import argparse import csv import itertools from collections import defaultdict from dataclasses import asdict, dataclass from functools import partial from typing import Callable, List, Optional, Tuple import numpy as np from tabulate import tabulate from tqdm import tqdm import torch import torch.nn.functional as F from torch.nn.attention.flex_attention import ( _create_empty_block_mask, create_block_mask, create_mask, flex_attention, ) torch._dynamo.config.automatic_dynamic_shapes = False # Needed since changing args to function causes recompiles torch._dynamo.config.cache_size_limit = 1000 from torch._inductor.runtime.benchmarking import benchmarker def benchmark_torch_function_in_microseconds(func: Callable, *args, **kwargs) -> float: # warmup for _ in range(5): func(*args, **kwargs) return benchmarker.benchmark_gpu(lambda: func(*args, **kwargs)) * 1e3 @dataclass(frozen=True) class ExperimentConfig: shape: Tuple[int] score_mod: Callable mask_mod: Callable dtype: torch.dtype calculate_bwd_time: bool cal_bandwidth: bool def __post_init__(self): assert ( len(self.shape) == 6 ), "Shape must be of length 6" # [B, Hq, M, Hkv, N, D] def asdict(self): # Convert the dataclass instance to a dictionary d = asdict(self) # Remove the 'calculate_bwd_time' and `cal_bandwidth` key d.pop("calculate_bwd_time", None) d.pop("cal_bandwidth", None) d["shape(B,Hq,M,Hkv,N,D)"] = d.pop("shape") return d @dataclass(frozen=True) class Times: eager_time: float compiled_time: float @dataclass(frozen=True) class ExperimentResults: fwd_times: Times bwd_times: Optional[Times] @dataclass(frozen=True) class Experiment: config: ExperimentConfig results: ExperimentResults def asdict(self): dict1 = self.config.asdict() dict2 = asdict(self.results) return {**dict1, **dict2} def generate_inputs( batch_size: int, q_heads: int, q_sequence_length: int, kv_heads: int, kv_sequence_length: int, head_dim: int, dtype: torch.dtype, device: torch.device, requires_grad: bool, ): q_shape = (batch_size, q_sequence_length, q_heads * head_dim) kv_shape = (batch_size, kv_sequence_length, kv_heads * head_dim) assert q_heads % kv_heads == 0 num_h_groups = q_heads // kv_heads make_q = partial( torch.rand, q_shape, device=device, dtype=dtype, requires_grad=requires_grad ) make_kv = partial( torch.rand, kv_shape, device=device, dtype=dtype, requires_grad=requires_grad ) query = ( make_q().view(batch_size, q_sequence_length, q_heads, head_dim).transpose(1, 2) ) key = ( make_kv() .view(batch_size, kv_sequence_length, kv_heads, head_dim) .transpose(1, 2) ) value = ( make_kv() .view(batch_size, kv_sequence_length, kv_heads, head_dim) .transpose(1, 2) ) return query, key, value def run_single_experiment( config: ExperimentConfig, dynamic=False, max_autotune=False, ) -> ExperimentResults: device = torch.device("cuda") batch_size, q_heads, q_seq_len, kv_heads, kv_seq_len, head_dim = config.shape query, key, value = generate_inputs( batch_size, q_heads, q_seq_len, kv_heads, kv_seq_len, head_dim, config.dtype, device, requires_grad=config.calculate_bwd_time, ) kwargs = {} if get_func_name(config.mask_mod) == "causal": kwargs["is_causal"] = True def eager_sdpa(query, key, value, attn_mask): out = F.scaled_dot_product_attention(query, key, value, attn_mask, **kwargs) return out.reshape(batch_size, q_heads, q_seq_len, head_dim) if max_autotune: compiled_sdpa = torch.compile( flex_attention, dynamic=dynamic, mode="max-autotune-no-cudagraphs" ) else: compiled_sdpa = torch.compile(flex_attention, dynamic=dynamic) score_mod = config.score_mod mask_mod = config.mask_mod if mask_mod: block_mask = create_block_mask( mask_mod, 1, 1, q_seq_len, kv_seq_len, query.device ) else: block_mask = _create_empty_block_mask(query, key) if mask_mod and get_func_name(mask_mod) != "causal": attn_mask = create_mask(mask_mod, 1, 1, query.shape[-2], key.shape[-2]) else: attn_mask = None # Broadcast query/key for eager. b_key = torch.repeat_interleave(key, q_heads // kv_heads, dim=1) b_value = torch.repeat_interleave(value, q_heads // kv_heads, dim=1) forward_eager_time = benchmark_torch_function_in_microseconds( eager_sdpa, query, b_key, b_value, attn_mask ) forward_compiled_time = benchmark_torch_function_in_microseconds( compiled_sdpa, query, key, value, score_mod=score_mod, block_mask=block_mask, enable_gqa=True, ) out_eager = eager_sdpa(query, b_key, b_value, attn_mask) out_compile = compiled_sdpa( query, b_key, b_value, score_mod=score_mod, block_mask=block_mask, enable_gqa=True, ) if score_mod is None: torch.testing.assert_close(out_eager, out_compile, atol=1e-2, rtol=1e-2) if config.calculate_bwd_time: out_eager = eager_sdpa(query, b_key, b_value, attn_mask) dOut = torch.randn_like(out_eager) backward_eager_time = benchmark_torch_function_in_microseconds( out_eager.backward, dOut, retain_graph=True ) out_compile = compiled_sdpa( query, key, value, score_mod=score_mod, block_mask=block_mask, enable_gqa=True, ) dOut = torch.randn_like(out_compile) backward_compile_time = benchmark_torch_function_in_microseconds( out_compile.backward, dOut, retain_graph=True ) return ExperimentResults( fwd_times=Times(forward_eager_time, forward_compiled_time), bwd_times=Times(backward_eager_time, backward_compile_time), ) else: return ExperimentResults( fwd_times=Times(forward_eager_time, forward_compiled_time), bwd_times=None, ) def calculate_speedup(results: ExperimentResults, type: str) -> float: if type == "fwd": return results.fwd_times.eager_time / results.fwd_times.compiled_time elif type == "bwd": assert results.bwd_times is not None return results.bwd_times.eager_time / results.bwd_times.compiled_time else: raise ValueError(f"Invalid type {type}") def calculate_bandwidth( config: ExperimentConfig, results: ExperimentResults, type: str ) -> float: if type == "fwd": batch_size, q_heads, q_seq_len, kv_heads, kv_seq_len, head_dim = config.shape query_size = ( batch_size * q_heads * q_seq_len * head_dim * torch.finfo(config.dtype).bits / 8 ) kv_size = ( batch_size * kv_heads * kv_seq_len * head_dim * torch.finfo(config.dtype).bits / 8 * 2 ) output_size = query_size total_size = (query_size + kv_size + output_size) / 1e9 # In GB time_in_seconds = results.fwd_times.compiled_time / 1e6 return total_size / time_in_seconds / 1e3 else: raise ValueError(f"Invalid type {type}") def calculate_tflops(config: ExperimentConfig, results: ExperimentResults) -> float: (B, Hq, M, Hkv, N, D) = config.shape qk_flops = M * N * D * 2 softmax_flops = M * N * 2 # Not counting online softmax overhead o_flops = M * D * N * 2 # Not counting split k overhead total_flops = B * Hq * (qk_flops + softmax_flops + o_flops) return total_flops / results.fwd_times.compiled_time / 1e6 # in TFLOPs/ def get_func_name(func): if func is None: return "None" func_str = str(func) if "" in func_str: # For locally defined functions return func_str.split(".")[-1].split(" at ")[0] else: # For regular functions return func.__name__ def set_func_name(func, name): func.__name__ = name def get_average_speedups(results: List[Experiment], type: str): # Calculate speedups speedups = [calculate_speedup(r.results, type) for r in results] # Find indices of max and min speedups max_speedup_index = np.argmax(speedups) min_speedup_index = np.argmin(speedups) # Get the config dictionaries max_config_dict = results[max_speedup_index].config.asdict() min_config_dict = results[min_speedup_index].config.asdict() # Extract function names from score_mod strings max_config_dict["score_mod"] = get_func_name(max_config_dict["score_mod"]) max_config_dict["mask_mod"] = get_func_name(max_config_dict["mask_mod"]) min_config_dict["score_mod"] = get_func_name(min_config_dict["score_mod"]) min_config_dict["mask_mod"] = get_func_name(min_config_dict["mask_mod"]) # Create table data table_data = [ { "Type": "Average", "Speedup": np.mean(speedups), **dict.fromkeys(max_config_dict), }, {"Type": "Max", "Speedup": speedups[max_speedup_index], **max_config_dict}, {"Type": "Min", "Speedup": speedups[min_speedup_index], **min_config_dict}, ] return table_data def print_results(results: List[Experiment], save_path: Optional[str] = None): table_data = defaultdict(list) for experiment in results: for key, value in experiment.asdict().items(): if key == "fwd_times": for name, time in value.items(): table_data[f"fwd_{name}"].append(float(time)) elif key == "bwd_times": if experiment.config.calculate_bwd_time: for name, time in value.items(): table_data[f"bwd_{name}"].append(float(time)) else: table_data[key].append(value) # Calculate speedups fwd_speedups = [calculate_speedup(r.results, type="fwd") for r in results] table_data["fwd_speedup"] = fwd_speedups # Calculate mem + computational throughput if results[0].config.cal_bandwidth: fwd_bandwidth = [ calculate_bandwidth(r.config, r.results, type="fwd") for r in results ] table_data["fwd_mem_bw (TB/s)"] = fwd_bandwidth fwd_tflops = [calculate_tflops(r.config, r.results) for r in results] table_data["TFlops/s"] = fwd_tflops if results[0].config.calculate_bwd_time: bwd_speedups = [calculate_speedup(r.results, type="bwd") for r in results] table_data["bwd_speedup"] = bwd_speedups table_data["score_mod"] = [get_func_name(func) for func in table_data["score_mod"]] table_data["mask_mod"] = [get_func_name(func) for func in table_data["mask_mod"]] print(tabulate(table_data, headers="keys", tablefmt="github", floatfmt=".3f")) print("\n") print("FWD Speedups".center(125, "=")) print("\n") average_data = get_average_speedups(results, type="fwd") print(tabulate(average_data, headers="keys", tablefmt="github", floatfmt=".3f")) if results[0].config.calculate_bwd_time: print("\n") print("BWD Speedups".center(125, "=")) print("\n") average_data = get_average_speedups(results, type="bwd") print(tabulate(average_data, headers="keys", tablefmt="github", floatfmt=".3f")) if save_path is not None: with open(save_path, "w", newline="") as csvfile: writer = csv.DictWriter(csvfile, fieldnames=table_data.keys()) writer.writeheader() for i in range(len(next(iter(table_data.values())))): row = {k: v[i] for k, v in table_data.items()} writer.writerow(row) print(f"\nResults saved to {save_path}") def generate_score_mods(score_mods: List[str]) -> List[Callable | None]: def noop(score, b, h, m, n): return score def causal_mask(score, b, h, token_q, token_kv): return torch.where(token_q >= token_kv, score, float("-inf")) def relative_bias(score, b, h, m, n): return score + (m - n) def head_bias(score, b, h, m, n): return score + 2 * h function_dict = { "noop": None, "causal": None, "offset": None, "rel": relative_bias, "head_bias": head_bias, } return [function_dict[name] for name in score_mods] def generate_mask_mods(score_mods: List[str]) -> List[Callable | None]: def noop(b, h, m, n): return True def causal(b, h, m, n): return m >= n def gen_offset(off): def offset(b, h, m, n): return m + off >= n return offset mask_mod_dict = { "noop": None, "causal": causal, "offset": gen_offset, "rel": None, "head_bias": None, } return [mask_mod_dict[name] for name in score_mods] def generate_flash_configs( calculate_bwd: bool, dtype: torch.dtype, batch_sizes: List[int], num_heads: List[Tuple[int, int]], seq_lens: List[int], head_dims: List[int], score_mods_str: List[str], decoding: bool, kv_cache_size: List[int], cal_bandwidth: bool, ) -> List[ExperimentConfig]: assert not (calculate_bwd and decoding), "Decoding does not support backward" bs_seqlen_vals = [ (32, 512), (16, 1024), (8, 2048), (4, 4096), (2, 8192), (1, 16384), ] causal_vals = [False, True] headdim_vals = [64, 128] dim = 2048 score_mods = generate_score_mods(score_mods_str) mask_mods = generate_mask_mods(score_mods_str) all_configs = [] for ( (batch_size, seq_len), causal, head_dim, score_mod, mask_mod, ) in itertools.product( bs_seqlen_vals, causal_vals, headdim_vals, score_mods, mask_mods, ): num_heads = dim // head_dim if decoding: q_seq_len, kv_seq_len = 1, seq_len else: q_seq_len = kv_seq_len = seq_len all_configs.append( ExperimentConfig( shape=( batch_size, num_heads, q_seq_len, num_heads, kv_seq_len, head_dim, ), score_mod=score_mod, mask_mod=mask_mod, dtype=dtype, calculate_bwd_time=calculate_bwd, cal_bandwidth=cal_bandwidth, ) ) return all_configs def generate_experiment_configs( calculate_bwd: bool, dtype: torch.dtype, batch_sizes: List[int], num_heads: List[Tuple[int, int]], seq_lens: List[int], head_dims: List[int], score_mods_str: List[str], decoding: bool, kv_cache_size: List[int], cal_bandwidth: bool, ) -> List[ExperimentConfig]: assert not (calculate_bwd and decoding), "Decoding does not support backward" if decoding: q_kv_seq_lens = [(1, i) for i in seq_lens] # only testing query length == 1 else: q_kv_seq_lens = [(i, i) for i in seq_lens] # only testing q_len == kv_len dtypes = [dtype] score_mods = generate_score_mods(score_mods_str) mask_mods = generate_mask_mods(score_mods_str) all_configs = [] for ( bsz, (q_heads, kv_heads), (q_seq_len, kv_seq_len), head_dim, (score_mod, mask_mod), dtype, ) in itertools.product( kv_cache_size if kv_cache_size else batch_sizes, num_heads, q_kv_seq_lens, head_dims, zip(score_mods, mask_mods), dtypes, ): if kv_cache_size: head_size_bytes = torch.finfo(dtype).bits / 8 * head_dim bsz = int( (bsz * 1024 * 1024) // (kv_heads * kv_seq_len * head_size_bytes * 2) ) if bsz <= 0: continue assert q_heads % kv_heads == 0 if mask_mod and get_func_name(mask_mod) == "gen_offset": mask_mod = mask_mod(kv_seq_len // 2) all_configs.append( ExperimentConfig( shape=(bsz, q_heads, q_seq_len, kv_heads, kv_seq_len, head_dim), score_mod=score_mod, mask_mod=mask_mod, dtype=dtype, calculate_bwd_time=calculate_bwd, cal_bandwidth=cal_bandwidth, ) ) return all_configs def main(args): seed = 123 np.random.seed(seed) torch.manual_seed(seed) results = [] for config in tqdm( generate_experiment_configs( args.calculate_bwd, args.dtype, args.b, args.nh, args.s, args.d, args.mods, args.decoding, args.kv_cache_size, args.throughput, ) ): results.append( Experiment( config, run_single_experiment( config, dynamic=args.dynamic, max_autotune=args.max_autotune, ), ) ) print_results(results, args.save_path) def heads_input_type(s): try: hq, hkv = map(int, s.split(",")) return hq, hkv except Exception as e: raise argparse.ArgumentTypeError("Heads must be Hq,Hkv") from e if __name__ == "__main__": # Set up the argument parser parser = argparse.ArgumentParser( description="Run sweep over sizes and score mods for flex attention" ) parser.add_argument( "--dynamic", action="store_true", help="Runs a dynamic shapes version of compiled flex attention.", ) parser.add_argument( "--calculate-bwd", action="store_true", help="Calculate backward pass times" ) parser.add_argument("-dtype", type=str, help="dtype", default="bfloat16") parser.add_argument( "-b", type=int, nargs="+", help="batch sizes", default=[2, 8, 16] ) parser.add_argument( "-nh", type=heads_input_type, nargs="+", help="# of q-heads,kv-heads", default=[(16, 16), (16, 2)], ) parser.add_argument( "-s", type=int, nargs="+", help="sequence lengths", default=[512, 1024, 4096] ) parser.add_argument("-d", type=int, nargs="+", help="head dims", default=[64, 128]) parser.add_argument( "-mods", type=str, nargs="+", help="score mods", default=["noop", "causal", "rel", "head_bias"], ) parser.add_argument( "--max-autotune", action="store_true", help="Turn on max-autotune" ) parser.add_argument( "--decoding", action="store_true", help="Benchmark Decoding (query sequence length = 1)", ) parser.add_argument( "--kv-cache-size", type=int, nargs="+", required=False, help=""" key/value cache size in MiB. Ignores -b batch size and calculate batch size from kv_cache size instead when specified. """, ) parser.add_argument( "--throughput", action="store_true", help="Calculate kernel memory bandwidth & computational throughput. ", ) parser.add_argument( "--save-path", type=str, help="Path to save the results JSON file (optional)", default=None, ) # Parse arguments args = parser.parse_args() args.dtype = getattr(torch, args.dtype) main(args)