lfm 0.1.1

Rust ONNX inference for LiquidAI LFM2.5-VL (vision-language) models — implements the engine-agnostic llmtask::Task contract via llguidance for schema-constrained sampling
Documentation 
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

[package]
name          = "lfm"
version       = "0.1.1"
edition       = "2024"
rust-version  = "1.95"
description   = "Rust ONNX inference for LiquidAI LFM2.5-VL (vision-language) models — implements the engine-agnostic llmtask::Task contract via llguidance for schema-constrained sampling"
license       = "MIT OR Apache-2.0"
repository    = "https://github.com/findit-ai/lfm"
homepage      = "https://github.com/findit-ai/lfm"
documentation = "https://docs.rs/lfm"
keywords      = ["vlm", "onnx", "vision-language", "lfm", "constrained-decoding"]
categories    = ["api-bindings", "computer-vision", "multimedia::images"]
include      = [
  "src/**/*.rs",
  "examples/**/*.rs",
  "benches/**/*.rs",
  "models/*.json",
  "models/*.jinja",
  "build.rs",
  "Cargo.toml",
  "README.md",
  "CHANGELOG.md",
  "LICENSE-*",
]

[dependencies]
llmtask     = { version = "0.1", features = ["json", "regex"] }
ort         = { version = "2.0.0-rc.12", optional = true }
# `default-features = false` drops `progressbar` (training UI),
# `onig` (C++ regex engine) and `esaxx_fast` (C++ suffix-array
# Unigram trainer). lfm only uses `tokenizer.from_file` +
# encode/decode at inference time; the trainer paths are dead
# code for us. `fancy-regex` is the pure-Rust regex backend
# replacement for `onig` and is what JSON-defined BPE tokenizers
# need at runtime. Keeping `esaxx_fast` ON pulled the C++
# `esaxx-rs` into the link, which on Windows MSVC was built with
# `/MT` (static CRT) and clashed with `ort_sys` built with
# `/MD` (dynamic CRT) — `LNK1319: 1 mismatches detected`. This
# also matches the feature set `toktrie_hf_tokenizers` already
# uses on `tokenizers` 0.21 transitively.
tokenizers  = { version = "0.23", optional = true, default-features = false, features = ["fancy-regex"] }
llguidance  = { version = "1.7", optional = true }
toktrie     = { version = "1.7", optional = true }
toktrie_hf_tokenizers = { version = "1.7", optional = true }
minijinja   = { version = "2", optional = true, default-features = false, features = ["builtins", "json", "macros", "serde"] }
image       = { version = "0.25", default-features = false }
smol_str    = "0.3"
thiserror   = "2"
tracing     = "0.1"
serde       = { version = "1", features = ["derive"] }
serde_json  = { version = "1" }
fast_image_resize = "6.0.0"

[dev-dependencies]
serde_json = "1"

[target.'cfg(not(target_arch = "wasm32"))'.dev-dependencies]
criterion = "0.8"

[features]
default     = ["inference", "bundled", "decoders"]
# `inference` provides ORT, tokenizers, llguidance, minijinja. The Engine
# additionally requires `decoders` for image preprocessing — see
# src/lib.rs Engine gate. Use `inference` alone for the runtime types
# (Sampler, Decoder, KvCache, etc.) without the public Engine.
inference   = ["dep:ort", "dep:tokenizers", "dep:llguidance", "dep:toktrie", "dep:toktrie_hf_tokenizers", "dep:minijinja"]
# `bundled` ships the tokenizer + small JSON configs as include_bytes!.
# It implies both `inference` (for Engine machinery) AND `decoders`
# (for the `Engine::from_onnx_dir` constructor); without this, a
# `--features bundled` build would fail to expose Engine.
bundled     = ["inference", "decoders"]
decoders    = ["image/jpeg", "image/png"]
serde       = ["smol_str/serde", "llmtask/serde"]
cuda        = ["inference", "ort/cuda"]
tensorrt    = ["inference", "ort/tensorrt"]
directml    = ["inference", "ort/directml"]
rocm        = ["inference", "ort/rocm"]
coreml      = ["inference", "ort/coreml"]
integration = ["inference"]

[[test]]
name              = "integration"
path              = "tests/integration.rs"
required-features = ["integration"]

[[example]]
name              = "smoke"
# Engine + from_dir require all three: bundled (for tokenizer assets +
# from_dir gate), inference (for the runtime), decoders (for image
# decode). Round-27 fix: was incorrectly listed as `inference`-only.
required-features = ["bundled", "inference", "decoders"]
[[example]]
name              = "scene_analysis"
required-features = ["bundled", "inference", "decoders"]
[[example]]
name              = "preprocess_only"

[[bench]]
name    = "bench_preproc"
harness = false
[[bench]]
name    = "bench_tile_grid"
harness = false
[[bench]]
name              = "bench_chat_template"
harness           = false
required-features = ["inference"]

[profile.bench]
opt-level         = 3
debug             = false
codegen-units     = 1
lto               = 'thin'
incremental       = false
debug-assertions  = false
overflow-checks   = false
rpath             = false

[package.metadata.docs.rs]
features     = ["inference", "bundled", "decoders", "serde"]
rustdoc-args = ["--cfg", "docsrs"]

[lints.rust]
rust_2018_idioms     = "warn"
single_use_lifetimes = "warn"
unexpected_cfgs      = { level = "warn", check-cfg = ['cfg(docsrs)', 'cfg(tarpaulin)'] }