offgrid 0.3.0

Embeddable offline AI core: safe Rust API over llama.cpp
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

use std::ptr::NonNull;
use std::sync::Arc;

use offgrid_sys as sys;
use serde::{Deserialize, Serialize};

use crate::batch::Batch;
use crate::chat::default_threads;
use crate::{Error, Model, Result};

#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct EmbedParams {
    /// Max tokens per input (and per batch). 0 = the model's training context, capped at 8192.
    pub n_ctx: u32,
    /// Max inputs embedded together in one batch.
    pub max_batch_inputs: u32,
    /// L2-normalize output vectors (so dot product = cosine similarity).
    pub normalize: bool,
    pub n_threads: Option<i32>,
}

impl Default for EmbedParams {
    fn default() -> Self {
        Self { n_ctx: 0, max_batch_inputs: 16, normalize: true, n_threads: None }
    }
}

/// Embedding context for a model with a pooling head (e.g. Qwen3-Embedding, nomic-embed).
pub struct Embedder {
    model: Arc<Model>,
    ctx: NonNull<sys::llama_context>,
    batch: Batch,
    n_batch: usize,
    max_inputs: usize,
    normalize: bool,
    dim: usize,
}

unsafe impl Send for Embedder {}

impl Drop for Embedder {
    fn drop(&mut self) {
        let _guard = crate::backend::setup_lock();
        unsafe { sys::llama_free(self.ctx.as_ptr()) };
    }
}

impl Embedder {
    pub fn new(model: Arc<Model>, params: &EmbedParams) -> Result<Self> {
        let n_ctx = match params.n_ctx {
            0 => model.info().n_ctx_train.clamp(512, 8192),
            n => n,
        };
        let max_inputs = params.max_batch_inputs.max(1);
        let threads = params.n_threads.unwrap_or_else(default_threads);

        // The graph computes vocab-sized logits per ubatch token (~0.6 MB each for Qwen3), so use
        // small ubatches where inputs may span several (causal, last-token pooling).
        let arch = model.meta("general.architecture").unwrap_or_default();
        let causal = model.meta(&format!("{arch}.attention.causal")).is_none_or(|v| v != "false");
        let last_pooling = model.meta(&format!("{arch}.pooling_type")).as_deref() == Some("3");
        let mut cp = unsafe { sys::llama_context_default_params() };
        cp.n_ctx = n_ctx;
        cp.n_batch = n_ctx;
        cp.n_ubatch = if causal && last_pooling { n_ctx.min(512) } else { n_ctx };
        cp.n_seq_max = max_inputs;
        // Shared KV pool: a per-sequence split (n_ctx / n_seq_max) makes long inputs fail.
        cp.kv_unified = true;
        cp.n_threads = threads;
        cp.n_threads_batch = threads;
        cp.embeddings = true;
        cp.no_perf = true;

        let ctx = {
            let _guard = crate::backend::setup_lock();
            unsafe { sys::llama_init_from_model(model.as_ptr(), cp) }
        };
        let ctx = NonNull::new(ctx).ok_or(Error::ContextCreate)?;
        let this_pooling = unsafe { sys::llama_pooling_type(ctx.as_ptr()) };
        // NONE gives per-token vectors and RANK (rerankers) gives n_cls_out scores per
        // sequence, not n_embd floats: reading `dim` floats would overrun the buffer.
        if this_pooling == sys::llama_pooling_type_LLAMA_POOLING_TYPE_NONE
            || this_pooling == sys::llama_pooling_type_LLAMA_POOLING_TYPE_RANK
        {
            let _guard = crate::backend::setup_lock();
            unsafe { sys::llama_free(ctx.as_ptr()) };
            return Err(Error::NoEmbeddings);
        }
        let n_batch = unsafe { sys::llama_n_batch(ctx.as_ptr()) } as usize;
        let max_inputs = unsafe { sys::llama_n_seq_max(ctx.as_ptr()) } as usize;
        let m = model.as_ptr();
        let dim = match unsafe { sys::llama_model_n_embd_out(m) } {
            d if d > 0 => d as usize,
            _ => (unsafe { sys::llama_model_n_embd(m) }) as usize,
        };
        Ok(Self { model, ctx, batch: Batch::new(n_batch, 1), n_batch, max_inputs, normalize: params.normalize, dim })
    }

    /// Embedding vector length.
    pub fn dim(&self) -> usize {
        self.dim
    }

    pub fn model(&self) -> &Arc<Model> {
        &self.model
    }

    pub fn embed_one(&mut self, text: &str) -> Result<Vec<f32>> {
        Ok(self.embed(&[text])?.pop().expect("one input"))
    }

    /// Embed several inputs; returns one vector per input, in order.
    pub fn embed<S: AsRef<str>>(&mut self, inputs: &[S]) -> Result<Vec<Vec<f32>>> {
        let tokenized = inputs
            .iter()
            .map(|s| {
                let t = self.model.tokenize(s.as_ref(), true, false)?;
                if t.is_empty() {
                    return Err(Error::InvalidArgument("empty input".into()));
                }
                if t.len() > self.n_batch {
                    return Err(Error::ContextOverflow { needed: t.len(), n_ctx: self.n_batch });
                }
                Ok(t)
            })
            .collect::<Result<Vec<_>>>()?;

        let mut out = Vec::with_capacity(inputs.len());
        let mut start = 0;
        while start < tokenized.len() {
            // Pack as many inputs as fit the batch.
            let mut end = start;
            let mut total = 0;
            while end < tokenized.len() && end - start < self.max_inputs && total + tokenized[end].len() <= self.n_batch
            {
                total += tokenized[end].len();
                end += 1;
            }
            self.run_batch(&tokenized[start..end], &mut out)?;
            start = end;
        }
        Ok(out)
    }

    fn run_batch(&mut self, group: &[Vec<sys::llama_token>], out: &mut Vec<Vec<f32>>) -> Result<()> {
        let ctx = self.ctx.as_ptr();
        let mem = unsafe { sys::llama_get_memory(ctx) };
        if !mem.is_null() {
            unsafe { sys::llama_memory_clear(mem, true) };
        }
        self.batch.clear();
        for (seq, tokens) in group.iter().enumerate() {
            for (pos, &t) in tokens.iter().enumerate() {
                self.batch.add(t, pos as i32, &[seq as i32], true);
            }
        }
        let m = self.model.as_ptr();
        let rc = unsafe {
            if sys::llama_model_has_encoder(m) && !sys::llama_model_has_decoder(m) {
                sys::llama_encode(ctx, self.batch.raw())
            } else {
                sys::llama_decode(ctx, self.batch.raw())
            }
        };
        if rc != 0 {
            return Err(Error::Decode(rc));
        }
        for seq in 0..group.len() {
            let ptr = unsafe { sys::llama_get_embeddings_seq(ctx, seq as i32) };
            if ptr.is_null() {
                return Err(Error::NoEmbeddings);
            }
            let mut v = unsafe { std::slice::from_raw_parts(ptr, self.dim) }.to_vec();
            if self.normalize {
                let norm = v.iter().map(|x| x * x).sum::<f32>().sqrt();
                if norm > 0.0 {
                    v.iter_mut().for_each(|x| *x /= norm);
                }
            }
            out.push(v);
        }
        Ok(())
    }
}