inferd-engine 0.2.1

Backend trait and adapters for inferd. v0.1: mock + libllama (FFI). v0.2: cloud adapters.
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

[package]

name = "inferd-engine"

description = "Backend trait and adapters for inferd. v0.1: mock + libllama (FFI). v0.2: cloud adapters."

version.workspace      = true

edition.workspace      = true

rust-version.workspace = true

license.workspace      = true

repository.workspace   = true

homepage.workspace     = true

authors.workspace      = true

build = "build.rs"



[features]

# Default ships the mock backend only. The llama.cpp FFI backend is

# opt-in until M2a vendors the submodule and the build.rs lands; once

# stable, this becomes default for the daemon binary but stays

# feature-gated for crate consumers who only want the trait.

default              = []

llamacpp             = []

# Tier 3 integration tests in tests/llamacpp.rs only compile when this

# feature is on. Set INFERD_TEST_MODEL_PATH to a GGUF file to actually run

# them; otherwise they skip with an explanatory message.

llamacpp-integration = ["llamacpp"]

cuda                 = ["llamacpp"]

metal                = ["llamacpp"]

vulkan               = ["llamacpp"]

rocm                 = ["llamacpp"]

# OpenAI-compat outbound HTTP backend (Phase 5A). Pulls in reqwest +

# eventsource_stream. The daemon never *serves* HTTP — this is the

# narrow outbound carve-out wrapped behind the Backend trait

# (ADR 0006 §"cloud backends").

openai               = ["dep:reqwest", "dep:eventsource-stream", "dep:futures-util"]

# AWS Bedrock-runtime InvokeModelWithResponseStream backend (Phase

# 6B-5). Anthropic-on-Bedrock body shape only in v0.2.0; SigV4 +

# AWS_BEARER_TOKEN_BEDROCK auth; AWS event-stream framing parser.

# Same lean-core posture as `openai`: outbound HTTPS only, no extra

# inbound surface in the daemon.

bedrock              = ["dep:reqwest", "dep:futures-util", "dep:hmac", "dep:hex", "dep:url"]



[dependencies]

inferd-proto = { path = "../inferd-proto", version = "=0.2.1" }

async-trait  = { workspace = true }

tokio        = { workspace = true }

tokio-stream = "0.1"

futures-core = "0.3"

serde        = { workspace = true }

thiserror    = { workspace = true }

tracing      = { workspace = true }

sha2         = { workspace = true }

subtle       = { workspace = true }   # Constant-time hash compare (THREAT_MODEL F-5).

bytes        = { workspace = true }

# v2 multimodal: base64-decode the wire-shape RGB / PCM bytes before

# handing them to libmtmd. Pulled in unconditionally; tiny crate

# (no_std-compatible, ~10 KB binary growth).

base64       = "0.22"

# Used by the chat_template module to render tool-use input objects

# and tool-result content into Gemma 4's wire format. Already pulled

# in transitively via `inferd-proto`'s v2 ToolUseInput type alias.

serde_json   = { workspace = true }



# --- Optional: OpenAI-compat HTTP backend (Phase 5A). Pulled in only

# when the `openai` feature is on. rustls-only to avoid OpenSSL on

# Windows.

reqwest             = { version = "0.12", default-features = false, features = ["rustls-tls", "json", "stream"], optional = true }

eventsource-stream  = { version = "0.2", optional = true }

futures-util        = { version = "0.3", optional = true }



# --- Optional: Bedrock-invoke backend (Phase 6B-5). SigV4 needs HMAC-

# SHA-256 (hmac+sha2 — sha2 is already a workspace dep) and lowercase

# hex for the canonical request hash. `url` parses the configured

# endpoint into host/path/query for the canonical request shape.

hmac                = { version = "0.12", optional = true }

hex                 = { version = "0.4", optional = true }

url                 = { version = "2", optional = true }



[build-dependencies]

# Only invoked when the `llamacpp` feature is enabled; the build.rs

# checks env::var("CARGO_FEATURE_LLAMACPP") and short-circuits otherwise.

cmake   = "0.1"

bindgen = "0.71"



[dev-dependencies]

tokio    = { workspace = true, features = ["macros", "rt-multi-thread"] }

tempfile = "3"

# Tier 3 multimodal smoke tests act as a consumer: decode JPEG/PNG

# inputs to raw RGB before sending through the v2 wire shape (per

# ADR 0016, decoding lives consumer-side; the daemon never links a

# codec). image is dev-only.

image    = { version = "0.25", default-features = false, features = ["jpeg", "png"] }

base64   = "0.22"

# Mock HTTP server for the openai-compat adapter unit tests.

wiremock = "0.6"