Add help for run_benchmark (#2361)

xuzhao9 · facebook-github-bot · commit bb5294090a39 · 2024-07-08T17:52:20.000-07:00
Summary: Show more help messages for Tritonbench ``` $ python run_benchmark.py triton --help usage: run_benchmark.py [-h] [--op OP] [--mode {fwd,bwd,fwd_bwd}] [--bwd] [--fwd_bwd] [--device DEVICE] [--warmup WARMUP] [--iter ITER] [--csv] [--dump-csv] [--skip-print] [--plot] [--ci] [--metrics METRICS] [--only ONLY] [--baseline BASELINE] [--num-inputs NUM_INPUTS] [--keep-going] [--input-id INPUT_ID] [--test-only] [--dump-ir] options: -h, --help show this help message and exit --op OP Operator to benchmark. --mode {fwd,bwd,fwd_bwd} Test mode (fwd, bwd, or fwd_bwd). --bwd Run backward pass. --fwd_bwd Run both forward and backward pass. --device DEVICE Device to benchmark. --warmup WARMUP Num of warmup runs for reach benchmark run. --iter ITER Num of reps for each benchmark run. --csv Print result as csv. --dump-csv Dump result as csv. --skip-print Skip printing result. --plot Plot the result. --ci Run in the CI mode. --metrics METRICS Metrics to collect, split with comma. E.g., --metrics latency,tflops,speedup. --only ONLY Specify one or multiple operator implementations to run. --baseline BASELINE Override default baseline. --num-inputs NUM_INPUTS Number of example inputs. --keep-going --input-id INPUT_ID Specify the start input id to run. For example, --input-id 0 runs only the first available input sample.When used together like --input-id <X> --num-inputs <Y>, start from the input id <X> and run <Y> different inputs. --test-only Run this under test mode, potentially skipping expensive steps like autotuning. --dump-ir Dump Triton IR ``` ``` $ python run_benchmark.py triton --op gemm --num-inputs 1 --only triton_tutorial_matmul (M, N, K) triton_tutorial_matmul-latency --------------- -------------------------------- (256, 256, 256) 0.0033702 ``` Pull Request resolved: #2361 Reviewed By: jananisriram Differential Revision: D59374656 Pulled By: xuzhao9 fbshipit-source-id: 139f865895d7550a3475a1a8b4bed037a9ecc769
diff --git a/run_benchmark.py b/run_benchmark.py
@@ -24,7 +24,7 @@ def list_benchmarks() -> Dict[str, str]:
 
 def run():
     available_benchmarks = list_benchmarks()
-    parser = argparse.ArgumentParser(description="Run a TorchBench user benchmark")
+    parser = argparse.ArgumentParser(description="Run a TorchBench user benchmark", add_help=False)
     parser.add_argument(
         "bm_name",
         choices=available_benchmarks.keys(),
diff --git a/torchbenchmark/operators/addmm/operator.py b/torchbenchmark/operators/addmm/operator.py
@@ -1,6 +1,5 @@
-import csv
 import os
-import statistics
+import argparse
 from typing import Any, Callable, Generator, List, Optional, Tuple
 
 import numpy
@@ -70,8 +69,8 @@ class Operator(BenchmarkOperator):
     DEFAULT_METRICS = ["tflops", "best_config"]
     DEFAULT_PRECISION = "bf16"
 
-    def __init__(self, mode: str, device: str, extra_args: Optional[List[str]] = None):
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    def __init__(self, tb_args: argparse.Namespace, extra_args: Optional[List[str]] = None):
+        super().__init__(tb_args, extra_args)
         addmm_args = parse_args(self.extra_args)
         if addmm_args.m and addmm_args.n and addmm_args.k:
             self.shapes = [(addmm_args.m, addmm_args.k, addmm_args.n)]
diff --git a/torchbenchmark/operators/flash_attention/operator.py b/torchbenchmark/operators/flash_attention/operator.py
@@ -107,9 +107,8 @@ def parse_op_args(args: List[str]):
 class Operator(BenchmarkOperator):
     DEFAULT_PRECISION = "bf16"
 
-    def __init__(self, mode: str, device: str, extra_args: Optional[List[str]]=None):
-        # pass the framework level args (e.g., device, is_training, dtype) to the parent class
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    def __init__(self, tb_args: argparse.Namespace, extra_args: Optional[List[str]] = None):
+        super().__init__(tb_args, extra_args)
         args = parse_op_args(self.extra_args)
         self.BATCH = args.batch
         self.H = args.n_heads
diff --git a/torchbenchmark/operators/fp8_gemm/fp8_gemm.py b/torchbenchmark/operators/fp8_gemm/fp8_gemm.py
@@ -7,7 +7,7 @@
 
 from triton.runtime.jit import reinterpret
 
-from typing import Any
+from typing import Any, Optional, List
 
 from torchbenchmark.util.triton_op import (
     BenchmarkOperator,
@@ -27,8 +27,8 @@ def parse_args(args):
 class Operator(BenchmarkOperator):
     DEFAULT_METRICS = ["tflops", "gbps", "latency"]
 
-    def __init__(self, mode, device, extra_args):
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    def __init__(self, tb_args: argparse.Namespace, extra_args: Optional[List[str]] = None):
+        super().__init__(tb_args, extra_args)
         self.extra_args = parse_args(extra_args)
 
     def get_input_iter(self):
diff --git a/torchbenchmark/operators/fp8_gemm_blockwise/operator.py b/torchbenchmark/operators/fp8_gemm_blockwise/operator.py
@@ -108,8 +108,8 @@ class Operator(BenchmarkOperator):
     DEFAULT_METRICS = ["tflops", "speedup", "accuracy"]
     DEFAULT_PRECISION = "fp32"
 
-    def __init__(self, mode: str, device: str, extra_args: Optional[List[str]] = None):
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    def __init__(self, tb_args: argparse.Namespace, extra_args: Optional[List[str]] = None):
+        super().__init__(tb_args, extra_args)
         addmm_args = parse_args(self.extra_args)
         if addmm_args.m and addmm_args.n and addmm_args.k:
             self.shapes = [(addmm_args.m, addmm_args.n, addmm_args.k)]
diff --git a/torchbenchmark/operators/fp8_gemm_rowwise/operator.py b/torchbenchmark/operators/fp8_gemm_rowwise/operator.py
@@ -85,8 +85,8 @@ class Operator(BenchmarkOperator):
     DEFAULT_METRICS = ["tflops", "speedup", "accuracy"]
     DEFAULT_PRECISION = "fp32"
 
-    def __init__(self, mode: str, device: str, extra_args: Optional[List[str]] = None):
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    def __init__(self, tb_args: argparse.Namespace, extra_args: Optional[List[str]] = None):
+        super().__init__(tb_args, extra_args)
         addmm_args = parse_args(self.extra_args)
         if addmm_args.m and addmm_args.n and addmm_args.k:
             self.shapes = [(addmm_args.m, addmm_args.n, addmm_args.k)]
diff --git a/torchbenchmark/operators/gather_gemv/operator.py b/torchbenchmark/operators/gather_gemv/operator.py
@@ -4,6 +4,7 @@
 gather + gemv is the primary kernel driving mixtral perf.
 """
 
+import argparse
 import csv
 import os
 import statistics
@@ -38,8 +39,8 @@ def gbps(self, fn_name, example_inputs, metrics: BenchmarkOperatorMetrics):
             * 1e-6
         )
 
-    def __init__(self, mode: str, device: str, extra_args: List[str] = []):
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    def __init__(self, tb_args: argparse.Namespace, extra_args: Optional[List[str]] = None):
+        super().__init__(tb_args, extra_args)
 
     @register_benchmark(baseline=True)
     def test_0(self, p1, p2, p3) -> Callable:
diff --git a/torchbenchmark/operators/gemm/operator.py b/torchbenchmark/operators/gemm/operator.py
@@ -1,3 +1,4 @@
+import argparse
 import csv
 import os
 import statistics
@@ -78,8 +79,8 @@ class Operator(BenchmarkOperator):
     DEFAULT_METRICS = ["latency", "speedup", "accuracy", "tflops", "best_config"]
     DEFAULT_PRECISION = "fp16"
 
-    def __init__(self, mode: str, device: str, extra_args: Optional[List[str]] = None):
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    def __init__(self, tb_args: argparse.Namespace, extra_args: Optional[List[str]] = None):
+        super().__init__(tb_args, extra_args)
         gemm_args = parse_args(self.extra_args)
         if gemm_args.input:
             self.shapes = read_shapes_from_csv(gemm_args.input)
diff --git a/torchbenchmark/operators/int4_gemm/int4_gemm.py b/torchbenchmark/operators/int4_gemm/int4_gemm.py
@@ -12,7 +12,7 @@
 import triton.ops
 import triton.language as tl
 
-from typing import Any
+from typing import Any, Optional, List
 
 from torchbenchmark.util.triton_op import (
     BenchmarkOperator,
@@ -27,8 +27,8 @@
 class Operator(BenchmarkOperator):
     DEFAULT_METRICS = ["tflops", "gbps", "latency", "best_config"]
 
-    def __init__(self, mode, device, extra_args):
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    def __init__(self, tb_args: argparse.Namespace, extra_args: Optional[List[str]] = None):
+        super().__init__(tb_args, extra_args)
         # `Group size` and `inner K tiles` are defaults from gpt-fast.
         self.group_size = 32
         self.inner_k_tiles = 8
diff --git a/torchbenchmark/operators/jagged_mean/operator.py b/torchbenchmark/operators/jagged_mean/operator.py
@@ -101,8 +101,8 @@ class Operator(BenchmarkOperator):
         False  # enables GPU/CPU sync (for methods like NestedTensor unbind)
     )
 
-    def __init__(self, mode: str, device: str, extra_args: Optional[List[str]] = None):
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    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)
         )  # bias towards larger sizes, which are more representative of real-world shapes
diff --git a/torchbenchmark/operators/jagged_sum/operator.py b/torchbenchmark/operators/jagged_sum/operator.py
@@ -125,8 +125,8 @@ class Operator(BenchmarkOperator):
         False  # enables GPU/CPU sync (for methods like NestedTensor unbind)
     )
 
-    def __init__(self, mode: str, device: str, extra_args: Optional[List[str]] = None):
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    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)
         )  # bias towards larger sizes, which are more representative of real-world shapes
diff --git a/torchbenchmark/operators/sum/operator.py b/torchbenchmark/operators/sum/operator.py
@@ -151,8 +151,8 @@ class Operator(BenchmarkOperator):
 
     DEFAULT_METRICS = ["latency", "accuracy", "best_config"]
 
-    def __init__(self, mode: str, device: str, extra_args: Optional[List[str]] = None):
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    def __init__(self, tb_args: argparse.Namespace, extra_args: Optional[List[str]] = None):
+        super().__init__(tb_args, extra_args)
         args = parse_op_args(self.extra_args)
         self.input_dim = args.input_dim
         self.reduce_dim = args.reduce_dim
diff --git a/torchbenchmark/operators/template_attention/operator.py b/torchbenchmark/operators/template_attention/operator.py
@@ -1,4 +1,4 @@
-
+import argparse
 import csv
 import os
 import statistics
@@ -29,8 +29,8 @@
 class Operator(BenchmarkOperator):
     DEFAULT_METRICS = ["latency", "speedup", "accuracy"]
 
-    def __init__(self, mode: str, device: str, extra_args: List[str] = []):
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    def __init__(self, tb_args: argparse.Namespace, extra_args: Optional[List[str]] = None):
+        super().__init__(tb_args, extra_args)
         self.shapes = BUILDIN_SHAPES
 
     @register_benchmark(baseline=True)
diff --git a/torchbenchmark/operators/test_op/operator.py b/torchbenchmark/operators/test_op/operator.py
@@ -1,3 +1,4 @@
+import argparse
 from typing import Generator, List, Optional
 
 import torch
@@ -14,8 +15,8 @@ class Operator(BenchmarkOperator):
 
     DEFAULT_METRICS = ["test_metric"]
 
-    def __init__(self, mode: str, device: str, extra_args: Optional[List[str]] = None):
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    def __init__(self, tb_args: argparse.Namespace, extra_args: Optional[List[str]] = None):
+        super().__init__(tb_args, extra_args)
 
     @register_benchmark(label="new_op_label")
     def test_op(self, x: torch.Tensor):
diff --git a/torchbenchmark/operators/welford/operator.py b/torchbenchmark/operators/welford/operator.py
@@ -1,4 +1,4 @@
-
+import argparse
 import csv
 import os
 import statistics
@@ -38,8 +38,8 @@
 class Operator(BenchmarkOperator):
     DEFAULT_METRICS = ["latency", "speedup", "accuracy"]
 
-    def __init__(self, mode: str, device: str, extra_args: List[str] = []):
-        super().__init__(mode=mode, device=device, extra_args=extra_args)
+    def __init__(self, tb_args: argparse.Namespace, extra_args: Optional[List[str]] = None):
+        super().__init__(tb_args, extra_args)
         self.shapes = BUILDIN_SHAPES
 
     @register_benchmark()
diff --git a/torchbenchmark/util/triton_op.py b/torchbenchmark/util/triton_op.py
@@ -64,7 +64,7 @@
     "fp16": torch.float16,
     "bf16": torch.bfloat16,
 }
-
+_RANGE_NAME = "tritonbench_range"
 
 class Mode(Enum):
     FWD = "fwd"
@@ -380,7 +380,6 @@ def _inner(self, *args, **kwargs):
 
     return decorator
 
-
 def register_metric(
     # Metrics that only apply to non-baseline impls
     # E.g., accuracy, speedup
@@ -408,100 +407,47 @@ def _inner(self, *args, **kwargs):
 
     return decorator
 
-
-def parse_args(
-    default_metrics: List[str],
-    args: List[str],
-) -> Tuple[argparse.Namespace, List[str]]:
-    parser = argparse.ArgumentParser(allow_abbrev=False)
-    parser.add_argument(
-        "--metrics",
-        default=",".join(default_metrics),
-        help="Metrics to collect, split with comma. E.g., --metrics latency,tflops,speedup.",
-    )
-    parser.add_argument(
-        "--only",
-        default=None,
-        help="Specify one or multiple operator implementations to run."
-    )
-    parser.add_argument(
-        "--baseline",
-        type=str,
-        default=None,
-        help="Override default baseline."
-    )
-    parser.add_argument(
-        "--num-inputs",
-        type=int,
-        help="Number of example inputs.",
-    )
-    parser.add_argument(
-        "--keep-going",
-        action="store_true",
-    )
-    parser.add_argument(
-        "--input-id",
-        type=int,
-        default=0,
-        help="Specify the start input id to run. " \
-            "For example, --input-id 0 runs only the first available input sample." \
-            "When used together like --input-id <X> --num-inputs <Y>, start from the input id <X> " \
-            "and run <Y> different inputs."
-    )
-    parser.add_argument(
-        "--test-only",
-        action="store_true",
-        help="Run this under test mode, potentially skipping expensive steps like autotuning."
-    )
-    parser.add_argument(
-        "--dump-ir",
-        action="store_true",
-        help="Dump Triton IR",
-    )
-    return parser.parse_known_args(args)
-
 class PostInitProcessor(type):
     def __call__(cls, *args, **kwargs):
         obj = type.__call__(cls, *args, **kwargs)
         obj.__post__init__()
         return obj
 
-
-_RANGE_NAME = "tritonbench_range"
-
-
 class BenchmarkOperator(metaclass=PostInitProcessor):
     mode: Mode = Mode.FWD
     test: str = "eval"
     device: str = "cuda"
+    # By default, only collect latency metrics
+    # Each operator can override to define their own default metrics
+    DEFAULT_METRICS = ["latency"]
+    required_metrics: List[str]
     _input_iter: Optional[Generator] = None
     extra_args: List[str] = []
     example_inputs: Any = None
     use_cuda_graphs: bool = True
 
-    # By default, only collect latency metrics
-    # Each operator can override to define their own default metrics
-    DEFAULT_METRICS = ["latency"]
-
     """
     A base class for adding operators to torch benchmark.
     """
 
-    def __init__(self, mode: str, device: str, extra_args: Optional[List[str]]=None):
+    def __init__(self, tb_args: argparse.Namespace, extra_args: Optional[List[str]]=None):
         set_random_seed()
         self.name = _find_op_name_from_module_path(self.__class__.__module__)
         self._raw_extra_args = copy.deepcopy(extra_args)
+        self.tb_args = tb_args
         # we accept both "fwd" and "eval"
-        if mode == "fwd":
+        if self.tb_args.mode == "fwd":
             self.mode = Mode.FWD
-        elif mode == "fwd_bwd":
+        elif self.tb_args.mode == "fwd_bwd":
             self.mode = Mode.FWD_BWD
         else:
             assert (
-                mode == "bwd"
+                self.tb_args.mode == "bwd"
             ), f"We only accept 3 test modes: fwd(eval), fwd_bwd(train), or bwd."
             self.mode = Mode.BWD
-        self.dargs, unprocessed_args = parse_decoration_args(self, extra_args)
+        self.device = tb_args.device
+        self.required_metrics = list(set(tb_args.metrics.split(","))) if tb_args.metrics else self.DEFAULT_METRICS
+        self.dargs, self.extra_args = parse_decoration_args(self, extra_args)
         if self.name not in REGISTERED_X_VALS:
             REGISTERED_X_VALS[self.name] = "x_val"
         # This will be changed by the time we apply the decoration args
@@ -510,17 +456,11 @@ def __init__(self, mode: str, device: str, extra_args: Optional[List[str]]=None)
             [x for x in REGISTERED_METRICS.get(self.name, []) if x not in BUILTIN_METRICS]
         )
         self.DEFAULT_METRICS = list(set(self.DEFAULT_METRICS))
-        self.tb_args, self.extra_args = parse_args(
-            self.DEFAULT_METRICS,
-            unprocessed_args
-        )
         if self.tb_args.baseline:
             BASELINE_BENCHMARKS[self.name] = self.tb_args.baseline
-        self.required_metrics = list(set(self.tb_args.metrics.split(",")))
         self._only = _split_params_by_comma(self.tb_args.only)
         self._input_id = self.tb_args.input_id
         self._num_inputs = self.tb_args.num_inputs
-        self.device = device
 
     # Run the post initialization
     def __post__init__(self):
diff --git a/userbenchmark/triton/run.py b/userbenchmark/triton/run.py
