File size: 12,579 Bytes
e82122f | 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 | #!/usr/bin/env python3
"""
Unit tests for NVFP4 kernel functions.
This tests dequantization and GEMM operations in isolation before
attempting full model inference.
"""
import sys
import torch
import torch.nn.functional as F
# Import from local inference directory
from nvfp4_kernel import (
dequantize_nvfp4,
nvfp4_gemm_dequant,
NVFP4_LUT,
NVFP4_BLOCK_SIZE
)
# Constants from quantization script
FP4_MAX = 6.0
FP8_E4M3_MAX = 448.0
E2M1_BOUNDS = torch.tensor([0.25, 0.75, 1.25, 1.75, 2.5, 3.5, 5.0], dtype=torch.float32)
def compute_nvfp4_scales(fp32_weight, block_size=16):
"""
Compute two-level NVFP4 scaling factors.
Simplified version for testing.
"""
# Global scale
global_amax = fp32_weight.abs().max()
weight_scale_2 = global_amax / (FP4_MAX * FP8_E4M3_MAX)
if weight_scale_2.abs() < 1e-10:
weight_scale_2 = torch.tensor(1e-8, dtype=torch.float32, device=fp32_weight.device)
# Per-block scale
M = fp32_weight.shape[0] if fp32_weight.dim() > 1 else 1
N = fp32_weight.shape[-1]
# Pad if needed
N_padded = ((N + block_size - 1) // block_size) * block_size
if N_padded != N:
if fp32_weight.dim() == 1:
padded = torch.zeros(N_padded, dtype=fp32_weight.dtype, device=fp32_weight.device)
padded[:N] = fp32_weight
fp32_weight = padded
else:
padded = torch.zeros(M, N_padded, dtype=fp32_weight.dtype, device=fp32_weight.device)
padded[:, :N] = fp32_weight
fp32_weight = padded
# Reshape to blocks
if fp32_weight.dim() == 1:
weight_blocks = fp32_weight.view(-1, block_size)
else:
weight_blocks = fp32_weight.view(M, -1, block_size)
# Compute per-block amax
per_block_amax = weight_blocks.abs().amax(dim=-1)
per_block_scale = per_block_amax / (FP4_MAX * weight_scale_2)
per_block_scale = per_block_scale.clamp(min=1e-8)
# Convert to FP8 E4M3
try:
weight_scale = per_block_scale.to(torch.float8_e4m3fn)
except (RuntimeError, TypeError):
weight_scale = per_block_scale.to(torch.float32)
return weight_scale, weight_scale_2
def quantize_to_nvfp4_packed(fp32_weight, weight_scale, weight_scale_2, block_size=16):
"""
Quantize FP32 weight to NVFP4 packed uint8 format.
Simplified version for testing.
"""
device = fp32_weight.device
M = fp32_weight.shape[0] if fp32_weight.dim() > 1 else 1
N = fp32_weight.shape[-1]
# Pad if needed
N_padded = ((N + block_size - 1) // block_size) * block_size
if N_padded != N:
if fp32_weight.dim() == 1:
padded = torch.zeros(N_padded, dtype=fp32_weight.dtype, device=device)
padded[:N] = fp32_weight
fp32_weight = padded
else:
padded = torch.zeros(M, N_padded, dtype=fp32_weight.dtype, device=device)
padded[:, :N] = fp32_weight
fp32_weight = padded
# Reshape to blocks
if fp32_weight.dim() == 1:
weight_blocks = fp32_weight.view(-1, block_size)
else:
weight_blocks = fp32_weight.view(M, -1, block_size)
# Apply scaling
combined_scale = weight_scale.to(torch.float32) * weight_scale_2
scaled_weight = weight_blocks / combined_scale.unsqueeze(-1)
# Flatten
if fp32_weight.dim() == 1:
scaled_weight = scaled_weight.view(-1)
else:
scaled_weight = scaled_weight.view(M, -1)
# Get E2M1 bounds
e2m1_bounds = E2M1_BOUNDS.to(device)
# Extract sign and absolute values
sign_bit = (scaled_weight < 0).to(torch.uint8)
weight_abs = scaled_weight.abs()
# Quantize to E2M1 magnitude codes [0-7]
magnitude_code = torch.searchsorted(e2m1_bounds, weight_abs)
# Combine sign bit and magnitude
code = (sign_bit << 3) | magnitude_code.to(torch.uint8)
# Pack two 4-bit values per byte
N_current = code.shape[-1]
if N_current % 2 != 0:
# Pad to even
if code.dim() == 1:
padded = torch.zeros(N_current + 1, dtype=torch.uint8, device=device)
padded[:N_current] = code
code = padded
else:
padded = torch.zeros(M, N_current + 1, dtype=torch.uint8, device=device)
padded[:, :N_current] = code
code = padded
# Pack
if code.dim() == 1:
packed = (code[1::2] << 4) | code[0::2]
else:
packed = (code[:, 1::2] << 4) | code[:, 0::2]
return packed
def test_dequant_lookup_table():
"""Test 1: Verify NVFP4 lookup table values are correct."""
print("\n" + "=" * 70)
print("Test 1: NVFP4 Lookup Table")
print("=" * 70)
expected = [0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0,
-0.0, -0.5, -1.0, -1.5, -2.0, -3.0, -4.0, -6.0]
assert len(NVFP4_LUT) == 16, f"LUT should have 16 entries, got {len(NVFP4_LUT)}"
for i, (actual, expected_val) in enumerate(zip(NVFP4_LUT, expected)):
assert abs(actual - expected_val) < 1e-6, f"LUT[{i}] = {actual}, expected {expected_val}"
print(f" PASS: Lookup table correct: {NVFP4_LUT.tolist()[:8]}")
print(f" {NVFP4_LUT.tolist()[8:]}")
print(" PASS: Test 1 PASSED\n")
def test_dequant_simple():
"""Test 2: Simple dequantization with known values."""
print("=" * 70)
print("Test 2: Simple Dequantization")
print("=" * 70)
# Create simple test case: packed values representing [0, 1.0, 2.0, 3.0, ...]
# Codes: 0=0.0, 2=1.0, 4=2.0, 5=3.0, 6=4.0, 7=6.0
# Pack: (high << 4) | low
packed = torch.tensor([
[0x20, 0x54, 0x76, 0x00, 0x00, 0x00, 0x00, 0x00], # [0,2,4,5,6,7,0,0] -> [0,1,2,3,4,6,0,0]
], dtype=torch.uint8)
# Uniform scales for simplicity
scale = torch.ones(1, 1, dtype=torch.float8_e4m3fn)
scale_2 = torch.tensor([1.0], dtype=torch.float32)
result = dequantize_nvfp4(packed, scale, scale_2, dtype=torch.float32)
print(f" Packed: {packed[0].tolist()}")
print(f" Scales: scale={scale.shape}, scale_2={scale_2.item()}")
print(f" Result shape: {result.shape}")
print(f" Result values: {result[0].tolist()}")
# Expected: [0, 1, 2, 3, 4, 6, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
expected_values = [0, 1, 2, 3, 4, 6, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
for i, (val, expected) in enumerate(zip(result[0].tolist(), expected_values)):
assert abs(val - expected) < 0.01, f"Position {i}: got {val}, expected {expected}"
print(" PASS: Dequantization correct")
print(" PASS: Test 2 PASSED\n")
def test_quantize_dequantize_roundtrip():
"""Test 3: Quantize then dequantize, check error is acceptable."""
print("=" * 70)
print("Test 3: Quantization-Dequantization Roundtrip")
print("=" * 70)
# Create test tensor with values in representable range
M, N = 64, 256
torch.manual_seed(42)
fp32_weight = torch.randn(M, N, dtype=torch.float32) * 2.0 # Scale to ~[-6, 6]
print(f" Input shape: {fp32_weight.shape}")
print(f" Input range: [{fp32_weight.min():.3f}, {fp32_weight.max():.3f}]")
# Compute scales
scale, scale_2 = compute_nvfp4_scales(fp32_weight, block_size=16)
print(f" Scale shape: {scale.shape}, scale_2: {scale_2.item():.6e}")
# Quantize
packed = quantize_to_nvfp4_packed(fp32_weight, scale, scale_2, block_size=16)
print(f" Packed shape: {packed.shape} (expected [{M}, {N//2}])")
assert packed.shape == (M, N // 2), f"Packed shape mismatch"
# Dequantize
dequantized = dequantize_nvfp4(packed, scale, scale_2, dtype=torch.float32)
print(f" Dequantized shape: {dequantized.shape}")
assert dequantized.shape == (M, N), f"Dequantized shape mismatch"
# Compute error
error = (fp32_weight - dequantized).abs()
mean_error = error.mean().item()
max_error = error.max().item()
relative_error = (error / (fp32_weight.abs() + 1e-8)).mean().item()
print(f" Mean absolute error: {mean_error:.6f}")
print(f" Max absolute error: {max_error:.6f}")
print(f" Mean relative error: {relative_error:.6f}")
# For 4-bit quantization, we expect some error but should be reasonable
assert mean_error < 1.0, f"Mean error too high: {mean_error}"
assert relative_error < 0.5, f"Relative error too high: {relative_error}"
print(" PASS: Roundtrip error acceptable for 4-bit quantization")
print(" PASS: Test 3 PASSED\n")
def test_gemm_shapes():
"""Test 4: NVFP4 GEMM with various shapes."""
print("=" * 70)
print("Test 4: NVFP4 GEMM Shape Tests")
print("=" * 70)
test_cases = [
(32, 64, 128), # Small
(128, 256, 512), # Medium
(64, 512, 256), # Asymmetric
]
for M, N, K in test_cases:
print(f"\n Testing GEMM: [{M}, {K}] @ [{N}, {K}].T = [{M}, {N}]")
# Create input activation
x = torch.randn(M, K, dtype=torch.bfloat16)
# Create quantized weight
weight_fp32 = torch.randn(N, K, dtype=torch.float32) * 2.0
scale, scale_2 = compute_nvfp4_scales(weight_fp32, block_size=16)
packed_weight = quantize_to_nvfp4_packed(weight_fp32, scale, scale_2, block_size=16)
print(f" Input: {x.shape}, Weight: {packed_weight.shape}")
print(f" Scales: {scale.shape}, {scale_2.shape}")
# Run NVFP4 GEMM
result = nvfp4_gemm_dequant(x, packed_weight, scale, scale_2)
print(f" Output: {result.shape}")
assert result.shape == (M, N), f"Output shape mismatch: {result.shape} != ({M}, {N})"
# Verify no NaN/Inf
assert not torch.isnan(result).any(), "Output contains NaN"
assert not torch.isinf(result).any(), "Output contains Inf"
print(f" PASS: Shape correct, no NaN/Inf")
print("\n PASS: All GEMM shape tests passed")
print(" PASS: Test 4 PASSED\n")
def test_gemm_correctness():
"""Test 5: Verify NVFP4 GEMM output is close to reference."""
print("=" * 70)
print("Test 5: NVFP4 GEMM Correctness")
print("=" * 70)
M, N, K = 64, 128, 256
# Create test tensors
x = torch.randn(M, K, dtype=torch.bfloat16)
weight_fp32 = torch.randn(N, K, dtype=torch.float32) * 1.5
# Quantize weight
scale, scale_2 = compute_nvfp4_scales(weight_fp32, block_size=16)
packed_weight = quantize_to_nvfp4_packed(weight_fp32, scale, scale_2, block_size=16)
# Run NVFP4 GEMM
result_nvfp4 = nvfp4_gemm_dequant(x, packed_weight, scale, scale_2)
# Run reference GEMM with FP32
result_reference = F.linear(x, weight_fp32.to(torch.bfloat16))
print(f" NVFP4 GEMM output: {result_nvfp4.shape}, dtype={result_nvfp4.dtype}")
print(f" Reference output: {result_reference.shape}, dtype={result_reference.dtype}")
# Compute error
error = (result_nvfp4.float() - result_reference.float()).abs()
mean_error = error.mean().item()
max_error = error.max().item()
relative_error = (error / (result_reference.float().abs() + 1e-8)).mean().item()
print(f" Mean absolute error: {mean_error:.6f}")
print(f" Max absolute error: {max_error:.6f}")
print(f" Mean relative error: {relative_error:.6f}")
# Due to 4-bit quantization, expect significant error but not catastrophic
assert mean_error < 5.0, f"Mean error too high: {mean_error}"
assert relative_error < 1.0, f"Relative error too high: {relative_error}"
print(" PASS: NVFP4 GEMM output reasonably close to reference")
print(" PASS: Test 5 PASSED\n")
def main():
"""Run all NVFP4 kernel unit tests."""
print("\n" + "=" * 70)
print("NVFP4 Kernel Unit Tests")
print("=" * 70)
print("Testing NVFP4 quantization/dequantization and GEMM operations")
print("Expected runtime: < 30 seconds")
print("=" * 70)
try:
# Run all tests
test_dequant_lookup_table()
test_dequant_simple()
test_quantize_dequantize_roundtrip()
test_gemm_shapes()
test_gemm_correctness()
# Summary
print("=" * 70)
print("PASS: ALL TESTS PASSED")
print("=" * 70)
print("NVFP4 kernel functions are working correctly!")
print("Ready to proceed with full model testing.")
print("=" * 70)
return 0
except AssertionError as e:
print(f"\nFAIL: TEST FAILED: {e}")
import traceback
traceback.print_exc()
return 1
except Exception as e:
print(f"\nFAIL: UNEXPECTED ERROR: {e}")
import traceback
traceback.print_exc()
return 1
if __name__ == "__main__":
sys.exit(main())
|