File size: 10,489 Bytes
5fdc1d1 | 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 | """Benchmarks for the MaxSim kernel against naive PyTorch baselines.
This is the file the ``kernels`` CLI discovers and runs::
kernels benchmark erikkaum/maxsim # against the published kernel
just bench-local # against the local ./build
The workloads mirror ``scripts/cuda_bench_matrix.py`` (the README-number
generator) so the CLI tables and the README matrix describe the same shapes:
* Contrastive training -- the headline ColBERT fine-tuning step (forward +
backward), at the in-batch shape an actual PyLate batch hits.
* Padded inference -- the second-stage exact rerank (forward only).
* Packed inference -- the same rerank shape through the ragged pair API,
to surface the packing/layout overhead.
Naive baselines are identical to the matrix's: a ``torch.einsum`` that
materialises the full similarity tensor before ``max`` -- what the kernel is
meant to beat. fp16 inputs throughout; the matrix covers bf16 as well.
"""
from __future__ import annotations
import torch
from kernels.benchmark import Benchmark
_DTYPE = torch.float16
_DIM = 128
_SEED = 1234
# ---------------------------------------------------------------------------
# Naive PyTorch baselines (identical to scripts/cuda_bench_matrix.py).
# ---------------------------------------------------------------------------
def _naive_contrastive(q, docs, cu, d_lens):
"""All-pairs MaxSim over packed docs: materialises [Nq, Nb, Lq, Ld]."""
_Nq, _Lq, dim = q.shape
Nb = len(d_lens)
Ld_max = max(d_lens)
docs_padded = docs.new_zeros((Nb, Ld_max, dim))
offs = cu.to(torch.int64).cpu().tolist()
for i, ld_i in enumerate(d_lens):
docs_padded[i, :ld_i] = docs[offs[i] : offs[i + 1]]
sim = torch.einsum("qid,njd->qnij", q.float(), docs_padded.float())
return sim.max(dim=-1).values.sum(dim=-1)
def _naive_padded(q, d, qlen, dlen):
"""Per-query padded MaxSim: materialises [B, C, Lq, Ld]."""
_B, _C, Ld, _dim = d.shape
Lq = q.shape[1]
sim = torch.einsum("bid,bcjd->bcij", q.float(), d.float())
q_mask = torch.arange(Lq, device=q.device)[None, :] < qlen[:, None]
d_mask = torch.arange(Ld, device=q.device)[None, None, :] < dlen[:, :, None]
sim = sim.masked_fill(~d_mask[:, :, None, :], float("-inf"))
per_q_max = sim.max(dim=-1).values
per_q_max = per_q_max.masked_fill(~q_mask[:, None, :], 0.0)
return per_q_max.sum(dim=-1)
# ---------------------------------------------------------------------------
# Workload construction (shared, keyed off each Benchmark's shape attrs).
# ---------------------------------------------------------------------------
def _make_contrastive(self: Benchmark, Nq, Nb, Lq, Ld) -> None:
gen = torch.Generator().manual_seed(self.seed)
self.queries = torch.randn(Nq, Lq, _DIM, generator=gen, dtype=_DTYPE).to(self.device)
self.documents = torch.randn(Nb * Ld, _DIM, generator=gen, dtype=_DTYPE).to(self.device)
self.document_offsets = torch.arange(
0, (Nb + 1) * Ld, Ld, dtype=torch.int32, device=self.device
)
self.d_lens = [Ld] * Nb
def _make_padded(self: Benchmark, B, C, Lq, Ld) -> None:
gen = torch.Generator().manual_seed(self.seed)
self.queries = torch.randn(B, Lq, _DIM, generator=gen, dtype=_DTYPE).to(self.device)
self.documents = torch.randn(B, C, Ld, _DIM, generator=gen, dtype=_DTYPE).to(self.device)
self.query_lengths = torch.full((B,), Lq, dtype=torch.int32, device=self.device)
self.doc_lengths = torch.full((B, C), Ld, dtype=torch.int32, device=self.device)
def _make_packed(self: Benchmark, B, C, Lq, Ld) -> None:
"""Padded tensors plus a flattened CSR pair grid expressing the same work."""
_make_padded(self, B, C, Lq, Ld)
q, d = self.queries, self.documents
self.max_q_len = Lq
self.batch, self.candidates = B, C
self.q_flat = q.reshape(B * Lq, _DIM).contiguous()
self.d_flat = d.reshape(B * C * Ld, _DIM).contiguous()
self.q_offsets = torch.arange(0, (B + 1) * Lq, Lq, dtype=torch.int32, device=q.device)
self.d_offsets = torch.arange(0, (B * C + 1) * Ld, Ld, dtype=torch.int32, device=q.device)
pair_ids = torch.arange(B * C, dtype=torch.int32, device=q.device)
self.pair_query_ids = pair_ids // C
self.pair_document_ids = pair_ids
# ---------------------------------------------------------------------------
# Kernel / naive runners.
# ---------------------------------------------------------------------------
def _contrastive_train_kernel(self: Benchmark) -> torch.Tensor:
q = self.queries.detach().clone().requires_grad_(True)
d = self.documents.detach().clone().requires_grad_(True)
scores = self.kernel.score_contrastive_train(q, d, self.document_offsets)
scores.sum().backward()
return scores.detach()
def _contrastive_train_naive(self: Benchmark) -> torch.Tensor:
q = self.queries.detach().clone().requires_grad_(True)
d = self.documents.detach().clone().requires_grad_(True)
scores = _naive_contrastive(q, d, self.document_offsets, self.d_lens)
scores.sum().backward()
return scores.detach()
def _contrastive_ref(self: Benchmark) -> torch.Tensor:
return _naive_contrastive(
self.queries, self.documents, self.document_offsets, self.d_lens
)
def _padded_kernel(self: Benchmark) -> torch.Tensor:
return self.kernel.score_candidates_padded(
self.queries, self.documents, self.query_lengths, self.doc_lengths
)
def _padded_naive(self: Benchmark) -> torch.Tensor:
return _naive_padded(
self.queries, self.documents, self.query_lengths, self.doc_lengths
)
def _packed_kernel(self: Benchmark) -> torch.Tensor:
return self.kernel.score_pairs_packed(
self.q_flat,
self.q_offsets,
self.d_flat,
self.d_offsets,
self.pair_query_ids,
self.pair_document_ids,
max_q_len=self.max_q_len,
).view(self.batch, self.candidates)
# ---------------------------------------------------------------------------
# Contrastive training (forward + backward; the headline workload).
# ---------------------------------------------------------------------------
class ContrastiveLateOn(Benchmark):
"""In-batch contrastive training: Nq=Nb=32, Lq=32, Ld=80, dim=128."""
seed = _SEED
def setup(self) -> None:
_make_contrastive(self, Nq=32, Nb=32, Lq=32, Ld=80)
def benchmark_kernel(self) -> None:
self.out = _contrastive_train_kernel(self)
def benchmark_naive(self) -> None:
self.out = _contrastive_train_naive(self)
def verify_kernel(self) -> torch.Tensor:
return _contrastive_ref(self)
def verify_naive(self) -> torch.Tensor:
return _contrastive_ref(self)
class ContrastiveLongDocs(Benchmark):
"""Same in-batch shape but long docs (Ld=512) -- stresses retained state."""
seed = _SEED
def setup(self) -> None:
_make_contrastive(self, Nq=32, Nb=32, Lq=32, Ld=512)
def benchmark_kernel(self) -> None:
self.out = _contrastive_train_kernel(self)
def benchmark_naive(self) -> None:
self.out = _contrastive_train_naive(self)
def verify_kernel(self) -> torch.Tensor:
return _contrastive_ref(self)
def verify_naive(self) -> torch.Tensor:
return _contrastive_ref(self)
class ContrastiveBigBatch(Benchmark):
"""Doubled in-batch batch size: Nq=Nb=64, Lq=32, Ld=128, dim=128."""
seed = _SEED
def setup(self) -> None:
_make_contrastive(self, Nq=64, Nb=64, Lq=32, Ld=128)
def benchmark_kernel(self) -> None:
self.out = _contrastive_train_kernel(self)
def benchmark_naive(self) -> None:
self.out = _contrastive_train_naive(self)
def verify_kernel(self) -> torch.Tensor:
return _contrastive_ref(self)
def verify_naive(self) -> torch.Tensor:
return _contrastive_ref(self)
# ---------------------------------------------------------------------------
# Padded inference (second-stage rerank; forward only).
# ---------------------------------------------------------------------------
class PaddedRerank(Benchmark):
"""Padded rerank at a typical inference shape: B=32, K=50, Ld=180."""
seed = _SEED
def setup(self) -> None:
_make_padded(self, B=32, C=50, Lq=32, Ld=180)
def benchmark_kernel(self) -> None:
self.out = _padded_kernel(self)
def benchmark_naive(self) -> None:
self.out = _padded_naive(self)
def verify_kernel(self) -> torch.Tensor:
return _padded_naive(self)
def verify_naive(self) -> torch.Tensor:
return _padded_naive(self)
class PaddedHeavyRerank(Benchmark):
"""Padded rerank at K=100 candidates, Ld=256 -- larger compute envelope."""
seed = _SEED
def setup(self) -> None:
_make_padded(self, B=32, C=100, Lq=32, Ld=256)
def benchmark_kernel(self) -> None:
self.out = _padded_kernel(self)
def benchmark_naive(self) -> None:
self.out = _padded_naive(self)
def verify_kernel(self) -> torch.Tensor:
return _padded_naive(self)
def verify_naive(self) -> torch.Tensor:
return _padded_naive(self)
# ---------------------------------------------------------------------------
# Packed inference (same rerank shape via the ragged pair API; forward only).
# ---------------------------------------------------------------------------
class PackedRerank(Benchmark):
"""Rerank shape (B=32, K=50, Ld=180) expressed through the packed pair API."""
seed = _SEED
def setup(self) -> None:
_make_packed(self, B=32, C=50, Lq=32, Ld=180)
def benchmark_kernel(self) -> None:
self.out = _packed_kernel(self)
def benchmark_naive(self) -> None:
self.out = _padded_naive(self)
def verify_kernel(self) -> torch.Tensor:
return _padded_naive(self)
def verify_naive(self) -> torch.Tensor:
return _padded_naive(self)
class PackedHeavyRerank(Benchmark):
"""Rerank shape (B=32, K=100, Ld=256) through the packed pair API."""
seed = _SEED
def setup(self) -> None:
_make_packed(self, B=32, C=100, Lq=32, Ld=256)
def benchmark_kernel(self) -> None:
self.out = _packed_kernel(self)
def benchmark_naive(self) -> None:
self.out = _padded_naive(self)
def verify_kernel(self) -> torch.Tensor:
return _padded_naive(self)
def verify_naive(self) -> torch.Tensor:
return _padded_naive(self)
|