File size: 8,390 Bytes
b78342b | 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 | """Fused SM80/SM120 callables reuse dynamic text extents and Top-K ratios."""
import pytest
import torch
from h3_sparse_attention.sol_numerator_virtual_q import (
_dynamic_tensor_cache_signature,
)
def test_dynamic_signature_ignores_non_singleton_extents_and_compact_strides():
first = torch.empty((1, 73560, 56, 128), dtype=torch.bfloat16)
second = torch.empty((1, 73583, 56, 128), dtype=torch.bfloat16)
assert _dynamic_tensor_cache_signature(first) == _dynamic_tensor_cache_signature(second)
def test_dynamic_signature_preserves_static_abi_properties():
base = torch.empty((1, 128, 4, 128), dtype=torch.bfloat16)
different_dtype = torch.empty((1, 128, 4, 128), dtype=torch.float32)
different_order = torch.empty((1, 4, 128, 128), dtype=torch.bfloat16).permute(0, 2, 1, 3)
different_singletons = torch.empty((1, 128, 1, 128), dtype=torch.bfloat16)
signature = _dynamic_tensor_cache_signature(base)
assert signature != _dynamic_tensor_cache_signature(different_dtype)
assert signature != _dynamic_tensor_cache_signature(different_order)
assert signature != _dynamic_tensor_cache_signature(different_singletons)
@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")
@pytest.mark.parametrize(
"execution",
("threshold", "fused", "packed_external", "packed_external_no_route_qk"),
)
def test_sm120_compiled_callable_reuses_different_sink_lengths(execution):
if torch.cuda.get_device_capability() != (12, 0):
pytest.skip("SM120 required")
import h3_sparse_attention.sol_numerator_virtual_q as fused
torch.manual_seed(813)
video_tokens = 8192
def inputs(total_tokens):
heads = 2
blocks = (total_tokens + 63) // 64
q, k, v = [
torch.randn(
1,
total_tokens,
heads,
128,
device="cuda",
dtype=torch.bfloat16,
)
for _ in range(3)
]
ranges = torch.tensor(
[[0, video_tokens], [video_tokens, total_tokens]],
device="cuda",
dtype=torch.int64,
)
mapping = torch.tensor(
[0] * (video_tokens // 64)
+ [1] * (blocks - video_tokens // 64),
device="cuda",
dtype=torch.int64,
)
anchors = fused.build_virtual_anchors(q, ranges)
centroids = fused.reduce_virtual_key_centroids(k)
threshold = (
torch.zeros((1, blocks, heads), device="cuda", dtype=torch.float32)
if execution == "threshold"
else None
)
route = None
if execution.startswith("packed_external"):
route = torch.full(
(1, video_tokens // 64, heads, (blocks + 31) // 32),
-1,
device="cuda",
dtype=torch.int32,
)
return q, k, v, anchors, ranges, mapping, centroids, threshold, route
options = {
"force_local_blocks": False,
"_query_tokens": video_tokens,
"fused_topk_ratio": 0.1 if execution == "fused" else 0.0,
"skip_external_route_qk": execution == "packed_external_no_route_qk",
}
first = inputs(video_tokens + 64)
second = inputs(video_tokens + 81)
fused._FUSED_COMPILED.clear()
fused._FUSED_COMPILE_CALLS = 0
fused._FUSED_COMPILE_SECONDS = 0.0
fused._fused_virtual(
*first,
video_tokens,
64,
**options,
)
reused = fused._fused_virtual(
*second,
video_tokens,
81,
**options,
)[:, :video_tokens].clone()
reused_stats = fused.fused_compile_cache_stats()
assert reused_stats["entries"] == 1
assert reused_stats["compile_calls"] == 1
assert torch.isfinite(reused).all()
fused._FUSED_COMPILED.clear()
fresh = fused._fused_virtual(
*second,
video_tokens,
81,
**options,
)[:, :video_tokens]
torch.cuda.synchronize()
assert torch.equal(reused, fresh)
@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")
def test_sm120_compiled_callable_reuses_different_fused_topk_ratios(monkeypatch):
if torch.cuda.get_device_capability() != (12, 0):
pytest.skip("SM120 required")
import h3_sparse_attention.sol_numerator_virtual_q as fused
monkeypatch.setenv("H3_SM120_RUNTIME_TOPK_RATIO", "1")
torch.manual_seed(814)
video_tokens = 8192
sink_tokens = 64
total_tokens = video_tokens + sink_tokens
heads = 2
blocks = (total_tokens + 63) // 64
q, k, v = [
torch.randn(
1, total_tokens, heads, 128, device="cuda", dtype=torch.bfloat16
)
for _ in range(3)
]
ranges = torch.tensor(
[[0, video_tokens], [video_tokens, total_tokens]],
device="cuda",
dtype=torch.int64,
)
mapping = torch.tensor(
[0] * (video_tokens // 64) + [1] * (blocks - video_tokens // 64),
device="cuda",
dtype=torch.int64,
)
anchors = fused.build_virtual_anchors(q, ranges)
centroids = fused.reduce_virtual_key_centroids(k)
def run(ratio):
return fused._fused_virtual(
q, k, v, anchors, ranges, mapping, centroids, None, None,
video_tokens, sink_tokens,
force_local_blocks=False,
_query_tokens=video_tokens,
fused_topk_ratio=ratio,
)[:, :video_tokens].clone()
fused._FUSED_COMPILED.clear()
fused._FUSED_COMPILE_CALLS = 0
fused._FUSED_COMPILE_SECONDS = 0.0
first = run(0.1)
reused = run(0.2)
reused_stats = fused.fused_compile_cache_stats()
assert reused_stats["entries"] == 1
assert reused_stats["compile_calls"] == 1
assert not torch.equal(first, reused)
fused._FUSED_COMPILED.clear()
fresh = run(0.2)
torch.cuda.synchronize()
assert torch.equal(reused, fresh)
monkeypatch.setenv("H3_SM120_RUNTIME_TOPK_RATIO", "0")
fused._FUSED_COMPILED.clear()
static = run(0.2)
torch.cuda.synchronize()
assert torch.equal(reused, static)
@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")
def test_sm80_compiled_callable_reuses_text_lengths_and_topk_ratios(monkeypatch):
if torch.cuda.get_device_capability() != (8, 0):
pytest.skip("SM80 required")
import h3_sparse_attention.sol_numerator_virtual_q as fused
monkeypatch.setenv("H3_SM80_RUNTIME_TOPK_RATIO", "1")
torch.manual_seed(915)
video_tokens = 512
def inputs(sink_tokens):
total_tokens = video_tokens + sink_tokens
blocks = (total_tokens + 63) // 64
q, k, v = [
torch.randn(1, total_tokens, 2, 128, device="cuda", dtype=torch.bfloat16)
for _ in range(3)
]
ranges = torch.tensor(
[[0, video_tokens], [video_tokens, total_tokens]],
device="cuda", dtype=torch.int64,
)
mapping = torch.tensor(
[0] * (video_tokens // 64) + [1] * (blocks - video_tokens // 64),
device="cuda", dtype=torch.int64,
)
return (
q, k, v, fused.build_virtual_anchors(q, ranges), ranges, mapping,
fused.reduce_virtual_key_centroids(k),
)
def run(data, sink_tokens, ratio):
return fused._fused_virtual(
*data, None, None, video_tokens, sink_tokens,
force_local_blocks=False, _query_tokens=video_tokens,
fused_topk_ratio=ratio,
)[:, :video_tokens].clone()
shorter = inputs(64)
longer = inputs(81)
fused._FUSED_COMPILED.clear()
fused._FUSED_COMPILE_CALLS = 0
run(shorter, 64, 0.1)
first_ratio = run(longer, 81, 0.1)
second_ratio = run(longer, 81, 0.2)
assert fused.fused_compile_cache_stats()["entries"] == 1
assert fused.fused_compile_cache_stats()["compile_calls"] == 1
assert not torch.equal(first_ratio, second_ratio)
with pytest.raises(ValueError, match="fused_topk_ratio"):
run(longer, 81, 1.01)
fused._FUSED_COMPILED.clear()
fresh = run(longer, 81, 0.2)
monkeypatch.setenv("H3_SM80_RUNTIME_TOPK_RATIO", "0")
fused._FUSED_COMPILED.clear()
static = run(longer, 81, 0.2)
torch.cuda.synchronize()
assert torch.equal(second_ratio, fresh)
assert torch.equal(second_ratio, static)
|