File size: 13,304 Bytes
518db7a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 | """
Fused Triton kernels for mHC (manifold-constrained Hyper-Connections).
Fuses rearrange + einsum operations to reduce kernel launch overhead:
- Baseline: 9 kernels per mHC module (2 rearranges + 3 einsums + 4 softmax/sinkhorn ops)
- Fused: 3-4 kernels per mHC module (fused stream mixing + separate sinkhorn)
Performance target: 3-8x speedup for 48-layer model (96 mHC modules).
"""
import torch
import triton # type: ignore[import-untyped]
import triton.language as tl # type: ignore[import-untyped]
from typing import Optional
@triton.jit
def fused_stream_mixing_kernel( # type: ignore[no-untyped-def]
# Input pointers
x_ptr, # Residual input (batch, seq_len, num_streams, stream_dim)
transformed_ptr, # Transformed input (batch, seq_len, num_streams, stream_dim)
H_res_ptr, # Doubly stochastic matrix (num_streams, num_streams)
H_pre_ptr, # Pre-mixing matrix (num_streams, num_streams)
H_post_ptr, # Post-mixing matrix (num_streams, num_streams)
# Output pointer
output_ptr, # Output (batch, seq_len, num_streams, stream_dim)
# Dimensions
batch_size: tl.constexpr,
seq_len: tl.constexpr,
num_streams: tl.constexpr,
stream_dim: tl.constexpr,
# Block size
BLOCK_SIZE: tl.constexpr,
) -> None:
"""
Fused kernel for width_connection stream mixing.
Fuses the following operations:
1. residual_mixed = einsum(H_res, x_streams, "n m, b s n d -> b s m d")
2. pre_mixed = einsum(H_pre, transformed_streams, "n m, b s n d -> b s m d")
3. post_mixed = einsum(H_post, pre_mixed, "m n, b s m d -> b s n d")
4. output = residual_mixed + post_mixed
Each thread block processes a chunk of (batch, seq_len) positions.
"""
# Get thread block ID
pid = tl.program_id(0)
# Calculate batch and sequence position for this block
total_positions = batch_size * seq_len
num_blocks = tl.cdiv(total_positions, BLOCK_SIZE)
if pid >= num_blocks:
return
# Calculate batch and seq indices for this block
block_start = pid * BLOCK_SIZE
offsets = block_start + tl.arange(0, BLOCK_SIZE)
# Mask for valid positions
mask = offsets < total_positions
# Convert flat index to (batch, seq) coordinates
batch_idx = offsets // seq_len
seq_idx = offsets % seq_len
# Process each stream dimension
for s_out in range(num_streams):
for d in range(stream_dim):
# Initialize accumulator for this output position
residual_acc = tl.zeros([BLOCK_SIZE], dtype=tl.float32)
post_acc = tl.zeros([BLOCK_SIZE], dtype=tl.float32)
# Step 1: Compute residual_mixed[b, s, s_out, d]
# = sum over s_in: H_res[s_out, s_in] * x[b, s, s_in, d]
for s_in in range(num_streams):
# Load H_res[s_out, s_in]
h_res_val = tl.load(
H_res_ptr + s_out * num_streams + s_in
)
# Load x[batch_idx, seq_idx, s_in, d]
x_offset = (
batch_idx * seq_len * num_streams * stream_dim
+ seq_idx * num_streams * stream_dim
+ s_in * stream_dim
+ d
)
x_val = tl.load(
x_ptr + x_offset,
mask=mask,
other=0.0
)
residual_acc += h_res_val * x_val
# Step 2 & 3: Compute post_mixed[b, s, s_out, d]
# First compute pre_mixed[b, s, s_mid, d] (intermediate)
# Then multiply by H_post[s_mid, s_out]
for s_mid in range(num_streams):
# Compute pre_mixed[b, s, s_mid, d]
# = sum over s_in: H_pre[s_mid, s_in] * transformed[b, s, s_in, d]
pre_acc = 0.0
for s_in in range(num_streams):
# Load H_pre[s_mid, s_in]
h_pre_val = tl.load(
H_pre_ptr + s_mid * num_streams + s_in
)
# Load transformed[batch_idx, seq_idx, s_in, d]
transformed_offset = (
batch_idx * seq_len * num_streams * stream_dim
+ seq_idx * num_streams * stream_dim
+ s_in * stream_dim
+ d
)
transformed_val = tl.load(
transformed_ptr + transformed_offset,
mask=mask,
other=0.0
)
pre_acc += h_pre_val * transformed_val
# Now multiply by H_post[s_mid, s_out] and accumulate
h_post_val = tl.load(
H_post_ptr + s_mid * num_streams + s_out
)
post_acc += h_post_val * pre_acc
# Step 4: Combine residual and post-mixed
output_val = residual_acc + post_acc
# Store output[batch_idx, seq_idx, s_out, d]
output_offset = (
batch_idx * seq_len * num_streams * stream_dim
+ seq_idx * num_streams * stream_dim
+ s_out * stream_dim
+ d
)
tl.store(
output_ptr + output_offset,
output_val,
mask=mask
)
def fused_width_connection_triton(
x: torch.Tensor,
transformed: torch.Tensor,
H_res: torch.Tensor,
H_pre: torch.Tensor,
H_post: torch.Tensor,
) -> torch.Tensor:
"""
Fused Triton implementation of width_connection.
Args:
x: Residual input (batch, seq_len, num_streams, stream_dim)
transformed: Transformed input (batch, seq_len, num_streams, stream_dim)
H_res: Doubly stochastic matrix (num_streams, num_streams)
H_pre: Pre-mixing matrix (num_streams, num_streams)
H_post: Post-mixing matrix (num_streams, num_streams)
Returns:
output: Mixed features (batch, seq_len, num_streams, stream_dim)
"""
batch_size, seq_len, num_streams, stream_dim = x.shape
# Validate inputs
assert x.is_cuda, "Input must be on GPU"
assert x.is_contiguous(), "Input must be contiguous"
assert transformed.shape == x.shape, "Shape mismatch"
assert H_res.shape == (num_streams, num_streams), "H_res shape mismatch"
assert H_pre.shape == (num_streams, num_streams), "H_pre shape mismatch"
assert H_post.shape == (num_streams, num_streams), "H_post shape mismatch"
# Allocate output
output = torch.empty_like(x)
# Launch kernel
total_positions = batch_size * seq_len
BLOCK_SIZE = 128
grid = (triton.cdiv(total_positions, BLOCK_SIZE),)
fused_stream_mixing_kernel[grid](
x, transformed, H_res, H_pre, H_post, output,
batch_size, seq_len, num_streams, stream_dim,
BLOCK_SIZE=BLOCK_SIZE,
)
return output
def compare_with_einops_reference(
x: torch.Tensor,
transformed: torch.Tensor,
H_res: torch.Tensor,
H_pre: torch.Tensor,
H_post: torch.Tensor,
atol: float = 1e-5,
rtol: float = 1e-4,
) -> tuple[bool, float, Optional[str]]:
"""
Compare Triton kernel output with einops reference implementation.
Args:
x: Residual input (batch, seq_len, num_streams, stream_dim)
transformed: Transformed input (batch, seq_len, num_streams, stream_dim)
H_res: Doubly stochastic matrix (num_streams, num_streams)
H_pre: Pre-mixing matrix (num_streams, num_streams)
H_post: Post-mixing matrix (num_streams, num_streams)
atol: Absolute tolerance for comparison
rtol: Relative tolerance for comparison
Returns:
matches: True if outputs match within tolerance
max_diff: Maximum absolute difference
error_msg: Error message if mismatch, None otherwise
"""
from einops import einsum # type: ignore[import-not-found]
# Triton kernel output
triton_output = fused_width_connection_triton(
x, transformed, H_res, H_pre, H_post
)
# Reference einops implementation (from hyper_connections.py)
residual_mixed = einsum(
H_res, x, "n m, b s n d -> b s m d"
)
pre_mixed = einsum(
H_pre, transformed, "n m, b s n d -> b s m d"
)
post_mixed = einsum(
H_post, pre_mixed, "m n, b s m d -> b s n d"
)
einops_output = residual_mixed + post_mixed
# Compare
matches = torch.allclose(triton_output, einops_output, atol=atol, rtol=rtol)
max_diff = (triton_output - einops_output).abs().max().item()
error_msg = None
if not matches:
mean_diff = (triton_output - einops_output).abs().mean().item()
error_msg = (
f"Triton kernel output does not match einops reference!\n"
f"Max diff: {max_diff:.6e} (atol={atol}, rtol={rtol})\n"
f"Mean diff: {mean_diff:.6e}\n"
f"Triton output range: [{triton_output.min().item():.4f}, "
f"{triton_output.max().item():.4f}]\n"
f"Einops output range: [{einops_output.min().item():.4f}, "
f"{einops_output.max().item():.4f}]"
)
return matches, max_diff, error_msg
def benchmark_kernel_speedup(
batch_size: int = 16,
seq_len: int = 128,
dim: int = 512,
num_streams: int = 8,
num_warmup: int = 10,
num_iters: int = 100,
) -> tuple[float, float, float]:
"""
Benchmark Triton kernel vs einops reference.
Args:
batch_size: Batch size
seq_len: Sequence length
dim: Hidden dimension
num_streams: Number of streams
num_warmup: Warmup iterations
num_iters: Benchmark iterations
Returns:
einops_time_ms: Average einops time (milliseconds)
triton_time_ms: Average Triton time (milliseconds)
speedup: Speedup factor (einops_time / triton_time)
"""
from einops import einsum # type: ignore[import-not-found]
import time
assert torch.cuda.is_available(), "GPU required for benchmarking"
device = torch.device('cuda')
stream_dim = dim // num_streams
# Generate random inputs
x = torch.randn(batch_size, seq_len, num_streams, stream_dim, device=device)
transformed = torch.randn(batch_size, seq_len, num_streams, stream_dim, device=device)
H_res = torch.randn(num_streams, num_streams, device=device).softmax(dim=-1)
H_pre = torch.randn(num_streams, num_streams, device=device).softmax(dim=-1)
H_post = torch.randn(num_streams, num_streams, device=device).softmax(dim=-1)
# Warmup
for _ in range(num_warmup):
_ = fused_width_connection_triton(x, transformed, H_res, H_pre, H_post)
residual_mixed = einsum(H_res, x, "n m, b s n d -> b s m d")
pre_mixed = einsum(H_pre, transformed, "n m, b s n d -> b s m d")
post_mixed = einsum(H_post, pre_mixed, "m n, b s m d -> b s n d")
_ = residual_mixed + post_mixed
torch.cuda.synchronize()
# Benchmark einops
start = time.perf_counter()
for _ in range(num_iters):
residual_mixed = einsum(H_res, x, "n m, b s n d -> b s m d")
pre_mixed = einsum(H_pre, transformed, "n m, b s n d -> b s m d")
post_mixed = einsum(H_post, pre_mixed, "m n, b s m d -> b s n d")
output_einops = residual_mixed + post_mixed
torch.cuda.synchronize()
einops_time = (time.perf_counter() - start) / num_iters * 1000 # ms
# Benchmark Triton
start = time.perf_counter()
for _ in range(num_iters):
output_triton = fused_width_connection_triton(
x, transformed, H_res, H_pre, H_post
)
torch.cuda.synchronize()
triton_time = (time.perf_counter() - start) / num_iters * 1000 # ms
speedup = einops_time / triton_time
return einops_time, triton_time, speedup
if __name__ == "__main__":
"""Quick test of fused kernel correctness."""
print("Testing fused mHC Triton kernel...")
if not torch.cuda.is_available():
print("CUDA not available, skipping test")
exit(0)
device = torch.device('cuda')
# Test configuration
batch_size = 4
seq_len = 32
num_streams = 8
stream_dim = 64
# Generate test inputs
x = torch.randn(batch_size, seq_len, num_streams, stream_dim, device=device)
transformed = torch.randn(batch_size, seq_len, num_streams, stream_dim, device=device)
H_res = torch.randn(num_streams, num_streams, device=device).softmax(dim=-1)
H_pre = torch.randn(num_streams, num_streams, device=device).softmax(dim=-1)
H_post = torch.randn(num_streams, num_streams, device=device).softmax(dim=-1)
# Compare with reference
matches, max_diff, error_msg = compare_with_einops_reference(
x, transformed, H_res, H_pre, H_post
)
if matches:
print(f"✓ Correctness test PASSED (max diff: {max_diff:.6e})")
else:
print(f"✗ Correctness test FAILED")
print(error_msg)
exit(1)
# Benchmark
print("\nBenchmarking...")
einops_time, triton_time, speedup = benchmark_kernel_speedup()
print(f"Einops time: {einops_time:.3f} ms")
print(f"Triton time: {triton_time:.3f} ms")
print(f"Speedup: {speedup:.2f}x")
print("\n✓ All tests passed!")
|