frink-cli 0.49.0

llama.cpp-style CLI for the Frink inference engine
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

[package]
name = "frink-cli"
version.workspace = true
edition.workspace = true
license.workspace = true
repository.workspace = true
homepage.workspace = true
description = "llama.cpp-style CLI for the Frink inference engine"
readme = "README.md"
keywords = ["llm", "inference", "cli", "gguf", "frink"]
categories = ["command-line-utilities", "science"]

[[bin]]
name = "frink"
path = "src/main.rs"

[dependencies]
frink-gguf = { workspace = true }
frink-quant = { workspace = true }
frink-cuda = { workspace = true }
frink-metal = { workspace = true }
half = { workspace = true }
frink-core = { workspace = true }
frink-models = { workspace = true, features = ["hub"] }
frink-safetensors = { workspace = true }
# Optional, and off by default -- see the `serve` feature below.
frink-server = { workspace = true, optional = true }
clap = { workspace = true }
anyhow = { workspace = true }
serde_json = { workspace = true }
tracing = { workspace = true }
tracing-subscriber = { workspace = true }
rayon = { workspace = true }
byteorder = { workspace = true }
# For the SIGINT handler in `frink chat`: Ctrl-C stops the turn rather
# than the session. No new external dependency -- frink-core already
# uses it.
libc = { workspace = true }

[features]
default = ["serve"]
# Must forward to frink-models the same way `metal` does, or the CLI
# links a frink-core built *without* `cuda` -- no `cuda_dense_enabled`,
# no `apply_gpu` CUDA arm -- and `cargo build -p frink-cli --features
# cuda` does not compile at all. Only ever caught on a real CUDA host,
# because CI has none.
cuda = ["frink-cuda/cuda", "frink-models/cuda", "frink-server?/cuda"]
# Caps detection + Metal apply_gpu dispatch through frink-models.
metal = ["frink-metal/metal", "frink-models/metal", "frink-server?/metal"]
# `frink serve`: the OpenAI-compatible HTTP server, linked in from the
# frink-server library rather than reimplemented, so both front ends
# share one `ServerArgs` and one code path.
#
# ON BY DEFAULT. `cargo install frink-cli` gives you the whole tool:
# completions, `frink serve`, `frink download`, bench, verify. One
# binary that does everything is worth more than a small one that does
# part of it and makes you discover a feature flag to get the rest.
#
# It costs build time and crates (axum, tokio, hyper, rustls), and that
# is the trade. A library consumer who wants only the model code should
# depend on frink-models or frink-core directly rather than on the
# CLI, or pass `default-features = false`.
#
# The `?` in the cuda/metal features above still matters: they forward
# to frink-server only if it is in the graph, never enable it.
serve = ["dep:frink-server"]