Make autotuner take do_bench as a parameter

This makes the autotuner device-agnostic. Instead of having to know
about the existence of e.g. do_bench_cudagraph, it can let the callers
decide which backend-specific benchmarking function to use.

See discussion in #4417.

Author: int3

python/test/unit/hopper/test_flashattention.py

@@ -435,8 +435,6 @@ def test_op(Z, H, N_CTX, D_HEAD, dtype=torch.float16):
 @triton.testing.perf_report(configs)
 def bench_flash_attention(BATCH, H, N_CTX, D_HEAD, mode, provider, dtype=torch.float16, device="cuda"):
     assert mode in ['fwd', 'bwd']
-    warmup = 25
-    rep = 100
     if provider == "triton":
         q = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device="cuda", requires_grad=True)
         k = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device="cuda", requires_grad=True)
@@ -447,7 +445,7 @@ def bench_flash_attention(BATCH, H, N_CTX, D_HEAD, mode, provider, dtype=torch.f
             o = fn()
             do = torch.randn_like(o)
             fn = lambda: o.backward(do, retain_graph=True)
-        ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
+        ms = triton.testing.do_bench(fn)
         return ms
     if provider == "flash":
         lengths = torch.full((BATCH, ), fill_value=N_CTX, device=device)
@@ -459,7 +457,7 @@ def bench_flash_attention(BATCH, H, N_CTX, D_HEAD, mode, provider, dtype=torch.f
             o = fn()
             do = torch.randn_like(o)
             fn = lambda: o.backward(do, retain_graph=True)
-        ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
+        ms = triton.testing.do_bench(fn)
         return ms
 
 

python/test/unit/hopper/test_persistent_warp_specialized_fused-attention.py

@@ -361,10 +361,7 @@ def test_op(Z, H, N_CTX, D_HEAD, dtype=torch.float16):
 @triton.testing.perf_report(configs)
 def bench_flash_attention(BATCH, H, N_CTX, D_HEAD, mode, provider, dtype=torch.float16, device="cuda"):
     assert mode in ['fwd', 'bwd']
-    # warmup = 25
-    # rep = 100
-    warmup = 0
-    rep = 1
+    do_bench = lambda kernel, quantiles: triton.testing.do_bench(kernel, warmup=0, rep=1, quantiles=quantiles)
     if provider == "triton":
         q = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device="cuda", requires_grad=True)
         k = torch.randn((BATCH, H, N_CTX, D_HEAD), dtype=dtype, device="cuda", requires_grad=True)
@@ -375,7 +372,7 @@ def bench_flash_attention(BATCH, H, N_CTX, D_HEAD, mode, provider, dtype=torch.f
             o = fn()
             do = torch.randn_like(o)
             fn = lambda: o.backward(do, retain_graph=True)
-        ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
+        ms = triton.testing.do_bench(fn, do_bench=do_bench)
         return ms
     if provider == "flash":
         lengths = torch.full((BATCH, ), fill_value=N_CTX, device=device)
@@ -387,7 +384,7 @@ def bench_flash_attention(BATCH, H, N_CTX, D_HEAD, mode, provider, dtype=torch.f
             o = fn()
             do = torch.randn_like(o)
             fn = lambda: o.backward(do, retain_graph=True)
-        ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
+        ms = triton.testing.do_bench(fn, do_bench=do_bench)
         return ms
 
 

python/test/unit/language/test_decorator.py

@@ -33,7 +33,9 @@ def test_triton_heuristic(device):
     src = torch.empty(N, device=device)
     dst = torch.zeros(N, device=device)
 
-    @triton.autotune(configs=[triton.Config(kwargs={'BLOCK_SIZE': 32})], key=['N'], warmup=1, rep=1)
+    do_bench = lambda kernel, quantiles: triton.testing.do_bench(kernel, quantiles=quantiles, warmup=1, rep=1)
+
+    @triton.autotune(configs=[triton.Config(kwargs={'BLOCK_SIZE': 32})], key=['N'], do_bench=do_bench)
     @triton.heuristics({'EVEN_N': lambda nargs: nargs['N'] % 2 == 0})  # test kwargs
     @triton.heuristics({'EVEN_src': lambda nargs: nargs['src'].data_ptr() % 2 == 0})  # test args
     @triton.jit

python/test/unit/runtime/test_autotuner.py

@@ -5,15 +5,18 @@
 import pytest
 
 
-@pytest.mark.parametrize('use_cuda_graph', [False, True])
-def test_kwargs(use_cuda_graph: bool, device: str):
+def do_bench(kernel_call, quantiles):
+    return triton.testing.do_bench(kernel_call, quantiles=quantiles, warmup=1, rep=1)
+
+
+def test_kwargs(device: str):
     N = 1024
     src = torch.randn(N, device=device)
     dst = torch.empty(N, device=device)
 
     configs = [triton.Config(kwargs={'BLOCK_SIZE': 32}), triton.Config(kwargs={'BLOCK_SIZE': 128})]
 
-    @triton.autotune(configs=configs, key=['N'], warmup=1, rep=1, use_cuda_graph=use_cuda_graph)
+    @triton.autotune(configs=configs, key=['N'], do_bench=do_bench)
     @triton.jit
     def _kernel(dst, src, N, BLOCK_SIZE: tl.constexpr):
         offsets = tl.program_id(0) * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
@@ -31,7 +34,7 @@ def test_restore(device):
 
     configs = [triton.Config(kwargs={'BLOCK_SIZE': 32}), triton.Config(kwargs={'BLOCK_SIZE': 128})]
 
-    @triton.autotune(configs=configs, key=['N'], restore_value=['src'], warmup=1, rep=1)
+    @triton.autotune(configs=configs, key=['N'], restore_value=['src'], do_bench=do_bench)
     @triton.jit
     def _kernel(src, N, BLOCK_SIZE: tl.constexpr):
         offsets = tl.program_id(0) * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
@@ -61,7 +64,7 @@ def _post_hook(*args, exception):
             values["has_exception"] = True
         assert values["counter"] == 0
 
-    @triton.autotune(configs=configs, key=['N'], warmup=1, rep=1, pre_hook=_pre_hook, post_hook=_post_hook)
+    @triton.autotune(configs=configs, key=['N'], do_bench=do_bench, pre_hook=_pre_hook, post_hook=_post_hook)
     @triton.heuristics({"N_STAGES": lambda nargs: 100 if nargs['N'] == 4096 else 4})
     @triton.jit
     def _kernel(src, N, N_STAGES: tl.constexpr, BLOCK_SIZE: tl.constexpr):
@@ -112,7 +115,7 @@ def perf_model(*args, **kwargs):
     else:
         prune_configs_by = {'early_config_prune': early_config_prune}
 
-    @triton.autotune(configs=configs, key=['N'], prune_configs_by=prune_configs_by, warmup=1, rep=1)
+    @triton.autotune(configs=configs, key=['N'], prune_configs_by=prune_configs_by, do_bench=do_bench)
     @triton.jit
     def _kernel(dst, src, N, BLOCK_SIZE: tl.constexpr):
         offsets = tl.program_id(0) * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)

python/triton/backends/driver.py

@@ -1,4 +1,11 @@
 from abc import ABCMeta, abstractmethod, abstractclassmethod
+from typing import Callable, List, Protocol, Sequence
+
+
+class Benchmarker(Protocol):
+
+    def __call__(self, kernel_call: Callable, *, quantiles: List[float], **kwargs) -> Sequence[float]:
+        pass
 
 
 class DriverBase(metaclass=ABCMeta):
@@ -11,6 +18,13 @@ def is_active(self):
     def get_current_target(self):
         pass
 
+    @abstractmethod
+    def get_benchmarker(self) -> Benchmarker:
+        """
+        Return the benchmarking function that this backend should use by default.
+        """
+        raise NotImplementedError
+
     def __init__(self) -> None:
         pass
 

python/triton/runtime/autotuner.py

@@ -6,9 +6,9 @@
 import inspect
 from typing import Dict
 
-from ..testing import do_bench, do_bench_cudagraph
 from .jit import KernelInterface
 from .errors import OutOfResources
+from .driver import driver
 
 
 class Autotuner(KernelInterface):
@@ -24,9 +24,10 @@ def __init__(
         pre_hook=None,
         post_hook=None,
         prune_configs_by: Dict = None,
-        warmup=25,
-        rep=100,
+        warmup=None,
+        rep=None,
         use_cuda_graph=False,
+        do_bench=None,
     ):
         """
         :param prune_configs_by: a dict of functions that are used to prune configs, fields:
@@ -88,10 +89,35 @@ def _post_hook(args, exception):
         self.base_fn = fn
         while not inspect.isfunction(self.base_fn):
             self.base_fn = self.base_fn.fn
-        self.num_warmups = warmup
-        self.num_reps = rep
-        import torch
-        self.use_cuda_graph = use_cuda_graph and torch.cuda.is_available()
+
+        # If we got explicitly called via the old interface, raise a warning
+        # and proceed with the old behavior.
+        if warmup is not None or rep is not None or use_cuda_graph:
+            import warnings
+            warnings.warn("warmup, rep, and use_cuda_graph parameters are deprecated. See _ for details.",
+                          DeprecationWarning)
+            import torch
+            if use_cuda_graph and torch.cuda.is_available():
+                from ..testing import 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,
+                )
+                return
+
+            import triton.testing
+            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,
+                quantiles=quantiles,
+            )
+
+        if do_bench is None:
+            self.do_bench = driver.active.get_benchmarker()
+        else:
+            self.do_bench = do_bench
 
     def _bench(self, *args, config, **meta):
         from ..compiler.errors import CompileTimeAssertionFailure
@@ -125,11 +151,7 @@ def kernel_call():
             self.post_hook(args, exception=None)
 
         try:
-            if self.use_cuda_graph:
-                import torch
-                with torch.cuda.stream(torch.cuda.Stream()):
-                    return do_bench_cudagraph(kernel_call, rep=self.num_reps, quantiles=(0.5, 0.2, 0.8))
-            return do_bench(kernel_call, warmup=self.num_warmups, rep=self.num_reps, quantiles=(0.5, 0.2, 0.8))
+            return self.do_bench(kernel_call, quantiles=(0.5, 0.2, 0.8))
         except (OutOfResources, CompileTimeAssertionFailure):
             return [float("inf"), float("inf"), float("inf")]
 
@@ -262,7 +284,7 @@ def __str__(self):
 
 
 def autotune(configs, key, prune_configs_by=None, reset_to_zero=None, restore_value=None, pre_hook=None, post_hook=None,
-             warmup=25, rep=100, use_cuda_graph=False):
+             warmup=None, rep=None, use_cuda_graph=False, do_bench=None):
     """
     Decorator for auto-tuning a :code:`triton.jit`'d function.
 

python/tutorials/06-fused-attention.py

@@ -599,8 +599,6 @@ def test_op(Z, H, N_CTX, HEAD_DIM, causal, dtype=torch.float16):
 @triton.testing.perf_report(configs)
 def bench_flash_attention(BATCH, H, N_CTX, HEAD_DIM, causal, mode, provider, device="cuda"):
     assert mode in ["fwd", "bwd"]
-    warmup = 25
-    rep = 100
     dtype = torch.float16
     if "triton" in provider:
         q = torch.randn((BATCH, H, N_CTX, HEAD_DIM), dtype=dtype, device=device, requires_grad=True)
@@ -618,15 +616,15 @@ def bench_flash_attention(BATCH, H, N_CTX, HEAD_DIM, causal, mode, provider, dev
             o = fn()
             do = torch.randn_like(o)
             fn = lambda: o.backward(do, retain_graph=True)
-        ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
+        ms = triton.testing.do_bench(fn)
     if provider == "flash":
         qkv = torch.randn((BATCH, N_CTX, 3, H, HEAD_DIM), dtype=dtype, device=device, requires_grad=True)
         fn = lambda: flash_attn_func(qkv, causal=causal)
         if mode == "bwd":
             o = fn()
             do = torch.randn_like(o)
             fn = lambda: o.backward(do, retain_graph=True)
-        ms = triton.testing.do_bench(fn, warmup=warmup, rep=rep)
+        ms = triton.testing.do_bench(fn)
     flops_per_matmul = 2.0 * BATCH * H * N_CTX * N_CTX * HEAD_DIM
     total_flops = 2 * flops_per_matmul
     if causal:

third_party/amd/backend/driver.py

@@ -440,3 +440,7 @@ def get_current_target(self):
         arch = device_properties['arch']
         warp_size = device_properties['warpSize']
         return GPUTarget("hip", arch.split(':')[0], warp_size)
+
+    def get_benchmarker(self):
+        from triton.testing import do_bench
+        return do_bench

third_party/nvidia/backend/driver.py

@@ -383,3 +383,7 @@ def get_current_target(self):
     def is_active():
         import torch
         return torch.cuda.is_available() and (torch.version.hip is None)
+
+    def get_benchmarker(self):
+        from triton.testing import do_bench
+        return do_bench