libmir-cuda 0.1.0

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
156
157
158
159

use mircuda::{DeviceBuffer, bf16};

use super::{
    BatchedDecodeMoeLayer, CudaBackend, DecodeMoeBlockConfig, DecodeMoeBlockExecutor,
    DecodeMoeBlockWeights, NvFp4ExpertBank, PrefillMoeBlockBf16,
    graph::weights::CapturedBlockWeights,
};
use crate::{Error, PagedKvCache, Result};

/// Immutable model-layer resources shared by session-local CUDA executors.
///
/// Expert banks and captured weights retain existing device allocations.
/// [`Self::instantiate`] creates only mutable plans, scratch, K/V state, and a
/// graph for one session.
#[derive(Clone)]
pub struct DecodeMoeLayerTemplate {
    backend: CudaBackend,
    config: DecodeMoeBlockConfig,
    weights: CapturedBlockWeights,
    gate: NvFp4ExpertBank,
    up: NvFp4ExpertBank,
    down: NvFp4ExpertBank,
}

impl CudaBackend {
    /// Prepares immutable routed-MoE resources from uploaded checkpoint
    /// tensors.
    pub fn prepare_decode_moe_layer_template(
        &self,
        config: DecodeMoeBlockConfig,
        weights: DecodeMoeBlockWeights<'_>,
        gate: NvFp4ExpertBank,
        up: NvFp4ExpertBank,
        down: NvFp4ExpertBank,
    ) -> Result<DecodeMoeLayerTemplate> {
        validate_banks(config, &gate, &up, &down)?;
        Ok(DecodeMoeLayerTemplate {
            backend: self.clone(),
            config,
            weights: CapturedBlockWeights::prepare(self, weights)?,
            gate,
            up,
            down,
        })
    }
}

impl DecodeMoeLayerTemplate {
    /// Returns the fixed layer configuration.
    #[must_use]
    pub const fn config(&self) -> DecodeMoeBlockConfig {
        self.config
    }

    /// Creates private execution state for one session without copying weights.
    pub fn instantiate(
        &self,
        input: &DeviceBuffer<bf16>,
        output: &DeviceBuffer<bf16>,
    ) -> Result<DecodeMoeBlockExecutor> {
        let cache = self
            .backend
            .prepare_paged_kv(self.config.attention.layer, self.config.attention.cache)?;
        self.instantiate_with_cache(input, output, cache)
    }

    pub(in crate::backend) fn instantiate_with_cache(
        &self,
        input: &DeviceBuffer<bf16>,
        output: &DeviceBuffer<bf16>,
        cache: PagedKvCache,
    ) -> Result<DecodeMoeBlockExecutor> {
        let hidden = self.config.attention.hidden_size;
        if input.len() != hidden || output.len() != hidden {
            return Err(Error::InvalidDecoderKernel(
                "layer template requires hidden-sized input and output buffers",
            ));
        }
        let block = super::DecodeMoeBlockBf16::new_with_cache(
            &self.backend,
            self.config,
            self.gate.clone(),
            self.up.clone(),
            self.down.clone(),
            cache,
        )?;
        tracing::debug!(
            layer = self.config.attention.layer,
            hidden,
            experts = self.config.experts,
            "instantiated session-local CUDA MoE layer"
        );
        Ok(DecodeMoeBlockExecutor::new(
            block,
            input.clone(),
            self.weights.clone(),
            output.clone(),
        ))
    }

    /// Creates a fixed-token prefill executor sharing this layer's weights.
    pub fn instantiate_prefill(&self, tokens: usize) -> Result<PrefillMoeBlockBf16> {
        self.backend.prepare_prefill_moe_block_bf16(
            self.config,
            self.gate.clone(),
            self.up.clone(),
            self.down.clone(),
            tokens,
        )
    }

    /// Creates a fixed-row grouped-MoE decode executor without copying weights.
    pub fn instantiate_batch(&self, rows: usize) -> Result<BatchedDecodeMoeLayer> {
        let cache = self
            .backend
            .prepare_paged_kv(self.config.attention.layer, self.config.attention.cache)?;
        self.instantiate_batch_with_cache(rows, cache)
    }

    pub(in crate::backend) fn instantiate_batch_with_cache(
        &self,
        rows: usize,
        cache: PagedKvCache,
    ) -> Result<BatchedDecodeMoeLayer> {
        let block = super::BatchedDecodeMoeBlockBf16::new_with_cache(
            &self.backend,
            self.config,
            self.gate.clone(),
            self.up.clone(),
            self.down.clone(),
            rows,
            cache,
        )?;
        Ok(BatchedDecodeMoeLayer::new(block, self.weights.clone()))
    }
}

fn validate_banks(
    config: DecodeMoeBlockConfig,
    gate: &NvFp4ExpertBank,
    up: &NvFp4ExpertBank,
    down: &NvFp4ExpertBank,
) -> Result<()> {
    let gate = gate.config();
    let up = up.config();
    let down = down.config();
    let valid = gate == up
        && gate.experts == config.experts
        && gate.input_features == config.attention.hidden_size
        && gate.output_features == config.expert_intermediate
        && down.experts == config.experts
        && down.input_features == config.expert_intermediate
        && down.output_features == config.attention.hidden_size;
    if valid {
        Ok(())
    } else {
        Err(Error::InvalidNvFp4("expert banks differ from layer template configuration"))
    }
}