File size: 13,391 Bytes
2ece486 | 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 | """
ENGRAM Protocol — Compression Tests
Tests for kvcos.core.compression:
- FP16 passthrough
- Q8_0 round-trip accuracy & shape preservation
- PolarQuant round-trip accuracy & rotation invariants
- Dispatcher routing and Q4_0 fallback warning
- Edge cases: padding, single-element groups
"""
from __future__ import annotations
import warnings
import pytest
import torch
from kvcos.core.compression import (
Q8_GROUP_SIZE,
CompressionResult,
compress,
compress_fp16,
compress_polarquant,
compress_q8_0,
decompress,
decompress_fp16,
decompress_polarquant,
decompress_q8_0,
)
from kvcos.core.types import CompressionMethod
# ── FP16 Passthrough ──────────────────────────────────────────────────────────
class TestFP16:
"""FP16 passthrough: no quantization, just dtype normalization."""
def test_fp16_passthrough_shape(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
keys, _ = llama_kv_256
result = compress_fp16(keys)
assert result.data.shape == keys.shape
def test_fp16_passthrough_dtype(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
keys, _ = llama_kv_256
result = compress_fp16(keys)
assert result.data.dtype == torch.float16
def test_fp16_passthrough_exact(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
keys, _ = llama_kv_256
result = compress_fp16(keys)
assert torch.equal(result.data, keys.to(torch.float16))
def test_fp16_compression_ratio_one(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
keys, _ = llama_kv_256
result = compress_fp16(keys)
assert result.compression_ratio == 1.0
def test_fp16_method_tag(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
keys, _ = llama_kv_256
result = compress_fp16(keys)
assert result.method == CompressionMethod.FP16
def test_fp16_from_fp32(self) -> None:
"""FP32 input is cast to FP16."""
t = torch.randn(4, 8, 32, 128, dtype=torch.float32)
result = compress_fp16(t)
assert result.data.dtype == torch.float16
assert result.original_dtype == torch.float32
def test_fp16_decompress_identity(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
keys, _ = llama_kv_256
result = compress_fp16(keys)
out = decompress_fp16(result.data)
assert torch.equal(out, result.data)
# ── Q8_0 Quantization ────────────────────────────────────────────────────────
class TestQ8_0:
"""Q8_0: group quantization matching llama.cpp GGML_TYPE_Q8_0."""
def test_q8_0_shape_preserved(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
keys, _ = llama_kv_256
result = compress_q8_0(keys)
assert result.data.shape == keys.shape
def test_q8_0_output_dtype(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
"""Q8_0 stores dequantized bfloat16 for safetensors compat."""
keys, _ = llama_kv_256
result = compress_q8_0(keys)
assert result.data.dtype == torch.bfloat16
def test_q8_0_method_tag(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
keys, _ = llama_kv_256
result = compress_q8_0(keys)
assert result.method == CompressionMethod.Q8_0
def test_q8_0_metadata_group_size(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
keys, _ = llama_kv_256
result = compress_q8_0(keys)
assert result.metadata["q8_group_size"] == str(Q8_GROUP_SIZE)
def test_q8_0_round_trip_low_error(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
"""Q8_0 quantization error should be < 1% relative MSE."""
keys, _ = llama_kv_256
result = compress_q8_0(keys)
decompressed = decompress_q8_0(result.data)
original = keys.float()
restored = decompressed.float()
mse = ((original - restored) ** 2).mean()
signal_power = (original**2).mean()
relative_mse = (mse / signal_power).item()
assert relative_mse < 0.01, f"Q8_0 relative MSE {relative_mse:.6f} > 1%"
def test_q8_0_round_trip_values(
self, phi3_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
"""Q8_0 round-trip on Phi-3 (head_dim=96, needs padding)."""
keys, values = phi3_kv_256
for tensor in (keys, values):
result = compress_q8_0(tensor)
assert result.data.shape == tensor.shape
def test_q8_0_compression_ratio_fp32(self) -> None:
"""FP32 input → bfloat16 output gives 2x compression ratio."""
t = torch.randn(2, 4, 64, 128, dtype=torch.float32)
result = compress_q8_0(t)
assert abs(result.compression_ratio - 2.0) < 0.01
def test_q8_0_compression_ratio_fp16(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
"""FP16 input → bfloat16 output gives 1x ratio (same byte width)."""
keys, _ = llama_kv_256
result = compress_q8_0(keys)
assert abs(result.compression_ratio - 1.0) < 0.01
def test_q8_0_preserves_original_dtype(self) -> None:
t = torch.randn(4, 8, 32, 128, dtype=torch.float32)
result = compress_q8_0(t)
assert result.original_dtype == torch.float32
def test_q8_0_padding_dim_not_divisible(self) -> None:
"""Head dims not divisible by 32 get padded then unpadded."""
t = torch.randn(2, 4, 16, 96, dtype=torch.float16) # 96 = 3*32, exact
result = compress_q8_0(t)
assert result.data.shape == t.shape
t2 = torch.randn(2, 4, 16, 100, dtype=torch.float16) # 100 not div by 32
result2 = compress_q8_0(t2)
assert result2.data.shape == t2.shape
def test_q8_0_zero_tensor(self) -> None:
"""All-zero tensor should round-trip exactly."""
t = torch.zeros(2, 4, 16, 128, dtype=torch.float16)
result = compress_q8_0(t)
decompressed = decompress_q8_0(result.data)
assert torch.allclose(decompressed, t.to(torch.float16), atol=1e-6)
# ── PolarQuant ───────────────────────────────────────────────────────────────
class TestPolarQuant:
"""PolarQuant: MSE-optimal random rotation + Lloyd-Max at 3 bits.
QJL intentionally absent (D5).
"""
def test_polarquant_shape_preserved(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
keys, _ = llama_kv_256
result = compress_polarquant(keys)
assert result.data.shape == keys.shape
def test_polarquant_output_dtype(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
keys, _ = llama_kv_256
result = compress_polarquant(keys)
assert result.data.dtype == torch.bfloat16
def test_polarquant_method_tag(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
keys, _ = llama_kv_256
result = compress_polarquant(keys)
assert result.method == CompressionMethod.POLARQUANT
def test_polarquant_metadata_qjl_disabled(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
"""D5: QJL must be marked disabled in metadata."""
keys, _ = llama_kv_256
result = compress_polarquant(keys)
assert result.metadata["qjl_enabled"] == "false"
assert result.metadata["polarquant_bits"] == "3"
def test_polarquant_round_trip_bounded_error(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
"""PolarQuant 3-bit error should be < 15% relative MSE.
3-bit Lloyd-Max on rotated Gaussian: theoretical ~10% for 8 centroids.
Allow margin for rotation + dtype casting.
"""
keys, _ = llama_kv_256
result = compress_polarquant(keys)
decompressed = decompress_polarquant(result.data)
original = keys.float()
restored = decompressed.float()
mse = ((original - restored) ** 2).mean()
signal_power = (original**2).mean()
relative_mse = (mse / signal_power).item()
assert relative_mse < 0.15, f"PolarQuant relative MSE {relative_mse:.4f} > 15%"
def test_polarquant_worse_than_q8_0(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
"""3-bit PolarQuant should have higher error than 8-bit Q8_0."""
keys, _ = llama_kv_256
original = keys.float()
q8_result = compress_q8_0(keys)
pq_result = compress_polarquant(keys)
q8_mse = ((original - decompress_q8_0(q8_result.data).float()) ** 2).mean()
pq_mse = (
(original - decompress_polarquant(pq_result.data).float()) ** 2
).mean()
assert pq_mse > q8_mse, "PolarQuant 3-bit should be less accurate than Q8_0"
def test_polarquant_deterministic(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
"""Same input → same output (fixed seed rotation matrix)."""
keys, _ = llama_kv_256
r1 = compress_polarquant(keys)
r2 = compress_polarquant(keys)
assert torch.equal(r1.data, r2.data)
def test_polarquant_phi3_shape(
self, phi3_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
"""Phi-3 head_dim=96 works with PolarQuant."""
keys, _ = phi3_kv_256
result = compress_polarquant(keys)
assert result.data.shape == keys.shape
# ── Dispatcher ───────────────────────────────────────────────────────────────
class TestDispatcher:
"""compress() and decompress() dispatch to correct implementations."""
@pytest.mark.parametrize(
"method",
[CompressionMethod.FP16, CompressionMethod.Q8_0, CompressionMethod.POLARQUANT],
)
def test_compress_dispatches(self, method: CompressionMethod) -> None:
t = torch.randn(2, 4, 16, 128, dtype=torch.float16)
result = compress(t, method)
assert isinstance(result, CompressionResult)
assert result.method == method
@pytest.mark.parametrize(
"method",
[CompressionMethod.FP16, CompressionMethod.Q8_0, CompressionMethod.POLARQUANT],
)
def test_decompress_returns_fp16(self, method: CompressionMethod) -> None:
t = torch.randn(2, 4, 16, 128, dtype=torch.float16)
result = compress(t, method)
out = decompress(result.data, method)
assert out.dtype == torch.float16
def test_q4_0_warns_and_falls_back(self) -> None:
"""D5: Q4_0 emits warning and uses Q8_0 instead."""
t = torch.randn(2, 4, 16, 128, dtype=torch.float16)
with warnings.catch_warnings(record=True) as w:
warnings.simplefilter("always")
result = compress(t, CompressionMethod.Q4_0)
assert len(w) == 1
assert "Q4_0" in str(w[0].message)
assert "92%" in str(w[0].message)
assert result.method == CompressionMethod.Q8_0
def test_unknown_method_raises(self) -> None:
t = torch.randn(2, 4, 16, 128, dtype=torch.float16)
with pytest.raises(ValueError, match="Unknown compression method"):
compress(t, "invalid_method") # type: ignore[arg-type]
def test_decompress_unknown_raises(self) -> None:
t = torch.randn(2, 4, 16, 128, dtype=torch.float16)
with pytest.raises(ValueError, match="Unknown compression method"):
decompress(t, "invalid_method") # type: ignore[arg-type]
# ── Round-trip Integration ───────────────────────────────────────────────────
class TestRoundTrip:
"""Full compress → decompress round-trip through dispatcher."""
@pytest.mark.parametrize(
"method",
[CompressionMethod.FP16, CompressionMethod.Q8_0, CompressionMethod.POLARQUANT],
)
def test_round_trip_shape_preserved(self, method: CompressionMethod) -> None:
t = torch.randn(4, 8, 64, 128, dtype=torch.float16)
result = compress(t, method)
out = decompress(result.data, method)
assert out.shape == t.shape
def test_round_trip_both_kv(
self, llama_kv_256: tuple[torch.Tensor, torch.Tensor]
) -> None:
"""Compress and decompress both keys and values."""
keys, values = llama_kv_256
for tensor in (keys, values):
for method in (CompressionMethod.FP16, CompressionMethod.Q8_0):
result = compress(tensor, method)
out = decompress(result.data, method)
assert out.shape == tensor.shape
assert out.dtype == torch.float16
|