kernel/resolution/
format.rs

1//! Format and capability facts derived from a model's files.
2
3use serde::{Deserialize, Serialize};
4
5use crate::records::{Capability, ExecutionMode, Modality};
6
7/// A recognized model format — the on-disk weight formats plus the logical
8/// sources (a diffusers pipeline directory, an Ollama store entry, a built-in or
9/// remote-endpoint model), and `Unknown` for anything unrecognized.
10#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize)]
11#[serde(rename_all = "kebab-case")]
12pub enum ModelFormat {
13    /// A GGUF weight file (llama.cpp).
14    Gguf,
15    /// A legacy GGML `.bin` weight file.
16    GgmlBin,
17    /// A safetensors weight directory.
18    Safetensors,
19    /// An MLX-format safetensors weight directory.
20    MlxSafetensors,
21    /// A diffusers pipeline directory (a `model_index.json`).
22    Diffusers,
23    /// An entry in a local Ollama store.
24    OllamaStore,
25    /// A built-in (platform-provided) model.
26    Builtin,
27    /// A remote inference endpoint.
28    Endpoint,
29    /// An unrecognized format.
30    Unknown,
31}
32
33/// The modality, capabilities, and execution shape implied by a GGUF
34/// architecture.
35#[derive(Debug, Clone, PartialEq)]
36pub struct GgufArchitectureProfile {
37    /// The model's primary modality.
38    pub modality: Modality,
39    /// What the model can do.
40    pub capabilities: Vec<Capability>,
41    /// How its runtime delivers output.
42    pub execution: ExecutionMode,
43}
44
45/// Facts read from a GGUF header.
46#[derive(Debug, Clone, PartialEq)]
47pub struct GgufFacts {
48    /// The `general.architecture` value, if present.
49    pub architecture: Option<String>,
50    /// The resolved context length, if the header declared one.
51    pub context_length: Option<i64>,
52    /// Whether the header carries a chat template.
53    pub has_chat_template: bool,
54}
55
56/// The known-architecture profile for a GGUF `general.architecture` value.
57///
58/// The table names the architectures that are *not* a text chat model, which is
59/// what an architecture missing from it is taken to be: the great majority of
60/// them are, and a name nobody has taught this table is far more likely to be
61/// another text model than a component. What is listed here is the exceptions:
62/// the encoders that only embed, the models that can also see, and the pieces
63/// that are half of a pipeline and serve nothing on their own.
64pub fn gguf_architecture_profile(architecture: &str) -> Option<GgufArchitectureProfile> {
65    let profile = |modality, capabilities, execution| GgufArchitectureProfile {
66        modality,
67        capabilities,
68        execution,
69    };
70    let embedding = || {
71        Some(profile(
72            Modality::embedding(),
73            vec![Capability::embed()],
74            ExecutionMode::Stream,
75        ))
76    };
77    let sees = || {
78        Some(profile(
79            Modality::text(),
80            vec![
81                Capability::chat(),
82                Capability::complete(),
83                Capability::see(),
84            ],
85            ExecutionMode::Stream,
86        ))
87    };
88    // A piece of a pipeline: it has a modality but nothing can be asked of it
89    // directly, so no runtime offers to serve it and it is never mistaken for a
90    // model that answers.
91    let component = |modality| Some(profile(modality, vec![], ExecutionMode::Sync));
92    match architecture {
93        "whisper" => Some(profile(
94            Modality::audio(),
95            vec![Capability::transcribe()],
96            ExecutionMode::Stream,
97        )),
98        "qwen2vl" | "qwen3vl" | "qwen3vlmoe" | "mllama" | "cogvlm" | "hunyuan-vl" => sees(),
99        "bert" | "nomic-bert" | "nomic-bert-moe" | "jina-bert-v2" | "jina-bert-v3" | "neo-bert"
100        | "modern-bert" | "eurobert" | "gemma-embedding" | "llama-embed" | "t5encoder" => {
101            embedding()
102        }
103        "clip" => component(Modality::vision()),
104        // The vocoder half of a text-to-speech pair: it turns another model's
105        // tokens into audio and cannot be prompted.
106        "wavtokenizer-dec" => component(Modality::audio()),
107        _ => None,
108    }
109}
110
111/// The default profile for an Ollama chat model with no more specific match.
112pub fn ollama_chat_profile() -> GgufArchitectureProfile {
113    GgufArchitectureProfile {
114        modality: Modality::text(),
115        capabilities: vec![Capability::chat(), Capability::complete()],
116        execution: ExecutionMode::Stream,
117    }
118}
119
120/// The default profile for an Ollama vision-capable chat model.
121pub fn ollama_vision_profile() -> GgufArchitectureProfile {
122    GgufArchitectureProfile {
123        modality: Modality::text(),
124        capabilities: vec![
125            Capability::chat(),
126            Capability::complete(),
127            Capability::see(),
128        ],
129        execution: ExecutionMode::Stream,
130    }
131}
132
133/// The profile for an Ollama model: vision when it ships a projector, otherwise
134/// the GGUF architecture's profile (read from `weight_path`) if recognized, else
135/// the plain chat default.
136///
137/// `weight_path` must be absolute — a leading `~` is not expanded (the Ollama
138/// scanner always passes a resolved blob path;
139/// a `~`-relative path would fail to open and fall through to the chat default).
140pub fn ollama_profile(has_projector: bool, weight_path: Option<&str>) -> GgufArchitectureProfile {
141    if has_projector {
142        return ollama_vision_profile();
143    }
144    if let Some(path) = weight_path
145        && let Some(architecture) =
146            crate::resolution::gguf::gguf_general_architecture(std::path::Path::new(path))
147        && let Some(profile) = gguf_architecture_profile(&architecture)
148    {
149        return profile;
150    }
151    ollama_chat_profile()
152}
kernel/resolution/format.rs

kernel/resolution/
format.rs
