File size: 9,934 Bytes
b3a3b15 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 |
import math
import torch
from functools import partial
import pandas as pd
import matplotlib.pyplot as plt
from HybridTensor.utils.utils import get_gpu_name, create_results_directory
from tqdm import tqdm # For progress bars
# from HybridTensor.modules.MLP import StandardMLPBlock, SelectiveMLP, SelectiveMLPTriton
# from HybridTensor.utils.utils import sparse_index
def benchmark_mlp_fwd(x, model, index_vec = None, iterations = 100, print_result = False):
class_name_ = model.__class__.__name__
if index_vec is not None:
model = partial(model, index_vec=index_vec)
# warm up, this also compiles the triton kernel before measuring the execution time
for _ in range(10):
out = model(x)
torch.cuda.synchronize()
start = torch.cuda.Event(enable_timing=True)
start.record()
for _ in range(iterations):
out = model(x)
torch.cuda.synchronize()
end = torch.cuda.Event(enable_timing=True)
end.record()
torch.cuda.synchronize()
elapsed_time = start.elapsed_time(end)/iterations
if print_result:
print(f"{class_name_} Execution time: {elapsed_time} ms")
return out, elapsed_time
def generate_index_sizes(hidden_features):
index_sizes = []
idx = 0
while idx < hidden_features:
idx += 512
index_sizes.append(min(idx, hidden_features))
return index_sizes
def save_results_to_csv(df, filename_prefix='mlp_profiling_results', results_dir = create_results_directory("results")):
"""
Saves the profiling results DataFrame to a CSV file within the specified results directory.
Parameters:
- df (pd.DataFrame): The DataFrame containing profiling results.
- filename_prefix (str): The prefix for the CSV filename.
- results_dir (Path): The Path object for the results directory.
"""
# Retrieve the GPU name
gpu_name = get_gpu_name()
# Define the filename with GPU name
filename = f"{filename_prefix}_{gpu_name}.csv"
# Define the full path for the CSV file
csv_path = results_dir / filename
# Save the DataFrame to the CSV file
df.to_csv(csv_path, index=False)
print(f"Results saved to {csv_path}")
def plot_results(df, output_prefix='mlp_profiling', results_dir=create_results_directory("results")):
"""
Plots the profiling results and saves the plot image within the specified results directory.
Parameters:
- df (pd.DataFrame): The DataFrame containing profiling results.
- output_prefix (str): The prefix for the plot filename.
- results_dir (Path): The Path object for the results directory.
"""
plt.figure(figsize=(12, 6))
# Plot Execution Time
plt.subplot(1, 2, 1)
plt.plot(df['index_size'], df['standard_time'], label='Standard MLP', marker='o')
plt.plot(df['index_size'], df['selective_cutlass_time'], label='Selective MLP Cutlass', marker='o')
plt.plot(df['index_size'], df['selective_triton_time'], label='Selective MLP Triton', marker='o')
plt.xlabel('Index Size')
plt.ylabel('Execution Time (ms)')
plt.title('Execution Time vs. Index Size')
plt.legend()
plt.grid(True)
# Plot Speedup
plt.subplot(1, 2, 2)
plt.plot(df['index_size'], df['cutlass_speedup'], label='Cutlass Speedup', marker='o')
plt.plot(df['index_size'], df['triton_speedup'], label='Triton Speedup', marker='o')
plt.xlabel('Index Size')
plt.ylabel('Speedup')
plt.title('Speedup vs. Index Size')
plt.legend()
plt.grid(True)
plt.tight_layout()
# Retrieve the GPU name
gpu_name = get_gpu_name()
# Define the filename with GPU name
plot_filename = f"{output_prefix}_{gpu_name}.png"
# Define the full path for the plot image
plot_path = results_dir / plot_filename
# Save the plot
plt.savefig(plot_path)
plt.show()
print(f"Plots saved as {plot_path}")
def cuda_profiler(func, *args, warmup_runs=10, timed_runs=1000, **kwargs):
"""
Generic profiler function to measure execution time of a given function.
Parameters:
- func: The function to be profiled.
- *args: Positional arguments to be passed to the function.
- warmup_runs: Number of warm-up runs (default: 2).
- timed_runs: Number of timed iterations (default: 10).
- **kwargs: Keyword arguments to be passed to the function.
Returns:
- float: The average execution time of the function in milliseconds.
"""
# Warm-up phase
for _ in range(warmup_runs):
out = func(*args, **kwargs)
# Synchronize before starting the timer to ensure accurate measurements
torch.cuda.synchronize()
# Create CUDA events for measuring execution time
start_event = torch.cuda.Event(enable_timing=True)
end_event = torch.cuda.Event(enable_timing=True)
# Record execution times for the given number of runs
start_event.record()
for _ in range(timed_runs):
# Execute the function
out = func(*args, **kwargs)
# Wait for the events to be completed
# torch.cuda.synchronize()
end_event.record()
torch.cuda.synchronize()
# Calculate elapsed time for this iteration
elapsed_time = start_event.elapsed_time(end_event)
# Calculate average time per run
avg_time = elapsed_time / timed_runs
return out, avg_time
from HybridTensor.triton.select_attn_v1 import select_attn
from HybridTensor.utils.utils import generate_BH_index
from HybridTensor.triton.references.attention_proj_sparse import qkv_proj_sparse, out_proj_sparse
def _sim_cache_update(k, v, qkv, seq_len):
k[:, -1, ...] = qkv[:, :, 1]
v[:, -1, ...] = qkv[:, :, 2]
def mha_inference_simulation(B, in_features, seq_len, head_density, active_density):
'''
Simulates the execultion time of a standard MHA layer and a selective MHA layer with sparse projection and select_attn.
Parameters:
- B: batch size
- in_features: number of features
- seq_len: sequence length
- head_density: the percentage of heads that are active per batch
- active_density: the percentage of active heads per layer (aggregate active heads in all batches)
'''
# Test parameters
H = in_features // 128 # Number of heads
G = 1 # Group size
M = 1 # Sequence length for queries
Mk = seq_len # Sequence length for keys/values
Kq = 128 # Embedding size for queries
Kkv = 128 # Embedding size for keys/values
dtype = torch.float16
device = 'cuda' if torch.cuda.is_available() else 'cpu'
device = 'cuda:0'
dtype = torch.float16
x = torch.rand(B, in_features, dtype=dtype).to(device)
proj_dim = 3 * in_features
# Define the Linear layer
qkv_project = torch.nn.Linear(in_features, proj_dim, dtype=dtype).to(device)
out_project = torch.nn.Linear(in_features, in_features, dtype=dtype).to(device)
weight = torch.randn(3, H, Kkv, in_features, device=device, dtype=dtype)
bias = torch.randn(3, H, Kkv, device=device, dtype=dtype)
n_active_heads = math.ceil(H * active_density)
head_idx = torch.randperm(H, device=device, dtype=torch.int32)[:n_active_heads]
batch_idx = torch.stack([
torch.arange(B, dtype=torch.int32, device=device)
for _ in range(n_active_heads)
])
print(f"Batch size: {B}, Total heads: {H}, Features: {in_features}, Seq len: {seq_len}")
print(f"Total active heads: {n_active_heads}")
print(f"Head density in SelectAttn: {head_density}")
print("====================================")
# Inference simulation
qkv, qkv_project_time = cuda_profiler(qkv_project, x)
print(f"qkv projection time: {qkv_project_time:.3f} ms")
qkv_sel, qkv_sel_proj_time = cuda_profiler(qkv_proj_sparse, x, weight, head_idx, batch_idx, bias)
print(f"qkv projection time: {qkv_sel_proj_time:.3f} ms")
# Generate random tensors for q, k, v
q = torch.randn(B, M, G, H, Kq, dtype=dtype, device=device)
k = torch.randn(B, Mk, G, H, Kkv, dtype=dtype, device=device)
v = torch.randn(B, Mk, G, H, Kkv, dtype=dtype, device=device)
# need to update kv cache with the new k, v
_sim_cache_update(k, v, qkv, seq_len)
_, kv_cache_update_time = cuda_profiler(_sim_cache_update, k, v, qkv, seq_len)
print(f"KV cache update time: {kv_cache_update_time:.3f} ms")
scale = 1 / (Kq ** 0.5)
batch_head_index_1 = generate_BH_index(B, H, math.ceil(H * 1))
triton_sel_output, attn_time = cuda_profiler(select_attn, q, k, v, scale, batch_head_index_1)
print(f"Attention time: {attn_time:.3f} ms")
batch_head_index_2 = generate_BH_index(B, H, math.ceil(H * head_density))
triton_sel_output_2, select_attn_time = cuda_profiler(select_attn, q, k, v, scale, batch_head_index_2)
print(f"SelectAttn time: {select_attn_time:.3f} ms")
triton_sel_output_2, view_time = cuda_profiler(triton_sel_output_2.view, B, in_features)
# Out projection
out, out_project_time = cuda_profiler(out_project, triton_sel_output_2)
print(f"out projection time: {out_project_time:.3f} ms")
standard_time = qkv_project_time + attn_time + out_project_time
select_time = qkv_project_time + select_attn_time + out_project_time
select_time_sparse_project = qkv_sel_proj_time + select_attn_time + out_project_time
print("====================================")
print(f"Standard time: {standard_time:.3f} ms")
print(f"Select time: {select_time:.3f} ms")
print(f"Select time with sparse project: {select_time_sparse_project:.3f} ms")
print("====================================")
print(f"Selective Speedup: {standard_time / select_time:.3f}")
print(f"Selective Speedup with sparse project: {standard_time / select_time_sparse_project:.3f}")
# free cuda memory
del qkv, qkv_sel, q, k, v, triton_sel_output, triton_sel_output_2, out
torch.cuda.empty_cache()
|