libmir-cuda 0.2.0

CUDA inference backend for libmir
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

use models::{
    layout::{DecoderConfig, EncoderConfig},
    semantic::MixerSpec,
};
use runtime::trace::{TraceDecoder, TraceWeights};

use super::{LoadedModel, mlp_layout};

pub(super) fn attention_counts(model: &LoadedModel) -> (usize, usize) {
    if let Some(semantic) = model.semantic() {
        return semantic
            .decoder
            .layers
            .iter()
            .fold((0, 0), |(full, sliding), layer| match &layer.mixer {
                MixerSpec::SoftmaxAttention(attention) if attention.window.is_some() => {
                    (full, sliding + 1)
                },
                MixerSpec::SoftmaxAttention(_) => (full + 1, sliding),
                MixerSpec::LinearAttention(_) => (full, sliding),
            });
    }
    (model.encoder.as_ref().map_or(0, |encoder| encoder.num_hidden_layers), 0)
}

pub(super) fn execution(
    decoder: Option<&DecoderConfig>,
    encoder: Option<&EncoderConfig>,
    full: usize,
    sliding: usize,
) -> TraceDecoder {
    if let Some(decoder) = decoder {
        return decoder_trace(decoder, full, sliding);
    }
    encoder.map_or_else(empty_execution, encoder_trace)
}

pub(super) fn weights(model: &LoadedModel) -> TraceWeights {
    if let Some(encoder) = model.encoder.as_ref() {
        return TraceWeights {
            token_embeddings: "new.embeddings.word_embeddings".into(),
            final_norm: "post-attention and post-MLP LayerNorm".into(),
            output_head: "pooler tanh and scalar classifier".into(),
            output_tied: false,
            layer_count: encoder.num_hidden_layers,
            attention_layout: "packed QKV with bidirectional F16 attention".into(),
            mlp_layout: mlp_layout(model).into(),
            linear_bias_count: encoder.num_hidden_layers.saturating_mul(4).saturating_add(2),
        };
    }
    let Some(decoder) = model.decoder.as_ref() else {
        return empty_weights();
    };
    TraceWeights {
        token_embeddings: "config-discovered BF16 embedding".into(),
        final_norm: "config-discovered BF16 RMSNorm".into(),
        output_head: "tied or independent checkpoint projection".into(),
        output_tied: decoder.tie_word_embeddings,
        layer_count: decoder.num_hidden_layers,
        attention_layout: "fused QKV with causal paged GQA".into(),
        mlp_layout: mlp_layout(model).into(),
        linear_bias_count: 0,
    }
}

fn encoder_trace(encoder: &EncoderConfig) -> TraceDecoder {
    TraceDecoder {
        layers: encoder.num_hidden_layers,
        hidden_size: encoder.hidden_size,
        intermediate_size: encoder.intermediate_size,
        vocab_size: encoder.vocab_size,
        attention_heads: encoder.num_attention_heads,
        kv_heads: encoder.num_attention_heads,
        head_dim: encoder.head_dim,
        global_head_dim: None,
        global_kv_heads: None,
        full_attention_layers: encoder.num_hidden_layers,
        sliding_attention_layers: 0,
        max_position_embeddings: encoder.max_position_embeddings,
        rope_theta: encoder.rope_theta,
        full_attention_rope_theta: None,
        sliding_attention_rope_theta: None,
        sliding_window: None,
        num_experts: None,
        top_k_experts: None,
        moe_intermediate_size: None,
        hidden_activation: Some(encoder.hidden_activation.clone()),
        final_logit_softcapping: None,
    }
}

fn decoder_trace(decoder: &DecoderConfig, full: usize, sliding: usize) -> TraceDecoder {
    TraceDecoder {
        layers: decoder.num_hidden_layers,
        hidden_size: decoder.hidden_size,
        intermediate_size: decoder.intermediate_size,
        vocab_size: decoder.vocab_size,
        attention_heads: decoder.num_attention_heads,
        kv_heads: decoder.num_key_value_heads,
        head_dim: decoder.head_dim,
        global_head_dim: decoder.global_head_dim,
        global_kv_heads: decoder.num_global_key_value_heads,
        full_attention_layers: full,
        sliding_attention_layers: sliding,
        max_position_embeddings: decoder.max_position_embeddings,
        rope_theta: decoder.rope_theta,
        full_attention_rope_theta: decoder.full_attention_rope_theta,
        sliding_attention_rope_theta: decoder.sliding_attention_rope_theta,
        sliding_window: decoder.sliding_window,
        num_experts: decoder.num_experts,
        top_k_experts: decoder.top_k_experts,
        moe_intermediate_size: decoder.moe_intermediate_size,
        hidden_activation: decoder.hidden_activation.clone(),
        final_logit_softcapping: decoder.final_logit_softcapping,
    }
}

fn empty_execution() -> TraceDecoder {
    TraceDecoder {
        layers: 0,
        hidden_size: 0,
        intermediate_size: 0,
        vocab_size: 0,
        attention_heads: 0,
        kv_heads: 0,
        head_dim: 0,
        global_head_dim: None,
        global_kv_heads: None,
        full_attention_layers: 0,
        sliding_attention_layers: 0,
        max_position_embeddings: 0,
        rope_theta: None,
        full_attention_rope_theta: None,
        sliding_attention_rope_theta: None,
        sliding_window: None,
        num_experts: None,
        top_k_experts: None,
        moe_intermediate_size: None,
        hidden_activation: None,
        final_logit_softcapping: None,
    }
}

fn empty_weights() -> TraceWeights {
    TraceWeights {
        token_embeddings: "unknown".into(),
        final_norm: "unknown".into(),
        output_head: "unknown".into(),
        output_tied: false,
        layer_count: 0,
        attention_layout: "unknown".into(),
        mlp_layout: "unknown".into(),
        linear_bias_count: 0,
    }
}