File size: 9,997 Bytes
1717684 | 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 | #!/usr/bin/env python3
"""
Test suite for GauS: Differentiable Scheduling via Gaussian Reparameterization.
Integrated with Twill's kernel descriptions for direct comparison.
Tests:
1. Basic Gaussian reparameterization (P_i^d computation)
2. Regular scheduling on simple DAGs
3. Modulo scheduling on Twill's FMHA kernels (compare to Twill ILP+SMT)
4. Scalability: 100-node, 1000-node random DAGs
"""
import sys
import time
import numpy as np
sys.path.insert(0, '/app')
from twill.gaus_solver import (
GauSSolver, GausGraph, GausResult,
compute_asap, compute_alap, gaussian_cdf,
twill_graph_to_gaus, gaus_solve_twill_graph,
generate_random_dag,
)
from twill.kernels import (
flash_attention_forward_simplified,
flash_attention_forward_hopper,
flash_attention_forward_blackwell,
simple_gemm_pipeline,
)
from twill.twill_solver import twill_solve
def test_gaussian_cdf():
"""Test basic Gaussian CDF computation."""
print("\n" + "=" * 70)
print("TEST: Gaussian CDF + P_i^d Computation")
print("=" * 70)
import torch
# CDF at 0 should be 0.5
assert abs(gaussian_cdf(torch.tensor(0.0)).item() - 0.5) < 1e-6
# CDF at large positive should be ~1
assert abs(gaussian_cdf(torch.tensor(5.0)).item() - 1.0) < 1e-4
# CDF at large negative should be ~0
assert abs(gaussian_cdf(torch.tensor(-5.0)).item()) < 1e-4
# Test P_i^d: single node at ΞΌ=2, Ο=0.5
graph = GausGraph(num_nodes=1, edges=[])
solver = GauSSolver(graph, D=6)
mu = torch.tensor([2.0])
sigma = torch.tensor([0.5])
P = solver._compute_P(mu, sigma)
print(f" ΞΌ=2.0, Ο=0.5, D=6")
print(f" P = {P[0].detach().numpy().round(4)}")
print(f" Sum P = {P[0].sum().item():.6f} (should be ~1.0)")
print(f" Argmax P = {P[0].argmax().item()} (should be 2)")
assert P[0].argmax().item() == 2, "Peak should be at ΞΌ=2"
assert abs(P[0].sum().item() - 1.0) < 0.01, "Probabilities should sum to ~1"
# As Ο β 0, P should be a delta at round(ΞΌ)
sigma_small = torch.tensor([0.01])
P_sharp = solver._compute_P(mu, sigma_small)
print(f" Ο=0.01: P[2]={P_sharp[0, 2].item():.6f} (should be ~1.0)")
assert P_sharp[0, 2].item() > 0.99
print("β Gaussian CDF test passed")
return True
def test_asap_alap():
"""Test ASAP/ALAP computation."""
print("\n" + "=" * 70)
print("TEST: ASAP / ALAP Computation")
print("=" * 70)
# Chain: 0 -> 1 -> 2
graph = GausGraph(
num_nodes=3,
edges=[(0, 1), (1, 2)],
node_names=["A", "B", "C"],
)
asap = compute_asap(graph)
alap = compute_alap(graph, D=5)
print(f" Chain A->B->C, D=5")
print(f" ASAP: {asap}") # Expected: [0, 1, 2]
print(f" ALAP: {alap}") # Expected: [2, 3, 4]
assert list(asap) == [0, 1, 2], f"ASAP wrong: {asap}"
assert list(alap) == [2, 3, 4], f"ALAP wrong: {alap}"
print("β ASAP/ALAP test passed")
return True
def test_regular_scheduling():
"""Test regular (non-modulo) scheduling."""
print("\n" + "=" * 70)
print("TEST: Regular Scheduling (Formulation A)")
print("=" * 70)
# Diamond: 0 -> 1, 0 -> 2, 1 -> 3, 2 -> 3
graph = GausGraph(
num_nodes=4,
edges=[(0, 1), (0, 2), (1, 3), (2, 3)],
resource_weights=np.array([1, 1, 1, 1], dtype=np.float64),
node_names=["A", "B", "C", "D"],
)
solver = GauSSolver(graph, D=6, lr=0.05)
result = solver.solve_regular(max_iters=500, legalize_every=100, verbose=True)
print(f"\n Result: {result}")
# Verify dependencies
s = result.schedule
assert s[1] > s[0], f"B must be after A: {s[1]} > {s[0]}"
assert s[2] > s[0], f"C must be after A: {s[2]} > {s[0]}"
assert s[3] > s[1], f"D must be after B: {s[3]} > {s[1]}"
assert s[3] > s[2], f"D must be after C: {s[3]} > {s[2]}"
assert result.is_feasible, "Schedule should be feasible"
print("β Regular scheduling test passed")
return True
def test_modulo_scheduling_simple():
"""Test modulo scheduling on a simple loop body."""
print("\n" + "=" * 70)
print("TEST: Modulo Scheduling (Formulation C) β Simple")
print("=" * 70)
# Simple loop: A -> B -> C, with C -> C loop-carried
graph = GausGraph(
num_nodes=3,
edges=[(0, 1), (1, 2)],
back_edges=[(2, 2, 1)], # C -> C with Ξ΄=1
resource_weights=np.array([1, 1, 1], dtype=np.float64),
node_names=["S", "P", "O"],
)
D = 8
II = 2
solver = GauSSolver(graph, D=D, lr=0.02)
result = solver.solve_modulo(II=II, R_cap=1.0, max_iters=1000, verbose=True)
print(f"\n Result: {result}")
# Verify dependencies
s = result.schedule
assert s[1] > s[0], f"P must be after S"
assert s[2] > s[1], f"O must be after P"
print("β Modulo scheduling (simple) test passed")
return True
def test_twill_comparison_simplified_fa():
"""Compare GauS vs Twill on simplified Flash Attention."""
print("\n" + "=" * 70)
print("TEST: GauS vs Twill β Simplified Flash Attention")
print("=" * 70)
graph = flash_attention_forward_simplified()
# Twill solution
print("--- Twill (ILP + SMT) ---")
t0 = time.time()
twill_result = twill_solve(
graph, max_I=5, enable_cost_normalization=False,
enable_memory_constraints=False, enable_warp_constraints=False,
verbose=False,
)
twill_time = time.time() - t0
if twill_result:
print(f" Twill: I={twill_result.I}, schedule={twill_result.schedule}, time={twill_time:.3f}s")
# GauS solution
print("\n--- GauS (Differentiable) ---")
gaus_graph, name_to_idx = twill_graph_to_gaus(graph, D=10)
solver = GauSSolver(gaus_graph, D=10, lr=0.02)
gaus_result = solver.solve_modulo(
II=2, R_cap=1.0, max_iters=1500, legalize_every=200, verbose=True,
)
print(f"\n GauS: {gaus_result}")
# Compare
print(f"\n--- Comparison ---")
if twill_result:
print(f" Twill: I={twill_result.I}, time={twill_time:.3f}s")
print(f" GauS: II={gaus_result.initiation_interval}, "
f"feasible={gaus_result.is_feasible}, time={gaus_result.solve_time_seconds:.3f}s")
return True
def test_twill_comparison_hopper():
"""Compare GauS vs Twill on Hopper FMHA forward."""
print("\n" + "=" * 70)
print("TEST: GauS vs Twill β Hopper FMHA Forward")
print("=" * 70)
graph = flash_attention_forward_hopper()
# Twill
print("--- Twill ---")
t0 = time.time()
twill_result = twill_solve(
graph, max_I=10, enable_cost_normalization=False,
enable_memory_constraints=False, enable_warp_constraints=False,
verbose=False,
)
twill_time = time.time() - t0
if twill_result:
print(f" Twill: I={twill_result.I}, schedule={twill_result.schedule}, time={twill_time:.3f}s")
# GauS
print("\n--- GauS ---")
gaus_result = gaus_solve_twill_graph(
graph, target_II=4, D=20, max_iters=2000, verbose=True,
)
print(f"\n--- Comparison ---")
if twill_result:
print(f" Twill: I={twill_result.I}, time={twill_time:.3f}s")
print(f" GauS: II={gaus_result.initiation_interval}, "
f"feasible={gaus_result.is_feasible}, time={gaus_result.solve_time_seconds:.3f}s")
return True
def test_scalability():
"""Test GauS scalability on larger graphs."""
print("\n" + "=" * 70)
print("TEST: Scalability β Random DAGs")
print("=" * 70)
for n_nodes in [50, 200, 1000]:
print(f"\n--- {n_nodes} nodes ---")
graph = generate_random_dag(
num_nodes=n_nodes,
edge_density=min(0.3, 10.0 / n_nodes), # Keep sparse for large graphs
max_weight=2,
num_back_edges=max(1, n_nodes // 20),
seed=42,
)
D = n_nodes + 10
II = max(2, n_nodes // 10)
solver = GauSSolver(graph, D=D, lr=0.01)
t0 = time.time()
result = solver.solve_modulo(
II=II, R_cap=float(n_nodes // 5),
max_iters=min(1000, n_nodes * 5),
legalize_every=200,
verbose=False,
)
elapsed = time.time() - t0
print(f" |V|={n_nodes}, |E|={len(graph.edges)}, D={D}, II={II}")
print(f" Time: {elapsed:.2f}s")
print(f" Feasible: {result.is_feasible}")
print(f" Violations: {result.num_violations}")
print(f" Schedule range: [{min(result.schedule.values())}, {max(result.schedule.values())}]")
print("\nβ Scalability test passed")
return True
if __name__ == "__main__":
print("β" + "β" * 68 + "β")
print("β" + " GauS Test Suite ".center(68) + "β")
print("β" + " Differentiable Scheduling via Gaussian Reparameterization ".center(68) + "β")
print("β" + " (arXiv:2602.20427) ".center(68) + "β")
print("β" + "β" * 68 + "β")
results = {}
start = time.time()
results["Gaussian CDF"] = test_gaussian_cdf()
results["ASAP/ALAP"] = test_asap_alap()
results["Regular Scheduling"] = test_regular_scheduling()
results["Modulo Scheduling (Simple)"] = test_modulo_scheduling_simple()
results["GauS vs Twill: Simplified FA"] = test_twill_comparison_simplified_fa()
results["GauS vs Twill: Hopper FMHA"] = test_twill_comparison_hopper()
results["Scalability"] = test_scalability()
elapsed = time.time() - start
print("\n" + "=" * 70)
print("TEST SUMMARY")
print("=" * 70)
for name, passed in results.items():
status = "β PASS" if passed else "β FAIL"
print(f" {status} {name}")
print(f"\nTotal time: {elapsed:.2f}s")
print(f"Passed: {sum(results.values())}/{len(results)}")
sys.exit(0 if all(results.values()) else 1)
|