feat(neuron,cortex-core): source-aware loader (scheme:org/name)
All checks were successful
CI / CUDA type-check (push) Successful in 46s
CI / Format (push) Successful in 32s
build-prerelease / Resolve version stamps (push) Successful in 42s
CI / Clippy (push) Successful in 2m40s
build-prerelease / Build cortex binary (push) Successful in 4m23s
CI / Test (push) Successful in 5m28s
CI / Build cortex SRPM (push) Has been skipped
CI / Build neuron SRPM (push) Has been skipped
CI / Publish cortex to COPR (push) Has been skipped
CI / Publish neuron to COPR (push) Has been skipped
CI / Bump version in source (push) Has been skipped
build-prerelease / Build neuron-blackwell (push) Successful in 5m39s
build-prerelease / Package cortex RPM (push) Successful in 1m19s
build-prerelease / Build neuron-ampere (push) Successful in 7m53s
build-prerelease / Build neuron-ada (push) Successful in 5m18s
build-prerelease / Package helexa-neuron-ampere RPM (push) Successful in 2m59s
build-prerelease / Package helexa-neuron-ada RPM (push) Successful in 3m6s
build-prerelease / Package helexa-neuron-blackwell RPM (push) Successful in 3m44s
build-prerelease / Publish to rpm.lair.cafe (unstable) (push) Successful in 1m2s

Phase 1 of plan-source-aware-loader-preflight. Makes neuron's
loader treat `huggingface:org/name` and `helexa:org/name` as
first-class distinct sources with per-source endpoint + cache,
while staying backwards-compatible with bare `org/name` ids.
Zero behavior change for existing operator configs.

Motivation: helexa is adding an EU-hosted registry
(`registry.helexa.ai`) alongside HF. Both speak HF-compatible
wire format, but the bytes, jurisdiction, trust root, and cache
namespace are distinct. The loader needs to disambiguate which
registry serves a given model id, and to keep their caches from
colliding on disk when both happen to host the same `org/name`.

What lands:

- `cortex-core::source` — new module. `ModelSourceId { scheme,
  org, name }` with `FromStr` accepting both `scheme:org/name`
  and bare `org/name`. `Display` round-trips. `repo_path()`
  emits the `org/name` half for the hf-hub `Api::model(...)`
  call regardless of which scheme/endpoint we're hitting.
  Rejects malformed input with typed `ParseError` variants
  (empty scheme, missing slash, scheme with `/`, name with
  `:`, etc.).

- `neuron::config::CandleHarnessConfig` gains
  `default_source: Option<String>` and
  `sources: HashMap<String, SourceConfig>`. `SourceConfig`
  mirrors what `hf_hub::ApiBuilder` consumes: endpoint URL,
  optional `auth_env` (env var name read at startup so secrets
  stay out of TOML), and optional cache_dir. Defaults
  synthesise a `huggingface` entry pointing at
  `https://huggingface.co` with the legacy `hf_cache` field as
  its cache_dir — so existing configs that only set `hf_cache`
  keep working unchanged.

- `CandleHarness::new(bind_url, &CandleHarnessConfig)` replaces
  `CandleHarness::new(bind_url, hf_cache)`. Resolves every
  configured source's auth env var and cache dir up front so
  `hf_api_for(scheme)` is a pure HashMap lookup on the hot
  load path. Only the `huggingface` scheme gets the legacy
  `HF_HUB_CACHE`/`HF_HOME` env-var fallback chain; other
  schemes resolve to whatever the operator typed.

- `hf_api()` -> `hf_api_for(scheme)`. Builds an
  `hf_hub::Api` with the source's endpoint, cache_dir, and
  auth token. Errors with a useful message naming the
  configured schemes when an unknown scheme is requested.

- `CandleHarness::load_model` parses `spec.model_id` into a
  `ModelSourceId`, substitutes `default_source` for bare ids,
  and threads the parsed source through `preflight`,
  `resolve_files`, `resolve_dense_files`, `load_arch_gguf`,
  `load_arch_dense`, and `load_tp`. The hf-hub `Api::model()`
  call now uses `source_id.repo_path()` so registry calls hit
  the right URL shape regardless of scheme.

- `preflight()` signature gains a `&ModelSourceId` parameter
  (it's the canonical id for log lines and error display);
  `RepoFetchFailed.model_id` etc. now carry the
  scheme-qualified form so operator-visible errors echo
  exactly what was configured.

- `neuron.example.toml` documents the new
  `[harness.candle.sources.*]` table with commented-out
  examples for `huggingface` (explicit override) and `helexa`.

Tests:

- 13 new unit tests in `cortex-core::source` covering parse /
  display round-trip, default-scheme substitution semantics,
  and every `ParseError` variant.
- 6 new unit tests in `neuron::config` covering the
  `effective_sources` synth (legacy `hf_cache` carry-through,
  explicit override preservation, helexa-alongside-huggingface)
  and `effective_default_source` fallback.
- 2 new unit tests in `harness::candle::tests` covering
  multi-scheme `hf_api_for` routing, including the
  "unknown scheme" error path naming configured schemes.
- Preflight integration tests updated to construct
  `ModelSourceId` and assert against the scheme-qualified
  error form.

CI gate: cargo fmt --check, cargo clippy --workspace
--all-targets -- -D warnings, cargo test --workspace (all 24
test groups ok, zero failures).

Out of scope (Phase 3):
- Cortex catalogue `source` field — independent of Phase 1+2,
  ships when the registry comes online.
- `helexa` source endpoint itself — separate project; this
  PR adds the client-side rails only.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
2026-06-01 13:42:11 +03:00
parent 61adff347a
commit d4e1b05956
8 changed files with 780 additions and 67 deletions

View File

@@ -7,4 +7,5 @@ pub mod metrics;
pub mod node;
pub mod openai;
pub mod responses;
pub mod source;
pub mod translate;

View File

@@ -0,0 +1,267 @@
//! Scheme-qualified model identifiers.
//!
//! cortex/neuron historically resolves every model id through hf-hub
//! against `https://huggingface.co`. Helexa is adding an EU-hosted
//! registry (`registry.helexa.ai`) alongside HF — both speak the same
//! HF-compatible wire format, but the bytes, jurisdiction, and trust
//! root differ. Model ids therefore need a scheme:
//!
//! - `huggingface:Qwen/Qwen3.6-27B` — HF-hosted bytes
//! - `helexa:Qwen/Qwen3.6-27B-Uncensored` — helexa registry bytes
//! - `helexa:SomeOperator/CustomFinetune` — operator publishing
//! under the helexa namespace; same scheme handles all `org/name`
//! pairs hosted in that registry.
//!
//! Bare `org/name` parses with an empty scheme; the caller (typically
//! a harness) substitutes its configured default scheme so existing
//! configs keep working through the transition.
use serde::{Deserialize, Serialize};
use std::fmt;
use std::str::FromStr;
/// Parsed `scheme:org/name`. Bare `org/name` produces an empty scheme
/// — call `with_default_scheme` (or check `is_scheme_unset`) to
/// resolve before using.
#[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize)]
pub struct ModelSourceId {
pub scheme: String,
pub org: String,
pub name: String,
}
/// Errors from `ModelSourceId::from_str`. Carries the offending input
/// so log lines / API errors can echo what the operator typed.
#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
pub enum ParseError {
#[error("empty model id")]
Empty,
#[error("model id '{0}' is missing the '/' between org and name")]
MissingSlash(String),
#[error("model id '{0}' has an empty scheme before ':'")]
EmptyScheme(String),
#[error("model id '{0}' has an empty org")]
EmptyOrg(String),
#[error("model id '{0}' has an empty name")]
EmptyName(String),
#[error("model id '{0}' has a scheme containing '/' which is reserved for org/name")]
SchemeContainsSlash(String),
#[error("model id '{0}' has a name containing ':' which is reserved for the scheme prefix")]
NameContainsColon(String),
}
impl ModelSourceId {
/// Construct directly from already-validated parts. Used by tests
/// and call sites that have the fields separately; the public API
/// for parsing user input is `FromStr`.
pub fn new(scheme: impl Into<String>, org: impl Into<String>, name: impl Into<String>) -> Self {
Self {
scheme: scheme.into(),
org: org.into(),
name: name.into(),
}
}
/// True when this id parsed from a bare `org/name` (no scheme
/// prefix). The harness substitutes its configured default in
/// `with_default_scheme` before resolving against a registry.
pub fn is_scheme_unset(&self) -> bool {
self.scheme.is_empty()
}
/// Substitute `default` for an empty scheme. No-op when the scheme
/// is already set. Returns self by value so it composes neatly:
/// `id.parse::<ModelSourceId>()?.with_default_scheme("huggingface")`.
pub fn with_default_scheme(mut self, default: &str) -> Self {
if self.scheme.is_empty() {
self.scheme = default.to_string();
}
self
}
/// The `org/name` half — what an hf-hub `Api::model(...)` call
/// expects regardless of which scheme/endpoint we're hitting.
pub fn repo_path(&self) -> String {
format!("{}/{}", self.org, self.name)
}
}
impl fmt::Display for ModelSourceId {
fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
if self.scheme.is_empty() {
write!(f, "{}/{}", self.org, self.name)
} else {
write!(f, "{}:{}/{}", self.scheme, self.org, self.name)
}
}
}
impl FromStr for ModelSourceId {
type Err = ParseError;
fn from_str(s: &str) -> Result<Self, Self::Err> {
if s.is_empty() {
return Err(ParseError::Empty);
}
// Scheme split. Only the *first* colon counts — anything after
// belongs to org/name (and would be rejected separately because
// `:` isn't allowed there).
let (scheme, rest) = match s.split_once(':') {
Some((scheme, rest)) => {
if scheme.is_empty() {
return Err(ParseError::EmptyScheme(s.to_string()));
}
if scheme.contains('/') {
return Err(ParseError::SchemeContainsSlash(s.to_string()));
}
(scheme.to_string(), rest)
}
None => (String::new(), s),
};
let (org, name) = rest
.split_once('/')
.ok_or_else(|| ParseError::MissingSlash(s.to_string()))?;
if org.is_empty() {
return Err(ParseError::EmptyOrg(s.to_string()));
}
if name.is_empty() {
return Err(ParseError::EmptyName(s.to_string()));
}
if name.contains(':') {
return Err(ParseError::NameContainsColon(s.to_string()));
}
Ok(Self {
scheme,
org: org.to_string(),
name: name.to_string(),
})
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn parses_qualified() {
let id: ModelSourceId = "huggingface:Qwen/Qwen3.6-27B".parse().unwrap();
assert_eq!(id.scheme, "huggingface");
assert_eq!(id.org, "Qwen");
assert_eq!(id.name, "Qwen3.6-27B");
assert_eq!(id.repo_path(), "Qwen/Qwen3.6-27B");
assert!(!id.is_scheme_unset());
}
#[test]
fn parses_helexa_scheme() {
let id: ModelSourceId = "helexa:SomeOperator/Qwen3.6-27B-Uncensored"
.parse()
.unwrap();
assert_eq!(id.scheme, "helexa");
assert_eq!(id.org, "SomeOperator");
assert_eq!(id.name, "Qwen3.6-27B-Uncensored");
}
#[test]
fn parses_bare_id_with_empty_scheme() {
let id: ModelSourceId = "Qwen/Qwen3-30B-A3B-Instruct".parse().unwrap();
assert_eq!(id.scheme, "");
assert_eq!(id.org, "Qwen");
assert_eq!(id.name, "Qwen3-30B-A3B-Instruct");
assert!(id.is_scheme_unset());
}
#[test]
fn substitutes_default_scheme_only_when_unset() {
let id: ModelSourceId = "Qwen/Q3".parse().unwrap();
assert_eq!(id.with_default_scheme("huggingface").scheme, "huggingface");
let id: ModelSourceId = "helexa:Qwen/Q3".parse().unwrap();
assert_eq!(
id.with_default_scheme("huggingface").scheme,
"helexa",
"default substitution must not override an explicit scheme"
);
}
#[test]
fn display_roundtrips_qualified_id() {
let s = "helexa:Helexa/Qwen3.6-27B";
let id: ModelSourceId = s.parse().unwrap();
assert_eq!(id.to_string(), s);
}
#[test]
fn display_roundtrips_bare_id() {
let s = "Qwen/Q3";
let id: ModelSourceId = s.parse().unwrap();
assert_eq!(id.to_string(), s);
}
#[test]
fn rejects_empty() {
assert_eq!("".parse::<ModelSourceId>().unwrap_err(), ParseError::Empty);
}
#[test]
fn rejects_missing_slash() {
match "Qwen".parse::<ModelSourceId>().unwrap_err() {
ParseError::MissingSlash(s) => assert_eq!(s, "Qwen"),
other => panic!("expected MissingSlash, got {other:?}"),
}
match "huggingface:Qwen".parse::<ModelSourceId>().unwrap_err() {
ParseError::MissingSlash(s) => assert_eq!(s, "huggingface:Qwen"),
other => panic!("expected MissingSlash, got {other:?}"),
}
}
#[test]
fn rejects_empty_scheme() {
match ":Qwen/Q3".parse::<ModelSourceId>().unwrap_err() {
ParseError::EmptyScheme(s) => assert_eq!(s, ":Qwen/Q3"),
other => panic!("expected EmptyScheme, got {other:?}"),
}
}
#[test]
fn rejects_scheme_with_slash() {
match "hugg/ingface:Q/N".parse::<ModelSourceId>().unwrap_err() {
ParseError::SchemeContainsSlash(s) => assert_eq!(s, "hugg/ingface:Q/N"),
other => panic!("expected SchemeContainsSlash, got {other:?}"),
}
}
#[test]
fn rejects_empty_org_or_name() {
match "huggingface:/N".parse::<ModelSourceId>().unwrap_err() {
ParseError::EmptyOrg(_) => {}
other => panic!("expected EmptyOrg, got {other:?}"),
}
match "huggingface:Q/".parse::<ModelSourceId>().unwrap_err() {
ParseError::EmptyName(_) => {}
other => panic!("expected EmptyName, got {other:?}"),
}
}
#[test]
fn rejects_name_with_colon() {
match "huggingface:Q/N:weird"
.parse::<ModelSourceId>()
.unwrap_err()
{
ParseError::NameContainsColon(s) => assert_eq!(s, "huggingface:Q/N:weird"),
other => panic!("expected NameContainsColon, got {other:?}"),
}
}
#[test]
fn serde_roundtrips_via_struct() {
// We serialize as a struct (scheme/org/name fields) so the
// shape is self-describing in API payloads. Callers that want
// the compact `scheme:org/name` string use `Display`/`FromStr`.
let id = ModelSourceId::new("helexa", "Helexa", "Qwen3.6-27B");
let json = serde_json::to_string(&id).unwrap();
let back: ModelSourceId = serde_json::from_str(&json).unwrap();
assert_eq!(back, id);
}
}