File size: 16,157 Bytes
4a28d4d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
372
373
374
import torch

from lmdeploy.pytorch.engine.logits_process import SamplingInputs
from lmdeploy.pytorch.spec_decode.reject_sampler import (
    PLACEHOLDER_TOKEN_ID,
    _extract_outputs,
    rejection_greedy_sample_kernel,
    rejection_sample,
    sample_recovered_tokens_kernel,
    torch_greedy_rejection_sample,
)

device = 'cuda' if torch.cuda.is_available() else 'cpu'


def _make_peaked_logits(token_ids_2d, vocab_size):
    """Build logits where argmax(dim=-1) == token_ids_2d.

    token_ids_2d: list[list[int]] or Tensor [batch, num_spec]
    """
    if isinstance(token_ids_2d, torch.Tensor):
        token_ids_2d = token_ids_2d.tolist()
    batch_size = len(token_ids_2d)
    num_spec = len(token_ids_2d[0])
    logits = torch.full((batch_size, num_spec, vocab_size), -100.0, device=device)
    for b in range(batch_size):
        for s in range(num_spec):
            logits[b, s, token_ids_2d[b][s]] = 100.0
    return logits


def _extract_valid(output_token_ids):
    """Return list-of-lists of non-placeholder token ids."""
    out = []
    for row in output_token_ids:
        out.append(row[row != PLACEHOLDER_TOKEN_ID].tolist())
    return out


class TestRejectSample:
    """Tests for rejection_sample and torch_greedy_rejection_sample.

    Uses hand-crafted token ids so the expected output is obvious.
    """

    # ----- greedy: rejection_sample -----

    def test_greedy_all_match(self):
        """All draft == target -> accept all + bonus."""
        target_logits = _make_peaked_logits([[10, 20, 30], [10, 20, 30]], 64)
        draft = torch.tensor([[10, 20, 30], [10, 20, 30]], dtype=torch.long, device=device)
        bonus = torch.tensor([99, 88], dtype=torch.long, device=device)
        si = SamplingInputs(max_top_k=1)

        out, rej, last = rejection_sample(target_logits, draft, bonus, sampling_inputs=si)

        assert out.shape == (2, 4)
        assert (rej == 0).all()
        assert _extract_valid(out) == [[10, 20, 30, 99], [10, 20, 30, 88]]
        assert last.tolist() == [99, 88]

    def test_greedy_first_mismatch(self):
        """Mismatch at position 0 -> keep only target[0]."""
        target_logits = _make_peaked_logits([[10, 20, 30], [10, 20, 30]], 64)
        draft = torch.tensor([[11, 20, 30], [11, 20, 30]], dtype=torch.long, device=device)
        bonus = torch.tensor([99, 88], dtype=torch.long, device=device)
        si = SamplingInputs(max_top_k=1)

        out, rej, last = rejection_sample(target_logits, draft, bonus, sampling_inputs=si)

        assert (rej == 3).all()
        assert _extract_valid(out) == [[10], [10]]
        assert last.tolist() == [10, 10]

    def test_greedy_middle_mismatch(self):
        """Mismatch at position 2 -> keep target[0:3], no bonus."""
        target_logits = _make_peaked_logits([[10, 20, 30], [10, 20, 30]], 64)
        draft = torch.tensor([[10, 20, 31], [10, 20, 31]], dtype=torch.long, device=device)
        bonus = torch.tensor([99, 88], dtype=torch.long, device=device)
        si = SamplingInputs(max_top_k=1)

        out, rej, last = rejection_sample(target_logits, draft, bonus, sampling_inputs=si)

        assert (rej == 1).all()
        assert _extract_valid(out) == [[10, 20, 30], [10, 20, 30]]
        assert last.tolist() == [30, 30]

    def test_greedy_mixed_mismatch_positions(self):
        """Each row has a different first-mismatch position.

        Row 0: all match       -> [10, 20, 30, 99]
        Row 1: mismatch at 0   -> [10]
        Row 2: mismatch at 1   -> [10, 20]
        Row 3: mismatch at 2   -> [10, 20, 30]
        """
        target_logits = _make_peaked_logits([[10, 20, 30]] * 4, 64)
        draft = torch.tensor([
            [10, 20, 30],
            [11, 20, 30],
            [10, 21, 30],
            [10, 20, 31],
        ],
                             dtype=torch.long,
                             device=device)
        bonus = torch.tensor([99, 88, 77, 66], dtype=torch.long, device=device)
        si = SamplingInputs(max_top_k=1)

        out, rej, last = rejection_sample(target_logits, draft, bonus, sampling_inputs=si)

        assert rej.tolist() == [0, 3, 2, 1]
        assert _extract_valid(out) == [
            [10, 20, 30, 99],
            [10],
            [10, 20],
            [10, 20, 30],
        ]
        assert last.tolist() == [99, 10, 20, 30]

    def test_greedy_all_rejected(self):
        """Every draft wrong -> only first target token kept."""
        target_logits = _make_peaked_logits([[10, 20, 30], [10, 20, 30]], 64)
        draft = torch.tensor([[11, 21, 31], [11, 21, 31]], dtype=torch.long, device=device)
        bonus = torch.tensor([99, 88], dtype=torch.long, device=device)
        si = SamplingInputs(max_top_k=1)

        out, rej, last = rejection_sample(target_logits, draft, bonus, sampling_inputs=si)

        assert (rej == 3).all()
        assert _extract_valid(out) == [[10], [10]]
        assert last.tolist() == [10, 10]

    # ----- greedy: compare rejection_sample vs torch_greedy_rejection_sample -----

    def test_greedy_triton_vs_torch_all_match(self):
        """Both implementations agree when all drafts match."""
        target_logits = _make_peaked_logits([[10, 20, 30]] * 4, 64)
        draft = torch.tensor([[10, 20, 30]] * 4, dtype=torch.long, device=device)
        bonus = torch.tensor([99, 88, 77, 66], dtype=torch.long, device=device)
        si = SamplingInputs(max_top_k=1)

        t_out, t_rej, t_last = rejection_sample(target_logits, draft, bonus, sampling_inputs=si)
        p_out, p_rej, p_last = torch_greedy_rejection_sample(target_logits, draft, bonus, sampling_inputs=si)

        assert torch.equal(t_out, p_out)
        assert torch.equal(t_rej, p_rej)
        assert torch.equal(t_last, p_last)

    def test_greedy_triton_vs_torch_mixed(self):
        """Both implementations agree with mixed mismatch positions."""
        target_logits = _make_peaked_logits([[10, 20, 30]] * 4, 64)
        draft = torch.tensor(
            [
                [10, 20, 30],  # all match
                [11, 20, 30],  # pos 0
                [10, 21, 30],  # pos 1
                [10, 20, 31],  # pos 2
            ],
            dtype=torch.long,
            device=device)
        bonus = torch.tensor([99, 88, 77, 66], dtype=torch.long, device=device)
        si = SamplingInputs(max_top_k=1)

        t_out, t_rej, t_last = rejection_sample(target_logits, draft, bonus, sampling_inputs=si)
        p_out, p_rej, p_last = torch_greedy_rejection_sample(target_logits, draft, bonus, sampling_inputs=si)

        assert torch.equal(t_out, p_out), f'output mismatch: triton={t_out}, torch={p_out}'
        assert torch.equal(t_rej, p_rej), f'num_rejected: triton={t_rej}, torch={p_rej}'
        assert torch.equal(t_last, p_last), f'last_token: triton={t_last}, torch={p_last}'

    def test_greedy_triton_vs_torch_large_batch(self):
        """Both implementations agree with large batch + random mismatches."""
        batch_size, num_spec, vocab = 32, 6, 128
        target_logits = torch.randn(batch_size, num_spec, vocab, device=device)
        target_argmax = target_logits.argmax(dim=-1)
        draft = target_argmax.clone()

        torch.manual_seed(42)
        flip = torch.rand(batch_size, num_spec) < 0.3
        for i in range(batch_size):
            for j in range(num_spec):
                if flip[i, j]:
                    draft[i, j] = (target_argmax[i, j] + 1) % vocab

        bonus = torch.randint(0, vocab, (batch_size, ), device=device)
        si = SamplingInputs(max_top_k=1)

        t_out, t_rej, t_last = rejection_sample(target_logits, draft, bonus, sampling_inputs=si)
        p_out, p_rej, p_last = torch_greedy_rejection_sample(target_logits, draft, bonus, sampling_inputs=si)

        assert torch.equal(t_out, p_out)
        assert torch.equal(t_rej, p_rej)
        assert torch.equal(t_last, p_last)

    # ----- mixed greedy/random -----

    def test_mixed_batch(self):
        """Mixed batch: greedy (top_k=1) and random sequences."""
        target_logits = _make_peaked_logits([[5, 6, 7], [5, 6, 7], [5, 6, 7], [5, 6, 7]], 32)
        draft = target_logits.argmax(dim=-1)
        bonus = torch.tensor([99, 88, 77, 66], dtype=torch.long, device=device)

        top_k = torch.tensor([1, 10, 1, 10], device=device)
        si = SamplingInputs(max_top_k=10, top_k=top_k)

        out, rej, last = rejection_sample(target_logits, draft, bonus, sampling_inputs=si)

        # Greedy rows (0, 2) must accept all since draft == target
        assert rej[0] == 0
        assert rej[2] == 0
        assert last[0] == 99
        assert last[2] == 77

    # ----- random: no draft_probs (MTP/n-gram) -----

    def test_random_no_draft_probs(self):
        """Random sampling with draft_probs=None (MTP/n-gram mode)."""
        target_logits = _make_peaked_logits([[5, 6, 7], [5, 6, 7]], 32)
        draft = torch.tensor([[5, 6, 7], [5, 6, 7]], dtype=torch.long, device=device)
        bonus = torch.tensor([99, 88], dtype=torch.long, device=device)

        top_k = torch.tensor([10, 10], device=device)
        si = SamplingInputs(max_top_k=10, top_k=top_k)

        out, rej, last = rejection_sample(target_logits, draft, bonus, sampling_inputs=si, draft_probs=None)

        assert out.shape == (2, 4)
        assert (last >= 0).all()

    # ----- random: all reject (peaked target, wrong draft) -----

    def test_random_all_reject(self):
        """Random: draft picks wrong token, peaked target -> reject at pos 0."""
        vocab = 32
        target_logits = _make_peaked_logits([[5, 6, 7], [5, 6, 7]], vocab)
        draft = torch.tensor([[1, 1, 1], [1, 1, 1]], dtype=torch.long, device=device)
        draft_probs = torch.zeros(2, 3, vocab, device=device)
        draft_probs[:, :, 1] = 1.0
        bonus = torch.tensor([99, 88], dtype=torch.long, device=device)

        top_k = torch.tensor([10, 10], device=device)
        si = SamplingInputs(max_top_k=10, top_k=top_k)

        out, rej, last = rejection_sample(target_logits, draft, bonus, sampling_inputs=si, draft_probs=draft_probs)

        # Reject at pos 0 -> 1 recovered token per row
        valid = _extract_valid(out)
        for i in range(2):
            assert len(valid[i]) == 1, f'row {i}: expected 1 token, got {valid[i]}'

    # ----- _extract_outputs -----

    def test_extract_outputs(self):
        """Test _extract_outputs helper function."""
        output_token_ids = torch.tensor([
            [5, 6, 7, 8],  # all valid
            [10, 11, -1, -1],  # 2 valid
        ])
        out, rej, last = _extract_outputs(output_token_ids, num_spec_tokens=3)
        assert rej.tolist() == [0, 2]
        assert last.tolist() == [8, 11]


# ---------------------------------------------------------------------------
# Triton kernel tests (direct kernel invocations)
# ---------------------------------------------------------------------------


class TestTritonKernels:

    def test_greedy_kernel_mixed(self):
        """rejection_greedy_sample_kernel with mixed match/mismatch rows.

        Row 0: all match       -> [10, 20, 30, 40, 50]
        Row 1: mismatch at 2   -> [10, 20, 30, -1, -1]
        Row 2: mismatch at 0   -> [10, -1, -1, -1, -1]
        """
        batch_size, num_spec = 3, 4
        target_argmax = torch.tensor([[10, 20, 30, 40]] * 3, dtype=torch.long, device=device)
        draft = torch.tensor([
            [10, 20, 30, 40],
            [10, 20, 99, 40],
            [99, 20, 30, 40],
        ], dtype=torch.long, device=device)
        bonus = torch.tensor([50, 51, 52], dtype=torch.long, device=device)
        out = torch.full((batch_size, num_spec + 1), PLACEHOLDER_TOKEN_ID, dtype=torch.long, device=device)

        rejection_greedy_sample_kernel[(batch_size, )](out, draft, target_argmax, bonus, None, num_spec)

        assert out[0].tolist() == [10, 20, 30, 40, 50]
        assert out[1].tolist() == [10, 20, 30, PLACEHOLDER_TOKEN_ID, PLACEHOLDER_TOKEN_ID]
        assert out[2].tolist() == [
            10, PLACEHOLDER_TOKEN_ID, PLACEHOLDER_TOKEN_ID, PLACEHOLDER_TOKEN_ID, PLACEHOLDER_TOKEN_ID
        ]

    def test_greedy_kernel_with_is_greedy_mask(self):
        """rejection_greedy_sample_kernel skips non-greedy rows.

        Row 0: greedy, all match -> [10, 20, 30, 50]
        Row 1: not greedy        -> all placeholder (untouched)
        """
        batch_size, num_spec = 2, 3
        target_argmax = torch.tensor([[10, 20, 30]] * 2, dtype=torch.long, device=device)
        draft = target_argmax.clone()
        bonus = torch.tensor([50, 51], dtype=torch.long, device=device)
        out = torch.full((batch_size, num_spec + 1), PLACEHOLDER_TOKEN_ID, dtype=torch.long, device=device)
        is_greedy = torch.tensor([True, False], dtype=torch.bool, device=device)

        rejection_greedy_sample_kernel[(batch_size, )](out, draft, target_argmax, bonus, is_greedy, num_spec)

        assert out[0].tolist() == [10, 20, 30, 50]
        assert (out[1] == PLACEHOLDER_TOKEN_ID).all()

    def test_recovered_tokens_with_draft_probs(self):
        """sample_recovered_tokens_kernel: recovered token = argmax of
        max(0, target - draft) * inv_q.

        target = [0, 0, 0, 0.9, 0, 0.1, 0, 0]
        draft  = [0, 0, 0, 0.9, 0, 0,   0, 0]
        diff   = [0, 0, 0, 0,   0, 0.1, 0, 0]  -> token 5
        """
        batch_size, num_spec, vocab = 2, 2, 8
        target_probs = torch.zeros(batch_size, num_spec, vocab, device=device, dtype=torch.float32)
        target_probs[:, :, 3] = 0.9
        target_probs[:, :, 5] = 0.1

        draft_probs = torch.zeros_like(target_probs)
        draft_probs[:, :, 3] = 0.9

        draft_ids = torch.full((batch_size, num_spec), 3, dtype=torch.long, device=device)
        inv_q = torch.ones(batch_size, vocab, device=device, dtype=torch.float32)
        recovered = torch.empty(batch_size, num_spec, dtype=torch.long, device=device)

        sample_recovered_tokens_kernel[(batch_size, num_spec)](recovered,
                                                               draft_ids,
                                                               draft_probs,
                                                               target_probs,
                                                               inv_q,
                                                               num_spec,
                                                               vocab,
                                                               8,
                                                               NO_DRAFT_PROBS=False)

        assert (recovered == 5).all()

    def test_recovered_tokens_no_draft_probs(self):
        """sample_recovered_tokens_kernel with NO_DRAFT_PROBS: draft token
        is masked out, recovered = argmax of remaining target probs.

        target   = [0, 0, 0.6, 0, 0, 0.3, 0, 0.1]
        draft_id = 2 (masked out)
        remaining= [0, 0, 0,   0, 0, 0.3, 0, 0.1]  -> token 5
        """
        batch_size, num_spec, vocab = 2, 2, 8
        target_probs = torch.zeros(batch_size, num_spec, vocab, device=device, dtype=torch.float32)
        target_probs[:, :, 2] = 0.6
        target_probs[:, :, 5] = 0.3
        target_probs[:, :, 7] = 0.1

        draft_ids = torch.full((batch_size, num_spec), 2, dtype=torch.long, device=device)
        inv_q = torch.ones(batch_size, vocab, device=device, dtype=torch.float32)
        recovered = torch.empty(batch_size, num_spec, dtype=torch.long, device=device)

        sample_recovered_tokens_kernel[(batch_size, num_spec)](recovered,
                                                               draft_ids,
                                                               None,
                                                               target_probs,
                                                               inv_q,
                                                               num_spec,
                                                               vocab,
                                                               8,
                                                               NO_DRAFT_PROBS=True)

        assert (recovered == 5).all()