File size: 10,434 Bytes
ecadbd9 | 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 | import torch
import torch.nn as nn
from typing import Optional, Set
from peft.tuners.tuners_utils import BaseTunerLayer, check_adapters_to_merge
import einops
from layer import SharedMonarch, SingleSharedBlockDiag
def test_single():
# input B * c, output B * r
B = 20
block_size_r = 5
block_size_c = 4
r = 2
uni_w = B // r
model = SingleSharedBlockDiag(num_unique_weights=uni_w, share_factor=r, block_size_c=block_size_c, block_size_r=block_size_r)
x = torch.rand(2, 3, B*block_size_c)
y = model(x)
print(f"Weights Shape: {model.weights.shape}, input {x.shape}")
print(f"Output Shape: {y.shape}")
W_group_0 = model.weights[0]
x_blk_0 = x[0, 0, 0:block_size_c]
y_manual = x_blk_0 @ W_group_0.T
y_model = y[0, 0, 0:block_size_r]
print("\n--- Correctness Check ---")
print(f"Manual: {y_manual.detach().numpy()}")
print(f"Model : {y_model.detach().numpy()}")
# Assert close
assert torch.allclose(y_manual, y_model, atol=1e-6)
print(">> Logic Verified: Dense Matrix Multiplication per block is correct.")
def test_monarch():
rL, cL = 4, 6
rR, cR = 4, 2
share_factor = 2
model = SharedMonarch(share_factor, rR, cR, rL, cL)
x = torch.arange(cL * cR, dtype=torch.float32).unsqueeze(0)
y = model(x)
print(f"Input shape: {x.shape}")
print(f"Output shape: {y.shape}")
def manual_override_weights(model, val_r=1.0, val_l=1.0, method='identity'):
"""
Helper function to force weights to known values for testing.
"""
with torch.no_grad():
# --- Override Layer R ---
if method == 'identity':
# Create identity matrices for each block
# Weight shape: (num_unique, r, c). We want weights s.t. x @ W.T = x
# So W must be Identity.
for i in range(model.sama_R.num_unique_weights):
nn.init.eye_(model.sama_R.weights[i])
model.sama_R.weights[i] *= val_r # Optional scaling
elif method == 'seq':
# Fill with sequential integers to track position
nn.init.constant_(model.sama_R.weights, 0)
count = 1
for i in range(model.sama_R.num_unique_weights):
for r in range(model.sama_R.block_size_r):
for c in range(model.sama_R.block_size_c):
model.sama_R.weights[i, r, c] = count
count += 1
# --- Override Layer L ---
if method == 'identity':
for i in range(model.sama_L.num_unique_weights):
nn.init.eye_(model.sama_L.weights[i])
model.sama_L.weights[i] *= val_l
def test_monarch_correctness():
print("\n=== STARTING MONARCH LAYER DEBUG/TEST ===")
# 1. Setup minimal model to easy calculation
# N = 4 (2x2 grid).
# This is small enough to print everything.
rL, cL = 2, 2
rR, cR = 2, 2
share_factor = 1
model = SharedMonarch(share_factor, rR, cR, rL, cL)
# ==========================================
# TEST CASE 1: The "Passthrough" (Identity) Test
# ==========================================
print("\n--- Test 1: Identity/Reconstruction ---")
print("Goal: If R=Identity and L=Identity, output must equal input.")
print("Logic: x -> I -> P -> I -> P_T -> x")
# Force weights to Identity
manual_override_weights(model, method='identity')
# Create random input
x = torch.randn(1, cL * cR)
y = model(x)
print(f"Input: {x.detach().numpy().round(2)}")
print(f"Output: {y.detach().numpy().round(2)}")
if torch.allclose(x, y, atol=1e-6):
print(">> [PASS] Identity Test: P and P_T are correctly inverse.")
else:
print(">> [FAIL] Identity Test: Output does not match Input!")
return # Stop if basic logic fails
# ==========================================
# TEST CASE 2: The "Specific Path" (One-Hot) Test
# ==========================================
print("\n--- Test 2: Manual Calculation Trace ---")
print("Goal: Trace a specific value through the layers.")
# Setup:
# R will double the values (Scale = 2 * Identity)
# L will triple the values (Scale = 3 * Identity)
# Expected: Output = Input * 2 * 3 = Input * 6
manual_override_weights(model, val_r=2.0, val_l=3.0, method='identity')
# Input: One-hot vector at index 1 (value 1.0)
# [0, 1, 0, 0]
x_one_hot = torch.tensor([[0.0, 1.0, 0.0, 0.0]])
y_test = model(x_one_hot)
print(f"Input One-Hot: {x_one_hot.numpy()}")
print(f"Output: {y_test.detach().numpy()}")
expected_value = 1.0 * 2.0 * 3.0 # = 6.0
if torch.allclose(y_test[0, 1], torch.tensor(expected_value)):
print(f">> [PASS] Value Scaling: Input 1.0 became {y_test[0,1].item()} (Expected 6.0)")
else:
print(f">> [FAIL] Value Scaling: Expected 6.0 but got {y_test[0,1].item()}")
# ==========================================
# TEST CASE 3: The "Permutation Logic" Check
# ==========================================
print("\n--- Test 3: Permutation Logic (Complex) ---")
print("Goal: Ensure R acts on local blocks (neighbors) and L acts on distant blocks (strided).")
# Reset model
model = SharedMonarch(share_factor, rR, cR, rL, cL)
# Configuration:
# We want R to mix [0,1] and [2,3].
# We want L to mix [0,2] and [1,3] (after permute).
# Let's set R to be Identity (Pass through).
manual_override_weights(model, val_r=1.0, val_l=1.0, method='identity')
# Now CHANGE Layer L (The second one) manually.
# L has 2 blocks. Block size is 2.
# Block 0 of L acts on indices that WERE [0, 2] in original space (due to transpose).
# Let's make Block 0 of L sum its inputs: Matrix [[1, 1], [1, 1]]
with torch.no_grad():
# L weights shape: (num_unique, block_r, block_c) -> (2, 2, 2)
# Set Block 0 to all ones
model.sama_L.weights[0] = torch.ones(2, 2)
# Set Block 1 to Identity (keep simple)
nn.init.eye_(model.sama_L.weights[1])
# Input: Activate index 0 and 2.
# x = [1, 0, 1, 0]
x_input = torch.tensor([[1.0, 0.0, 1.0, 0.0]])
# Manual Trace Logic:
# 1. x -> R (Identity) -> [1, 0, 1, 0]
# 2. P (Permute):
# Original indices: 0, 1, 2, 3
# View (2,2): [[0,1], [2,3]]
# Transpose: [[0,2], [1,3]]
# New Vector sequence: [val_at_0, val_at_2, val_at_1, val_at_3]
# Values: [1, 1, 0, 0]
# 3. L (Block Diag):
# Vector is [1, 1] (Block 0 input) and [0, 0] (Block 1 input).
# Block 0 of L is "All Ones" matrix.
# Calculation: [1, 1] @ [[1,1],[1,1]].T = [2, 2]
# Block 1 of L is Identity.
# Calculation: [0, 0] @ I = [0, 0]
# Result after L: [2, 2, 0, 0] (This is in Permuted space!)
# 4. P_T (Un-permute):
# Current View (2,2): [[2, 2], [0, 0]]
# Transpose back: [[2, 0], [2, 0]]
# Flatten: [2, 0, 2, 0]
y_complex = model(x_input)
print(f"Input: {x_input.numpy()}")
print(f"Expected: [2., 0., 2., 0.]")
print(f"Actual: {y_complex.detach().numpy()}")
if torch.allclose(y_complex, torch.tensor([[2., 0., 2., 0.]])):
print(">> [PASS] Permutation Mixing Logic is correct!")
else:
print(">> [FAIL] Permutation Mixing Logic is flawed.")
import matplotlib.pyplot as plt
import torch
import numpy as np
def visualize_monarch_structure():
"""
Visualizes the effective weight matrices of the Monarch layers to prove
the structural properties (Sparsity & Butterfly pattern).
"""
# 1. Initialize Model
rL, cL = 4, 4
rR, cR = 4, 4
share_factor = 1
print(f"Visualizing Monarch Structure for N={cL*cR} (Grid {cL}x{cR})...")
model = SharedMonarch(share_factor, rR, cR, rL, cL)
# 2. Force weights to be random but significant (away from 0) for clear visualization
with torch.no_grad():
nn.init.uniform_(model.sama_R.weights, 0.5, 1.0)
nn.init.uniform_(model.sama_L.weights, 0.5, 1.0)
# 3. Create Identity Matrix to extract Effective Weights
input_dim = cL * cR
identity_matrix = torch.eye(input_dim)
# 4. Extract Matrices
with torch.no_grad():
# A. Extract Matrix R (Right Layer Only)
# We only run the first layer: x -> R
w_r_transposed = model.sama_R(identity_matrix)
w_r = w_r_transposed.T # Transpose back to standard (Out, In) format
# B. Extract Total Effective Matrix (R -> P -> L -> P_T)
w_total_transposed = model(identity_matrix)
w_total = w_total_transposed.T
# 5. Plotting
fig, axes = plt.subplots(1, 2, figsize=(12, 5))
# Plot A: The Local Block Diagonal Matrix (Layer R)
im1 = axes[0].imshow(w_r.abs().numpy(), cmap='Blues', interpolation='nearest')
axes[0].set_title(f"Layer R (Block Diagonal)\nBlocks of size {cR}x{cR}")
axes[0].set_xlabel("Input Index")
axes[0].set_ylabel("Intermediate Index")
# Plot B: The Full Monarch Matrix (Butterfly Structure)
im2 = axes[1].imshow(w_total.abs().numpy(), cmap='Reds', interpolation='nearest')
axes[1].set_title(f"Full Monarch (Butterfly)\nGlobal Mixing")
axes[1].set_xlabel("Input Index")
axes[1].set_ylabel("Output Index")
plt.tight_layout()
plt.savefig("monarch_viz1.png")
print("Graph saved to 'monarch_viz.png'. Please check your file explorer.")
plt.close()
def test_merge():
model = SharedMonarch(share_factor=1,
block_size_cR=3, block_size_rR=4,
block_size_cL=4, block_size_rL=5)
# Method 1 (Manual)
with torch.no_grad():
W_manual = model.get_delta_weight()
# Method 2: Identity Trick (Forward)
# Note: forward(I) returns là W^T.
I = torch.eye(12)
with torch.no_grad():
W_forward = model(I).T
print(f"Manual Shape: {W_manual.shape}") # (20, 12)
print(f"Forward Shape: {W_forward.shape}") # (20, 12)
diff = (W_manual - W_forward).abs().max()
print(f"Maximum Difference: {diff.item()}")
if diff < 1e-6:
print(">> Accurate!")
if __name__ == "__main__":
# test_monarch_correctness()
# visualize_monarch_structure()
test_merge() |