Gliner2 Rust

This project implements the Gliner2 model in rust with compatibility to the original weights and output of the python training.

cargo add gliner2
# and/or for a cli utility
cargo install gliner2

Development

Pre-commit

Git hooks run the same Rust checks as CI (cargo fmt, cargo clippy on the workspace) plus Ruff on first-party Python (for example under harness/). Paths under reference/ and .tickets/ are excluded from hooks.

Prerequisites: stable Rust with rustfmt and clippy (for example rustup component add rustfmt clippy).

Install pre-commit (either is fine):

uv tool install pre-commit

From the repository root, install the hooks once:

pre-commit install

Optionally validate the whole tree:

pre-commit run --all-files

If you must commit before fixing Clippy, you can skip that hook: SKIP=cargo-clippy git commit (use sparingly; CI still enforces warnings as errors).

Recorded speed (comparison harness)

The [harness/](harness/) scripts run the same release Rust binaries (harness_compare, harness_compare_mt on CPU) against the PyPI gliner2 package. Timing fields are wall-clock milliseconds from a single process: load_model_ms is one-time load; infer_ms is per-fixture forward work (entity harness sums all cases for the total row).

Reproduce (CPU vs CPU): from the repo root, with Hugging Face access for the default model:

uv sync --locked --directory harness
bash harness/run_all.sh
bash harness/run_multitask.sh

The shell wrappers call Python with CUDA_VISIBLE_DEVICES= and --device cpu so PyTorch does not use a discrete NVIDIA GPU and weights stay on CPU, matching the Rust side.

For apples-to-apples timing with the Rust single-forward path, Python uses **batch_size=1**: batch_extract_entities([text], …, batch_size=1) on the entity harness and batch_extract([text], schema, batch_size=1, …) on the multitask harness (instead of relying on extract / extract_entities defaults).

Reading python/rust: for infer times this is (python infer_ms) / (rust infer_ms) per case or for the total line. Values below 1 mean Python spent less time on that measure for these fixtures; above 1 mean Python was slower.

CPU vs CPU (recorded)

Model: fastino/gliner2-base-v1. Recorded: 2026-04-05 (Linux x86_64, local run; numbers vary by machine and load).

Entity harness ([harness/fixtures.json](harness/fixtures.json)) — metadata and per-case infer times:

	Rust	Python
`device_note`	`cpu`	`cpu`
`load_model_ms`	445.0	3874.4
Sum of `infer_ms` over cases	569.9	358.7
`python/rust` (total infer)	—	0.629×

Case id	rust `infer_ms`	python `infer_ms`	`python/rust`
`steve_jobs`	140.0	118.9	0.849×
`tim_cook_iphone`	151.1	85.6	0.566×
`sundar_pichai`	144.3	78.8	0.546×
`microsoft_windows`	134.4	75.4	0.561×

Multitask harness ([harness/fixtures_multitask.json](harness/fixtures_multitask.json)) — single fixture entities_plus_sentiment:

	Rust	Python
`device_note`	`cpu`	`cpu`
`load_model_ms`	409.5	4002.4
Sum of `infer_ms`	157.5	113.9
`python/rust` (total infer)	—	0.724×

These are short-fixture timings. Update the tables when you change the model, fixtures, or harness code in a way that affects performance.

Throughput (local only; not in CI)

These benchmarks are not run in GitHub Actions (see [.github/workflows/ci.yml](.github/workflows/ci.yml)). Run them on your machine when you need larger-sample timing.

The harness uses 64 samples by default, built by cycling texts from [harness/fixtures.json](harness/fixtures.json). Every sample uses the same entity label list ["company", "person", "product", "location", "date"] so Rust [batch_extract_entities](src/extract.rs) and PyPI batch_extract_entities can process the full set with **batch_size=64**. Sequential rows use 64× micro-batches of size 1 on both sides (Rust’s forward loop vs Python batch_extract_entities([t], …, batch_size=1)). Batched rows use one logical batch of 64 on each side.

uv sync --locked --directory harness
bash harness/run_throughput.sh

Optional: bash harness/run_throughput.sh [fixtures.json] [rust_seq_out.json] [rust_batch_out.json] [samples]. The script runs [harness/compare_throughput.py](harness/compare_throughput.py) on the three JSON outputs.

Recorded: 2026-04-05 (Linux x86_64, CPU, CUDA_VISIBLE_DEVICES= + --device cpu on Python). warmup_full_passes=2 over all samples before each timed pass.

Lane	`total_infer_ms` (64 samples)	samples/s	`python/rust` (infer)
Rust sequential (`batch_size` 1)	8813.2	7.26	—
Python sequential (`batch_size` 1)	4843.0	13.22	0.550×
Rust batched (`batch_size` 64)	6794.2	9.42	—
Python batched (`batch_size` 64)	1650.6	38.78	0.243×

Load times from that run: Rust sequential ~492 ms, Rust batched ~467 ms, Python ~2613 ms.

Re-run bash harness/run_throughput.sh to refresh; the script prints the same layout via [harness/compare_throughput.py](harness/compare_throughput.py).

GPU vs GPU (not recorded yet)

Fair comparison needs both implementations on the same device class (for example CUDA on the PyPI side and a GPU inference path in the Rust harness). That pairing is not wired into the harness yet, so no GPU numbers are published here.

	Rust	Python
Device	—	—
`load_model_ms`	—	—
Total `infer_ms`	—	—
`python/rust`	—	—

Usage

Like the Python implementation, this crate supports a full extraction API. You load the model once, build a SchemaTransformer from the tokenizer, then call Extractor methods.

Setup (load model + tokenizer)

use anyhow::Result;
use candle_core::Device;
use candle_nn::VarBuilder;
use candle_transformers::models::debertav2::Config as DebertaConfig;
use gliner2::config::{download_model, ExtractorConfig};
use gliner2::{Extractor, SchemaTransformer};

fn load_extractor(model_id: &str) -> Result<(Extractor, SchemaTransformer)> {
    let files = download_model(model_id)?;
    let device = Device::Cpu;
    let dtype = candle_core::DType::F32;

    let config: ExtractorConfig = serde_json::from_str(&std::fs::read_to_string(&files.config)?)?;
    let mut encoder_config: DebertaConfig =
        serde_json::from_str(&std::fs::read_to_string(&files.encoder_config)?)?;
    let transformer = SchemaTransformer::new(files.tokenizer.to_str().unwrap())?;
    encoder_config.vocab_size = transformer.tokenizer.get_vocab_size(true);

    let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[files.weights], dtype, &device)? };
    let extractor = Extractor::load(config, encoder_config, vb)?;
    Ok((extractor, transformer))
}

Entity extraction (`extract_entities`)

Same idea as Python extract_entities: pass label names; the returned serde_json::Value uses the formatted shape (entities → label → list of strings, when include_spans / include_confidence are false).

use gliner2::ExtractOptions;
use serde_json::json;

let (extractor, transformer) = load_extractor("fastino/gliner2-base-v1")?;
let text = "Apple CEO Tim Cook announced iPhone 15 in Cupertino.";

let entity_types = vec![
    "company".to_string(),
    "person".to_string(),
    "product".to_string(),
    "location".to_string(),
];

let opts = ExtractOptions::default();
let out = extractor.extract_entities(&transformer, text, &entity_types, &opts)?;
// e.g. {"entities":{"company":["Apple"],"person":["Tim Cook"], ...}}

// Optional: character spans + confidence (richer JSON, closer to Python with flags on)
let opts_rich = ExtractOptions {
    include_confidence: true,
    include_spans: true,
    ..Default::default()
};
let _out = extractor.extract_entities(&transformer, text, &entity_types, &opts_rich)?;

Text classification (`classify_text`)

One classification task per call. labels is a JSON array of class names, or an object mapping label → description (like Python).

use gliner2::ExtractOptions;
use serde_json::json;

let (extractor, transformer) = load_extractor("fastino/gliner2-base-v1")?;
let text = "The new phone is amazing and well worth the price.";

// Single-label: scalar string under the task name when format_results is true
let opts = ExtractOptions::default();
let out = extractor.classify_text(
    &transformer,
    text,
    "sentiment",
    json!(["positive", "negative", "neutral"]),
    &opts,
)?;
// e.g. {"sentiment":"positive"}

// Labels with optional descriptions (mirrors Python dict form)
let out2 = extractor.classify_text(
    &transformer,
    text,
    "topic",
    json!({
        "technology": "Tech products and software",
        "business": "Corporate or market news",
        "sports": "Athletics and games"
    }),
    &opts,
)?;

Relation extraction (`extract_relations`)

Pass relation names as a JSON array of strings, or a JSON object (name → description / config), matching Python relations(...).

use gliner2::ExtractOptions;
use serde_json::json;

let (extractor, transformer) = load_extractor("fastino/gliner2-base-v1")?;
let text = "Tim Cook works for Apple, based in Cupertino.";

let opts = ExtractOptions::default();

// List of relation types → formatted results under "relation_extraction"
let out = extractor.extract_relations(
    &transformer,
    text,
    json!(["works_for", "located_in"]),
    &opts,
)?;
// e.g. {"relation_extraction":{"works_for":[["Tim Cook","Apple"]],"located_in":[["Apple","Cupertino"]]}}

// Dict form (descriptions stored like Python; inference uses relation names)
let _out2 = extractor.extract_relations(
    &transformer,
    text,
    json!({
        "works_for": "Employment between person and organization",
        "founded": "Founder relationship"
    }),
    &opts,
)?;

Structured JSON (`extract_json`)

Field specs use the same string syntax as Python extract_json (name::dtype::[choices]::description).

use gliner2::ExtractOptions;
use serde_json::json;

let (extractor, transformer) = load_extractor("fastino/gliner2-base-v1")?;
let text = "iPhone 15 Pro costs $999 and is in stock.";

let structures = json!({
    "product_info": [
        "name::str",
        "price::str",
        "features::list",
        "availability::str::[in_stock|pre_order|sold_out]"
    ]
});
let out = extractor.extract_json(
    &transformer,
    text,
    &structures,
    &ExtractOptions::default(),
)?;

Multi-task builder (`create_schema` + `extract`)

Combines entities, classifications, relations, and structured fields in one encoder pass. Uses the same (extractor, transformer) and text as in the setup section.

use gliner2::{
    create_schema, ExtractOptions, Extractor, SchemaTransformer, ValueDtype,
};
use serde_json::json;

let mut s = create_schema();
s.entities(json!({
    "person": "Names of people",
    "company": "Organization names",
    "product": "Products or offerings",
}));
s.classification_simple("sentiment", json!(["positive", "negative", "neutral"]));
s.classification_simple("category", json!(["technology", "business", "finance", "healthcare"]));
s.relations(json!(["works_for", "founded", "located_in"]));
{
    let _ = s.structure("product_info")
        .field_str("name")
        .field_str("price")
        .field_list("features")
        .field_choices(
            "availability",
            vec![
                "in_stock".into(),
                "pre_order".into(),
                "sold_out".into(),
            ],
            ValueDtype::Str,
        );
}
let (schema_val, meta) = s.build();
let opts = ExtractOptions::default();
let out = extractor.extract(&transformer, text, &schema_val, &meta, &opts)?;

Batch inference

The crate mirrors Python’s batched entry points: records are preprocessed, padded into chunks of at most ExtractOptions::batch_size (default 8), the encoder runs once per chunk, span representations are computed with compute_span_rep_batched when needed, then each row is decoded. Results are returned in input order.

Set batch_size on ExtractOptions for any batch method (it only affects chunking, not single-sample extract_* calls).

Shared schema (one schema for every text)

Use the Extractor helpers; they build the same schema as the single-sample methods and call batch_extract internally.

use gliner2::ExtractOptions;
use serde_json::json;

let (extractor, transformer) = load_extractor("fastino/gliner2-base-v1")?;
let texts: Vec<String> = vec![
    "Apple CEO Tim Cook announced iPhone 15.".into(),
    "Google unveiled Gemini in Mountain View.".into(),
];

let entity_types: Vec<String> = ["company", "person", "product", "location"]
    .into_iter()
    .map(String::from)
    .collect();

let mut opts = ExtractOptions::default();
opts.batch_size = 16;

let results = extractor.batch_extract_entities(&transformer, &texts, &entity_types, &opts)?;
// Vec<serde_json::Value>, one formatted result per input line

let cls = extractor.batch_classify_text(
    &transformer,
    &texts,
    "sentiment",
    json!(["positive", "negative", "neutral"]),
    &opts,
)?;

let rels = extractor.batch_extract_relations(
    &transformer,
    &texts,
    json!(["works_for", "located_in"]),
    &opts,
)?;

let structures = json!({
    "product_info": ["name::str", "price::str"]
});
let json_results = extractor.batch_extract_json(&transformer, &texts, &structures, &opts)?;

Full schema + metadata (`batch_extract`)

For the same multitask flow as extract, build (schema_val, meta) once and run batch_extract with BatchSchemaMode::Shared, or pass per-row schemas and metadata with BatchSchemaMode::PerSample.

use gliner2::{batch_extract, create_schema, BatchSchemaMode, ExtractOptions};
use gliner2::schema::infer_metadata_from_schema;
use serde_json::{json, Value};

let (extractor, transformer) = load_extractor("fastino/gliner2-base-v1")?;
let texts: Vec<String> = vec!["First document.".into(), "Second document.".into()];

// Option A — shared multitask schema from the builder
let mut s = create_schema();
s.entities(json!({ "company": "", "person": "" }));
s.classification_simple("sentiment", json!(["positive", "negative", "neutral"]));
let (schema_val, meta) = s.build();

let opts = ExtractOptions {
    batch_size: 8,
    ..Default::default()
};

let out_shared = batch_extract(
    &extractor,
    &transformer,
    &texts,
    BatchSchemaMode::Shared {
        schema: &schema_val,
        meta: &meta,
    },
    &opts,
)?;

// Option B — per-text JSON schemas (e.g. from config); metadata from infer_metadata_from_schema
let schema_a: Value = json!({ "entities": { "person": "" } });
let schema_b: Value = json!({ "entities": { "location": "" } });
let schemas = vec![schema_a.clone(), schema_b.clone()];
let metas = vec![
    infer_metadata_from_schema(&schema_a),
    infer_metadata_from_schema(&schema_b),
];

let out_per = batch_extract(
    &extractor,
    &transformer,
    &texts,
    BatchSchemaMode::PerSample {
        schemas: &schemas,
        metas: &metas,
    },
    &opts,
)?;

For a shared schema you can also call extractor.batch_extract(&transformer, &texts, &schema_val, &meta, &opts) instead of the free function.

Lower-level reuse: after transform_extract you can run extract_from_preprocessed on one sample if you already have encoder outputs and span tensors; see src/extract.rs.

CLI specification

The command-line interface is not implemented yet. This section specifies the intended gliner2 binary (see default-run in Cargo.toml) so future work can match the library API and Python GLiNER2 behavior.

Install the binary with cargo install gliner2. Inference flags mirror [ExtractOptions](src/extract.rs) (threshold, format_results, include_confidence, include_spans, max_len).

Command overview

flowchart LR
  subgraph sub [Subcommands]
    entities[entities]
    classify[classify]
    relations[relations]
    jsonCmd[json]
    run[run]
  end
  gliner2[gliner2] --> entities
  gliner2 --> classify
  gliner2 --> relations
  gliner2 --> jsonCmd
  gliner2 --> run

Subcommand	Purpose	Library analogue
`gliner2 entities`	Named-entity extraction	`Extractor::extract_entities`, `Schema::entities`
`gliner2 classify`	Text classification (single- or multi-label)	`Extractor::classify_text`, `Schema::classification`
`gliner2 relations`	Relation extraction	`Extractor::extract_relations`, `Schema::relations`
`gliner2 json`	Structured JSON / field extraction	`Extractor::extract_json`, `Schema::extract_json_structures`
`gliner2 run`	Multitask: full engine schema in one pass	`Extractor::extract`

Top-level: gliner2 --help, gliner2 --version, and gliner2 <subcommand> --help.

Global options

These apply to every subcommand unless stated otherwise.

Flag	Description
`--model <HF_REPO_ID>`	Hugging Face model id (default: `fastino/gliner2-base-v1`, same as `harness/` scripts).
`--model-dir <DIR>`	Offline layout: `config.json`, `encoder_config/config.json`, `tokenizer.json`, `model.safetensors` (matches `ModelFiles` from `[download_model](src/config.rs)`).
`--config`, `--encoder-config`, `--tokenizer`, `--weights`	Explicit paths instead of `--model` / `--model-dir`.
`-q`, `-v` / `--log-level`	Quiet / verbose logging (exact mapping is implementation-defined).

Use either Hub resolution (--model) or a local layout (--model-dir or explicit file flags), not a conflicting mix; if both are given, the implementation should reject the invocation with a clear error.

Device and dtype are intentionally unspecified here until the library exposes them; do not document GPU flags until they exist.

Shared inference flags

Flag	Maps to	Default
`--threshold <float>`	`ExtractOptions::threshold`	`0.5`
`--max-len <N>`	`ExtractOptions::max_len`	unset
`--include-confidence`	`include_confidence`	off
`--include-spans`	`include_spans`	off
`--raw` / `--no-format-results`	`format_results = false`	formatted output (`true`)

Batching

The library implements tensor batch inference (Extractor::batch_extract*, ExtractOptions::batch_size); see Batch inference above. The CLI is not implemented yet; the contract below assumes the binary will drive those batched APIs for any input that produces more than one logical record (for example multi-line JSONL or plain text with --text-split line and multiple non-empty lines).

Flag	Description
`--batch-size <N>`	Maximum records per model batch. Default: 8 (implementation may choose a lower value on constrained devices, but must document any deviation).
`--batch-size 1`	Effectively sequential inference (debugging, peak memory limits, or until batched paths are stable).

Single-record inputs (one JSONL line, one JSON object, or --text-split full over an entire file) form a single batch of size 1.

Ordering: Output lines must follow the same order as input records, even when flushing internal batches.

Input and output

Input: final positional argument INPUT, or - for stdin.

Flag	Description
`--text-field <KEY>`	Field containing document text in JSON / JSONL records (default: `text`).
`--id-field <KEY>`	Field to pass through as record id when present (default: `id`).
`--text-split <MODE>`	Plain text: `full` (whole file) or `line` (one record per non-empty line). `sentence` / `char-chunk` reserved. Default: `full`.

Format	Detection / notes
JSONL	One JSON object per line. Text from `--text-field` (default: `text`). If the input object contains the id key named by `--id-field` (default: `id`), copy that field through to the output object.
JSON	A single object using the same field convention. For many records, use JSONL or preprocess (for example with `jq`).
Plain text	Controlled by `--text-split`: `full` (default for `.txt`) — entire file is one record; `line` — each non-empty line is one record (multiple lines ⇒ batching). `sentence` and `char-chunk` are reserved for a future release (segmentation semantics TBD).

Output: JSONL to stdout by default. --output <PATH> / -o <PATH> (use - for stdout). Optional --pretty: pretty-printed JSON when the implementation can buffer a single record or full result (for example one JSON object input or explicit single-line mode).

Format inference: From INPUT’s path suffix when possible: .jsonl → JSONL, .json → single JSON object, .txt (or other) → plain text with --text-split. For stdin (-), default input format is JSONL (one object per line).

Output record shape

Each output line is one JSON object, for example:

{"id":"optional","text":"...","result":{ }}

result matches Python / Rust **format_results** output for the task mix (entities, relation_extraction, classification keys, structured parents, etc.), consistent with the harness direction in harness/compare.py and multitask fixtures. If the input record has no id, omit id from the output (or use null; implementations should pick one behavior and document it).

Subcommands

`gliner2 entities`

Flag	Description
`--label <NAME>`	Repeatable entity type name.
`--labels-json <PATH>`	JSON array of names or object form accepted by `Schema::entities` (name → description string or `{ "description", "dtype", "threshold" }`).

Precedence: If any --label is given and --labels-json is given, exit with a usage error (do not merge).

`gliner2 classify`

Flag	Description
`--task <NAME>`	Required classification task name (JSON key in formatted output).
`--label <NAME>`	Repeatable class label.
`--labels-json <PATH>`	Array of labels or object label → description (Python-style).
`--multi-label`	Multi-label classification (`Schema::classification` with `multi_label: true`).
`--cls-threshold <float>`	Per-task classifier threshold (default `0.5`).

Same rule: do not combine --label with --labels-json.

`gliner2 relations`

Flag	Description
`--relation <NAME>`	Repeatable relation type name.
`--relations-json <PATH>`	JSON array of names or object form accepted by `Schema::relations`.

Do not pass both repeatable --relation and --relations-json.

`gliner2 json`

Flag	Description
`--structures <PATH>`	JSON file: object mapping structure name → array of field specs.
`--structures-json '<OBJECT>'`	Same object inline.

Field specs use the same grammar as Structured JSON (extract_json) above: strings like name::dtype::[choices]::description or JSON objects parsed by [parse_field_spec](src/schema.rs). Do not pass both --structures and --structures-json.

`gliner2 run`

Flag	Description
`--schema-file <PATH>`	Required. Full engine multitask schema (same shape as Python `GLiNER2.extract(text, schema)`). See `[harness/fixtures_multitask.json](harness/fixtures_multitask.json)` for a minimal example: `entities`, `classifications`, `relations`, `json_structures`, optional `entity_descriptions` / `json_descriptions`.

Each entry in classifications should include "true_label": ["N/A"] when mirroring Python; the harness script [harness/run_multitask_python.py](harness/run_multitask_python.py) sets this if missing.

Environment

**HF_TOKEN** — access to private or gated Hub models.
Cache and offline behavior follow Hugging Face Hub environment variables (HF_HOME, etc.); see upstream docs for the full list.

Exit codes

0 — success.
Non-zero — usage errors, I/O failures, model load failures, or inference errors.

Examples

# Entities: JSONL in → JSONL out (multi-record; default --batch-size 8 unless overridden)
gliner2 entities --label company --label person --batch-size 16 docs.jsonl --output out.jsonl

# Classify with labels from a file (JSONL input)
gliner2 classify --task sentiment --labels-json labels.json tweets.jsonl

# Relations
gliner2 relations --relation works_for --relation located_in article.txt

# Structured JSON (structures file matches extract_json object shape)
gliner2 json --structures product_fields.json --text-split full product_blurb.txt

# Multitask: JSONL file, custom text field
gliner2 run --schema-file schema.json --text-field body --batch-size 4 docs.jsonl

Minimal multitask schema file (trimmed from fixtures):

{
  "json_structures": [],
  "entities": { "company": "", "product": "" },
  "relations": [],
  "classifications": [
    {
      "task": "sentiment",
      "labels": ["positive", "negative", "neutral"],
      "multi_label": false,
      "cls_threshold": 0.5,
      "true_label": ["N/A"]
    }
  ]
}

Python Interface (Not implemented yet)

A Python package that wraps this Rust implementation (gliner2_rs) is planned if we can get rust performance to be better than Python; it is not implemented yet (this section is a placeholder).

# use your package manager of choice
uv add gliner2_rs

from gliner2_rs import Gliner2

gliner2 = Gliner2.from_pretrained('fastino/gliner2-base-v1')

text = "Apple CEO Tim Cook announced iPhone 15 in Cupertino yesterday."
result = extractor.extract_entities(text, ["company", "person", "product", "location"])

print(result)
# {'entities': {'company': ['Apple'], 'person': ['Tim Cook'], 'product': ['iPhone 15'], 'location': ['Cupertino']}}

gliner2 0.1.0