File size: 6,584 Bytes
d8bc908
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Unit tests for MultiHeadLatentAttention, ContextAttentionScheduler, RoPE utilities."""
import math
import torch
import sys
import os

sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))

from arbitor.attention.mla import (
    MultiHeadLatentAttention, apply_rotary_emb, precompute_freqs_cis,
)
from arbitor.attention.context_attention import ContextAttentionScheduler
from arbitor.attention.kv_ledger import KVLedger


def _default_mla():
    return MultiHeadLatentAttention(
        dim=256, n_heads=4, kv_lora_rank=16,
        qk_nope_head_dim=24, qk_rope_head_dim=8, v_head_dim=24,
    )


def test_mla_construct():
    mla = _default_mla()
    assert mla.dim == 256
    assert mla.n_heads == 4
    assert mla.kv_lora_rank == 16
    assert mla.qk_head_dim == 32
    print(" PASS test_mla_construct")


def test_mla_shape():
    mla = _default_mla()
    x = torch.randn(1, 4, 256)
    kv_cache = torch.randn(8, 16)
    pe_cache = torch.randn(8, 8)
    out = mla(x, kv_cache, pe_cache)
    assert out.shape == (1, 4, 256), f"shape {out.shape}"
    assert torch.isfinite(out).all()
    print(" PASS test_mla_shape")


def _get_wkv_b_eff(mla, n_heads, kv_lora_rank):
    """Get effective wkv_b weight from TernaryScaleTensor."""
    T = mla.wkv_b._get_T()
    S = mla.wkv_b._get_S()
    W = (T * S).view(n_heads, -1, kv_lora_rank)
    return W


def test_mla_absorb_vs_naive():
    for seed in [42, 123, 256]:
        torch.manual_seed(seed)
        dim = 128
        n_heads = 2
        kv_lora_rank = 8
        qk_nope = 16
        qk_rope = 8
        v_dim = 16

        mla = MultiHeadLatentAttention(
            dim=dim, n_heads=n_heads, kv_lora_rank=kv_lora_rank,
            qk_nope_head_dim=qk_nope, qk_rope_head_dim=qk_rope, v_head_dim=v_dim,
        )

        x = torch.randn(1, 4, dim)
        kv_cache = torch.randn(8, kv_lora_rank)
        pe_cache = torch.randn(8, qk_rope)
        absorb_out = mla(x, kv_cache, pe_cache)

        wkv_b = _get_wkv_b_eff(mla, n_heads, kv_lora_rank)
        kv_nope = torch.einsum("hdc,tc->thd", wkv_b[:, :qk_nope], kv_cache)
        kv_full_k = torch.cat([kv_nope, pe_cache.unsqueeze(1).expand(-1, n_heads, -1)], dim=-1)
        kv_full_v = torch.einsum("hdc,tc->thd", wkv_b[:, -v_dim:], kv_cache)
        naive = _naive_attention(mla, x, kv_full_k, kv_full_v, pe_cache)

        diff = (absorb_out - naive).abs().max().item()
        assert diff < 1e-4, f"seed={seed} diff={diff}"
    print(" PASS test_mla_absorb_vs_naive")


def _naive_attention(mla, x, kv_full_k, kv_full_v, pe_cache):
    bsz, seqlen, _ = x.shape
    q = mla.wq(mla.wq_norm(x))
    q = q.view(bsz, seqlen, mla.n_heads, mla.qk_head_dim)

    scores = torch.einsum("bshd,thd->bsht", q, kv_full_k) * mla.softmax_scale

    if seqlen > 1:
        causal = torch.triu(
            torch.full((seqlen, kv_full_k.shape[0]), float('-inf'), device=x.device),
            diagonal=1
        )
        scores = scores + causal.unsqueeze(0).unsqueeze(2)

    scores = scores.softmax(dim=-1, dtype=torch.float32)

    attn = torch.einsum("bsht,thd->bshd", scores, kv_full_v)
    attn = attn.flatten(2)
    return mla.wo(attn)


def test_mla_gradient_flow():
    mla = _default_mla()
    x = torch.randn(1, 4, 256, requires_grad=True)
    kv_cache = torch.randn(8, 16)
    pe_cache = torch.randn(8, 8)

    out = mla(x, kv_cache, pe_cache)
    loss = out.sum()
    loss.backward()

    assert x.grad is not None, "input grad is None"
    assert x.grad.abs().sum().item() > 0, "input grad is zero"
    print(" PASS test_mla_gradient_flow")


def test_mla_causal_mask():
    mla = _default_mla()
    x = torch.randn(1, 8, 256)
    kv_cache = torch.randn(12, 16)
    pe_cache = torch.randn(12, 8)

    out = mla(x, kv_cache, pe_cache, mask=None)
    assert out.shape == (1, 8, 256)

    mla2 = _default_mla()
    out2 = mla2(x, kv_cache, pe_cache, mask=None)
    assert torch.isfinite(out2).all()
    print(" PASS test_mla_causal_mask")


def test_apply_rotary_emb():
    x = torch.randn(1, 4, 2, 8)
    freqs_cis = torch.polar(
        torch.ones(4, 4),
        torch.linspace(0, math.pi, 4 * 4).reshape(4, 4),
    )
    out = apply_rotary_emb(x, freqs_cis)
    assert out.shape == (1, 4, 2, 8), f"shape {out.shape}"
    assert not torch.allclose(out, x), "rotation did nothing"
    print(" PASS test_apply_rotary_emb")


def test_precompute_freqs_cis():
    freqs = precompute_freqs_cis(dim=32, end=100)
    assert freqs.shape == (100, 16), f"shape {freqs.shape}"
    assert torch.is_complex(freqs)
    assert freqs.imag.abs().sum().item() > 0, "imag part is zero"
    print(" PASS test_precompute_freqs_cis")


def test_context_scheduler():
    scheduler = ContextAttentionScheduler(dim=256)
    x = torch.randn(1, 4, 256)

    ledger = KVLedger(32)
    for i in range(20):
        ledger.append(i)

    out = scheduler(x, ledger)
    assert out.shape == (1, 4, 256), f"shape {out.shape}"
    assert torch.isfinite(out).all()
    print(" PASS test_context_scheduler")


def test_context_scheduler_empty_ledger():
    scheduler = ContextAttentionScheduler(dim=256)
    x = torch.randn(1, 4, 256)
    ledger = KVLedger(32)

    out = scheduler(x, ledger)
    assert out.shape == (1, 4, 256)
    assert torch.isfinite(out).all()
    print(" PASS test_context_scheduler_empty_ledger")


def test_context_scheduler_gate():
    scheduler = ContextAttentionScheduler(dim=256)
    x = torch.randn(1, 4, 256)
    ledger = KVLedger(32)
    for i in range(20):
        ledger.append(i)

    out = scheduler(x, ledger)
    gate_val = torch.sigmoid(scheduler.gate(x.mean(dim=1, keepdim=True)))
    assert gate_val.shape == (1, 1, 1)
    assert 0 < gate_val.item() < 1
    print(" PASS test_context_scheduler_gate")


def test_context_scheduler_hca_shape_mismatch_regression():
    scheduler = ContextAttentionScheduler(dim=256)
    x = torch.randn(1, 4, 256)
    ledger = KVLedger(256)
    for i in range(160):
        ledger.append(i)

    out = scheduler(x, ledger)
    assert out.shape == (1, 4, 256), f"shape {out.shape}"
    assert torch.isfinite(out).all()
    print(" PASS test_context_scheduler_hca_shape_mismatch_regression")


if __name__ == "__main__":
    test_mla_construct()
    test_mla_shape()
    test_mla_absorb_vs_naive()
    test_mla_gradient_flow()
    test_mla_causal_mask()
    test_apply_rotary_emb()
    test_precompute_freqs_cis()
    test_context_scheduler()
    test_context_scheduler_empty_ledger()
    test_context_scheduler_gate()
    test_context_scheduler_hca_shape_mismatch_regression()
    print("\nAll MLA + scheduler tests PASS")