models/spotbugs-exclude.xml at main · integrallis/models · GitHub

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
<?xml version="1.0" encoding="UTF-8"?>
<FindBugsFilter>
    <!-- Suppress warnings about use of incubating Vector API -->
    <Match>
        <Bug pattern="THROWS_METHOD_THROWS_CLAUSE_BASIC_EXCEPTION"/>
    </Match>
    <!-- MemorySegment and internal-state fields throughout the inference pipeline are
         performance-critical views over mmap'd file data or shared state. Defensive copies
         are not feasible (would defeat zero-copy design). -->
    <Match>
        <Package name="com.integrallis.models.backend.purejava.llama"/>
        <Bug pattern="EI_EXPOSE_REP,EI_EXPOSE_REP2"/>
    </Match>
    <Match>
        <Package name="com.integrallis.models.backend.purejava.gguf"/>
        <Bug pattern="EI_EXPOSE_REP,EI_EXPOSE_REP2"/>
    </Match>
    <Match>
        <Package name="com.integrallis.models.backend.purejava.cache"/>
        <Bug pattern="EI_EXPOSE_REP,EI_EXPOSE_REP2"/>
    </Match>
    <Match>
        <Package name="com.integrallis.models.backend.purejava"/>
        <Bug pattern="EI_EXPOSE_REP,EI_EXPOSE_REP2"/>
    </Match>
    <!-- JFR reads event fields reflectively when the event is committed; nothing in Java reads them. -->
    <Match>
        <Class name="com.integrallis.models.backend.purejava.diagnostics.PerformanceCliffEvent"/>
        <Bug pattern="URF_UNREAD_FIELD"/>
    </Match>
    <!-- SopranoBackend.Step is a per-call transfer object. LlamaForwardPass already creates stable
         copies for it; returning those arrays without another defensive copy avoids duplicating
         the vocabulary and acoustic-frame buffers on every generated token. -->
    <Match>
        <Class name="com.integrallis.models.backend.purejava.soprano.SopranoBackend$Step"/>
        <Bug pattern="EI_EXPOSE_REP,EI_EXPOSE_REP2"/>
    </Match>
    <!-- bytes_to_unicode mapping: byte value is known <=255 by construction (loop 0..255) -->
    <Match>
        <Class name="com.integrallis.models.backend.purejava.tokenizer.GgufTokenizer"/>
        <Bug pattern="UC_USELESS_CONDITION"/>
    </Match>
    <!-- PureJavaBackend.load: metadata getString returning Optional is handled by orElse -->
    <Match>
        <Class name="com.integrallis.models.backend.purejava.PureJavaBackend"/>
        <Bug pattern="NP_NULL_ON_SOME_PATH_FROM_RETURN_VALUE"/>
    </Match>

    <!-- models-embedding: VectorCollectionEmbeddingSink intentionally exposes the wrapped
         VectorCollection and EmbeddingBackend via accessors so callers can interleave bulk
         operations and direct backend calls. Defensive copies of these references would defeat
         the bridge's purpose (the entire point is to share state with the downstream collection). -->
    <Match>
        <Class name="com.integrallis.models.embedding.VectorCollectionEmbeddingSink"/>
        <Or>
            <Bug pattern="EI_EXPOSE_REP"/>
            <Bug pattern="EI_EXPOSE_REP2"/>
        </Or>
    </Match>

    <!-- Lfm2ForwardPass hands back its own logits buffer, which is what every decoder in this
         backend does: the contract is that the array is valid until the next forward call, and the
         callers immediately read or clone it. Copying a vocabulary-sized array on every token would
         allocate once per generated token for no benefit, and the returned array is documented as
         transient on the method itself. -->
    <Match>
        <Class name="com.integrallis.models.backend.purejava.lfm2.Lfm2ForwardPass"/>
        <Method name="forward"/>
        <Bug pattern="EI_EXPOSE_REP"/>
    </Match>
    <!-- Gemma3nForwardPass hands back its own logits buffer for the same reason Lfm2ForwardPass does:
         the contract is that the array is valid until the next forward call, the callers read or clone
         it immediately, and copying a vocabulary-sized array per generated token would allocate once
         per token for no benefit. Documented on the method itself. -->
    <Match>
        <Class name="com.integrallis.models.backend.purejava.gemma3n.Gemma3nForwardPass"/>
        <Method name="forward"/>
        <Bug pattern="EI_EXPOSE_REP"/>
    </Match>
    <!-- Deepseek2ForwardPass hands back its own logits buffer, the same contract as every other decoder
         here: valid until the next forward call, read or cloned immediately by the caller. Documented on
         the method. -->
    <Match>
        <Class name="com.integrallis.models.backend.purejava.deepseek2.Deepseek2ForwardPass"/>
        <Method name="forward"/>
        <Bug pattern="EI_EXPOSE_REP"/>
    </Match>
    <!-- A ModelFleet deliberately shares its mutable ModelRouter so applications and framework
         adapters can publish live availability, load, cache, and call feedback to the same router
         that makes decisions. Returning a defensive copy would split that operational state. -->
    <Match>
        <Class name="com.integrallis.models.router.ModelFleet"/>
        <Method name="router"/>
        <Bug pattern="EI_EXPOSE_REP"/>
    </Match>
    <!-- An Attempt is the intentional lifecycle handle for an active reservation. Sharing its
         identity lets provider callbacks and cancellation settle the same reservation exactly once;
         a defensive copy would double-charge or lose that coordination. -->
    <Match>
        <Class name="com.integrallis.models.router.RoutingExecution"/>
        <Method name="beginAttempt"/>
        <Bug pattern="EI_EXPOSE_REP"/>
    </Match>
    <!-- models-decisions: every answer space's criteria() hands back a Map.copyOf result, built by
         AnswerSpaces.requireCriteria, so it is already immutable. Asserted behaviourally by
         AnswerSpaceCriteriaTest rather than trusted. Noul joins the list below because it now holds
         a map too. -->
    <Match>
        <Or>
            <Class name="com.integrallis.models.decisions.Noul"/>
            <Class name="com.integrallis.models.decisions.Choice"/>
            <Class name="com.integrallis.models.decisions.Score"/>
        </Or>
        <Bug pattern="EI_EXPOSE_REP"/>
    </Match>
    <!-- models-decisions: Choice.options and Score.levels hold List.copyOf results, which are
         genuinely immutable; SpotBugs cannot see through the factory. The immutability is pinned
         behaviourally by AnswerSpaceTest, which asserts both that the returned list rejects
         mutation and that the space copies the options it was handed. Copying again on every read
         would allocate once per decision for no gain. -->
    <Match>
        <Or>
            <Class name="com.integrallis.models.decisions.Choice"/>
            <Class name="com.integrallis.models.decisions.Score"/>
        </Or>
        <Bug pattern="EI_EXPOSE_REP"/>
    </Match>
    <!-- These hand back a DecisionArtifact, a final class of final fields with no array accessor.
         The value is immutable; SpotBugs cannot see that through the accessor. -->
    <Match>
        <Or>
            <Class name="com.integrallis.models.decisions.NoulEvaluation"/>
            <Class name="com.integrallis.models.decisions.TypedEvaluation"/>
        </Or>
        <Bug pattern="EI_EXPOSE_REP"/>
    </Match>
</FindBugsFilter>