=None, cache_results=False):
     """
     Decorator for auto-tuning a :code:`triton.jit`'d function.
 
@@ -372,12 +428,14 @@ def autotune(configs, key, prune_configs_by=None, reset_to_zero=None, restore_va
     :type rep: int
     :param do_bench: a benchmark function to measure the time of each run.
     :type do_bench: lambda fn, quantiles
+    :param cache_results: whether to cache autotune timings to disk.  Defaults to False.
+    "type cache_results: bool
     """
 
     def decorator(fn):
         return Autotuner(fn, fn.arg_names, configs, key, reset_to_zero, restore_value, pre_hook=pre_hook,
                          post_hook=post_hook, prune_configs_by=prune_configs_by, warmup=warmup, rep=rep,
-                         use_cuda_graph=use_cuda_graph, do_bench=do_bench)
+                         use_cuda_graph=use_cuda_graph, do_bench=do_bench, cache_results=cache_results)
 
     return decorator