File size: 11,286 Bytes
038acee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Literature-anchored specification tests.

Every expected value here is derived from a primary source (cited
per test; full references below) or from a derivation written out in
the docstring - never from the current output of this or the sibling repo. The minihack repo
carries the same assertions with the same inputs and tolerances wherever
the mathematics is parameter-free and shared.

Tolerances: closed-form checks use atol=1e-6 (an order of magnitude above
float32 round-off, which the parity probes measured at <4e-8 on these
functions). Statistical checks state their sampling distribution and use a
4-sigma bound (or the stated multiple) with the derivation in the docstring.

References:
- MDLM: Sahoo et al., "Simple and Effective Masked Diffusion Language
  Models", NeurIPS 2024. arXiv:2406.07524.
- Shi: Shi et al., "Simplified and Generalized Masked Diffusion for
  Discrete Data", NeurIPS 2024. arXiv:2406.04329.
- ReMDM: Wang et al., "Remasking Discrete Diffusion Models with
  Inference-Time Scaling", NeurIPS 2025. arXiv:2503.00307.
- Holtzman et al., "The Curious Case of Neural Text Degeneration",
  ICLR 2020. arXiv:1904.09751.
"""

from __future__ import annotations

import math

import jax
import jax.numpy as jnp
import numpy as np

from src.diffusion.forward import forward_process
from src.diffusion.loss import compute_loss
from src.diffusion.sampling import (
    _nucleus_sample,
    _sigma_max,
    sample_plan_inpainting,
    sigma_cap,
    sigma_conf,
    sigma_rescale,
)
from src.diffusion.schedules import (
    SCHEDULE_MAP,
    cosine_schedule,
    cosine_schedule_deriv,
    linear_schedule,
    linear_schedule_deriv,
)

ATOL = 1e-6
T_GRID = jnp.array([0.0, 1.0 / 3.0, 0.5, 2.0 / 3.0, 1.0])


# ---------------------------------------------------------------------------
# Noise schedules
# ---------------------------------------------------------------------------


def test_linear_schedule_closed_form():
    """alpha(t) = 1 - t; alpha'(t) = -1.

    Source: MDLM (Sahoo et al.) App E.1 eq (90) family; ReMDM Sec 3
    convention alpha(0)=1, alpha(1)=0. Same grid and tolerance as the
    minihack twin test.
    """
    expected = np.array([1.0, 2 / 3, 0.5, 1 / 3, 0.0])
    assert np.allclose(np.asarray(linear_schedule(T_GRID)), expected, atol=ATOL)
    assert np.allclose(np.asarray(linear_schedule_deriv(T_GRID)), -1.0, atol=ATOL)


def test_cosine_schedule_closed_form():
    """alpha(t) = cos(pi t / 2); alpha'(t) = -(pi/2) sin(pi t / 2).

    Source: MDLM App E.1 eq (92) ("Cosine"). Values: cos(pi/6)=sqrt(3)/2,
    cos(pi/4)=sqrt(2)/2, cos(pi/3)=1/2.
    """
    expected = np.array([1.0, math.sqrt(3) / 2, math.sqrt(2) / 2, 0.5, 0.0])
    assert np.allclose(np.asarray(cosine_schedule(T_GRID)), expected, atol=ATOL)
    expected_d = np.array(
        [0.0, -math.pi / 4, -(math.pi / 2) * math.sqrt(2) / 2,
         -(math.pi / 2) * math.sqrt(3) / 2, -math.pi / 2]
    )
    assert np.allclose(np.asarray(cosine_schedule_deriv(T_GRID)), expected_d, atol=ATOL)


def test_schedule_registry_names_follow_mdlm_e1():
    """The label "cosine" must denote MDLM eq (92), not eq (91)."""
    fn, deriv = SCHEDULE_MAP["cosine"]
    assert abs(float(fn(jnp.array(0.5))) - math.sqrt(2) / 2) < ATOL
    assert abs(float(deriv(jnp.array(0.5))) + (math.pi / 2) * math.sqrt(2) / 2) < ATOL


# ---------------------------------------------------------------------------
# Forward corruption
# ---------------------------------------------------------------------------


def test_forward_marginal_endpoints():
    """q(z_t|x) = Cat(alpha_t x + (1-alpha_t) m): alpha=1 identity,
    alpha=0 all-MASK, deterministically (uniform draws lie in [0,1)).

    Source: MDLM Sec 3.2.1.
    """
    rng = jax.random.PRNGKey(0)
    x0 = jax.random.randint(rng, (4, 32), 0, 17)
    z_keep = forward_process(rng, x0, jnp.ones(4), mask_id=17)
    assert bool(jnp.all(z_keep == x0))
    z_mask = forward_process(rng, x0, jnp.zeros(4), mask_id=17)
    assert bool(jnp.all(z_mask == 17))


def test_forward_marginal_rate():
    """Empirical mask rate at alpha=0.7 is 0.3 within 4 sigma.

    Source: MDLM Sec 3.2.1. N = 200*64 = 12800 Bernoulli(0.3) draws;
    sigma = sqrt(0.3*0.7/12800) = 0.00405; bound = 4 sigma = 0.0162.
    """
    rng = jax.random.PRNGKey(1)
    x0 = jnp.zeros((200, 64), dtype=jnp.int32)
    zt = forward_process(rng, x0, jnp.full(200, 0.7), mask_id=17)
    rate = float((zt == 17).mean())
    assert abs(rate - 0.3) < 0.0162


# ---------------------------------------------------------------------------
# Loss: NELBO estimator
# ---------------------------------------------------------------------------

_V, _H = 5, 8


def _uniform_apply(params, obs, z, t, rng):
    return jnp.zeros((obs.shape[0], _H, _V))


def _loss(t_pin, schedule_name, valid, B):
    fn, deriv = SCHEDULE_MAP[schedule_name]
    rng = jax.random.PRNGKey(0)
    x0 = jax.random.randint(rng, (B, _H), 0, _V)
    obs = jnp.zeros((B, 3))
    loss, info = compute_loss(
        _uniform_apply, None, rng, x0, obs, valid, _V, fn, deriv,
        t_min=t_pin, t_max=t_pin,
    )
    return float(loss)


def test_loss_all_masked_linear_t1():
    """Loss = w(1) * log V = log V with everything masked, uniform logits.

    Source: MDLM eq (10); Shi eq (4). Derivation: t pinned to 1 makes the
    forward step deterministic (alpha=0, all masked); uniform logits give
    CE = log V per position; sum/H = log V; linear w(1) = 1.
    """
    got = _loss(1.0, "linear", jnp.ones(4), B=4)
    assert abs(got - math.log(_V)) < 1e-5


def test_loss_weight_uses_analytic_derivative_cosine_t1():
    """Cosine at t=1: w(1) = -alpha'(1)/(1-alpha(1)) = pi/2, so loss =
    (pi/2) log V.

    Source: MDLM eq (10) with the eq (92) schedule: alpha'(1) =
    -(pi/2) sin(pi/2) = -pi/2 and alpha(1)=0. Pins the analytic-derivative
    form of the weight.
    """
    got = _loss(1.0, "cosine", jnp.ones(4), B=4)
    assert abs(got - (math.pi / 2) * math.log(_V)) < 1e-5


def test_loss_invalid_samples_contribute_zero():
    """A sample with valid=0 contributes exactly zero to the batch mean.

    Source: the validity mask is the benchmark-forced analogue of PAD
    exclusion (MDLM Sec 3.2.3: loss over masked positions of real data
    only). Derivation: B=2 at t=1 (all masked), valid=[1,0]:
    mean(log V, 0) = log V / 2.
    """
    got = _loss(1.0, "linear", jnp.array([1.0, 0.0]), B=2)
    assert abs(got - math.log(_V) / 2) < 1e-5


def test_loss_denominator_is_per_token_not_per_masked():
    """E[loss] at pinned t=0.5 (cosine) = w(0.5) * log V * (1 - alpha),
    which distinguishes the per-token denominator (1/H) from the
    per-masked-count denominator, whose expectation is
    w(0.5) * log V (3.41x larger here).

    Source: MDLM eq (8)/(10); Shi eq (4) (no division by the realised
    masked count). Derivation: per-sample loss = w * log V * n_hat/H with
    n_hat ~ Bin(H=8, 1-alpha=0.29289); mean over B=8192 samples has sigma
    = w*logV*std(n_hat)/H/sqrt(B) = w*logV*0.00178; the bound below is
    0.01*w*logV (~5.6 sigma), far smaller than the 2.41x separation
    between the two hypotheses.
    """
    alpha = math.sqrt(2) / 2
    w = (math.pi / 2) * math.sin(math.pi / 4) / (1 - alpha)
    expected = w * math.log(_V) * (1 - alpha)
    got = _loss(0.5, "cosine", jnp.ones(8192), B=8192)
    assert abs(got - expected) < 0.01 * w * math.log(_V)


# ---------------------------------------------------------------------------
# Reverse step: remasking schedules and the sigma bound
# ---------------------------------------------------------------------------


def test_sigma_strategies_closed_form_and_bound():
    """Same grid and assertions as the minihack twin (ReMDM eq (7), Sec 4.1)."""
    eta = 0.5
    for k in range(1, 10):
        a_t = jnp.array(1 - k / 10)
        a_s = jnp.array(1 - (k + 1) / 10)
        smax = min(1.0, (1 - float(a_s)) / float(a_t))
        assert abs(float(_sigma_max(a_t, a_s)) - smax) < ATOL
        assert abs(float(sigma_rescale(a_t, a_s, eta)) - eta * smax) < ATOL
        assert abs(float(sigma_cap(a_t, a_s, eta)) - min(eta, smax)) < ATOL


def test_conf_strategy_softmax_of_stored_psi():
    """sigma_conf = softmax(-psi) * eta * sigma_max over unmasked positions,
    zero at masked; lower psi => higher remask probability; sums to
    eta * sigma_max.

    Source: ReMDM Sec 4.1 (Confidence-Based Schedule).
    """
    eta = 0.5
    a_t, a_s = jnp.array(0.5), jnp.array(0.1)
    smax = float(_sigma_max(a_t, a_s))
    psi = jnp.array([[0.9, 0.2, jnp.inf, 0.5]])
    unmasked = jnp.array([[True, True, False, True]])
    sigma = np.asarray(sigma_conf(a_t, a_s, eta, psi, unmasked))
    assert sigma[0, 2] == 0.0
    assert sigma[0, 1] > sigma[0, 3] > sigma[0, 0]
    assert abs(sigma[0, [0, 1, 3]].sum() - eta * smax) < 1e-5
    assert (sigma <= smax + ATOL).all()


# ---------------------------------------------------------------------------
# Nucleus filtering
# ---------------------------------------------------------------------------


def test_nucleus_sample_support_and_frequencies():
    """Nucleus sampling draws only from the smallest prefix with cumulative
    mass >= p, with renormalised probabilities.

    Source: ReMDM Sec 5 (nucleus sampling, Holtzman et al.). For probs
    [0.5, 0.3, 0.15, 0.05] and p=0.9 the support is {0,1,2} with
    renormalised probs [0.5263, 0.3158, 0.1579]. Statistical: 4096 draws;
    per-token sigma = sqrt(p_i(1-p_i)/4096) <= 0.0078; bound = 4 sigma
    = 0.0313.
    """
    probs = jnp.array([0.5, 0.3, 0.15, 0.05])
    logits = jnp.log(jnp.broadcast_to(probs, (4096, 1, 4)))
    tokens = np.asarray(_nucleus_sample(jax.random.PRNGKey(0), logits, 0.9)).ravel()
    counts = np.bincount(tokens, minlength=4) / tokens.size
    assert counts[3] == 0.0, "out-of-nucleus token was sampled"
    renorm = np.array([0.5, 0.3, 0.15]) / 0.95
    assert np.all(np.abs(counts[:3] - renorm) < 0.0313)


# ---------------------------------------------------------------------------
# Prefix locking through the full corrected chain
# ---------------------------------------------------------------------------


def test_prefix_lock_survives_loop_mode_chain():
    """A locked prefix is bit-identical after a full ReMDM chain with loop
    mode and the conf strategy active, and the output contains no MASK.

    Source: planning-as-inpainting (Diffuser Sec 3.3: conditioned values
    are fixed throughout denoising) on top of ReMDM Alg 1 / Sec 4.2 loop.
    Uses a uniform-logits stub model.
    """
    B, H, V = 3, 8, 5
    fn, _ = SCHEDULE_MAP["cosine"]

    def apply_fn(params, obs, z, t, rng):
        return jnp.zeros((obs.shape[0], H, V))

    history = jnp.tile(jnp.arange(H, dtype=jnp.int32) % V, (B, 1))
    hist_len = jnp.array([0, 3, H], dtype=jnp.int32)
    plan = sample_plan_inpainting(
        apply_fn, None, jax.random.PRNGKey(0), jnp.zeros((B, 3)),
        history, hist_len, V, H, diffusion_steps=6, schedule_fn=fn,
        remask_strategy="conf", eta=0.5, use_loop=True, t_on=0.7, t_off=0.3,
        temperature=0.5, top_p=0.95,
    )
    plan = np.asarray(plan)
    assert (plan != V).all(), "output contains MASK tokens"
    assert (plan[1, :3] == np.asarray(history)[1, :3]).all()
    assert (plan[2] == np.asarray(history)[2]).all()