hugo/DS4Server
1
0
Fork 0
Code Issues 2 Pull Requests Actions Packages Projects Releases Wiki Activity

Keep runtime files inside application support

Browse Source
This commit is contained in:
Georg Bauer
2026-07-28 18:34:08 +02:00
parent 31f6426eef
commit 9cf1c70a67
8 changed files with 141 additions and 108 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
src/engine/metal/glm.rs
View File

@@ -3221,7 +3221,11 @@ mod tests {
use crate::engine::{GLM, Model, ReasoningMode};
use crate::model::ModelChoice;
use crate::settings::EngineSsdSettings;
use std::path::Path;
fn installed_glm_path() -> std::path::PathBuf {
crate::model::engine_artifacts(ModelChoice::Glm52, false, false, &crate::app::models_path())
.model
}
#[test]
fn dsa_indexer_schedule_matches_the_reference() {
@@ -3258,16 +3262,14 @@ mod tests {
#[test]
#[ignore = "requires the 197 GiB GLM 5.2 checkpoint and Apple Metal"]
fn resident_and_streamed_glm_match_ds4_decode_oracles() {
let path = std::env::var("DS4_GLM_MODEL").unwrap_or_else(|_| {
"../ds4/models/GLM-5.2-UD-IQ2_XXS_RoutedIQ2XXS_blk78Q2K.gguf".into()
});
if !Path::new(&path).is_file() {
eprintln!("skipping unavailable GLM fixture: {path}");
let path = installed_glm_path();
if !path.is_file() {
eprintln!("skipping unavailable GLM fixture: {}", path.display());
return;
}
let prompt = b"Complete the C statement with the next exact token only:\nreturn snprintf(buf, sizeof(buf), \"%d\", value";
for streamed in [false, true] {
let model = Model::open_main(Path::new(&path), ModelChoice::Glm52).unwrap();
let model = Model::open_main(&path, ModelChoice::Glm52).unwrap();
let tokens = model.tokenize(std::str::from_utf8(prompt).unwrap());
let executor = GlmExecutor::open(
model,
@@ -3329,15 +3331,13 @@ mod tests {
#[test]
#[ignore = "requires the 197 GiB GLM 5.2 checkpoint and Apple Metal"]
fn streamed_glm_uses_ds4_indexed_prefill_for_long_prompts() {
let path = std::env::var("DS4_GLM_MODEL").unwrap_or_else(|_| {
"../ds4/models/GLM-5.2-UD-IQ2_XXS_RoutedIQ2XXS_blk78Q2K.gguf".into()
});
if !Path::new(&path).is_file() {
eprintln!("skipping unavailable GLM fixture: {path}");
let path = installed_glm_path();
if !path.is_file() {
eprintln!("skipping unavailable GLM fixture: {}", path.display());
return;
}
let prompt = "Complete each C statement. Example: return snprintf(buf, sizeof(buf), \"%d\", value); Example: return snprintf(buf, sizeof(buf), \"%d\", value); Example: return snprintf(buf, sizeof(buf), \"%d\", value); Example: return snprintf(buf, sizeof(buf), \"%d\", value); Example: return snprintf(buf, sizeof(buf), \"%d\", value); Now complete exactly: return snprintf(buf, sizeof(buf), \"%d\", value";
let model = Model::open_main(Path::new(&path), ModelChoice::Glm52).unwrap();
let model = Model::open_main(&path, ModelChoice::Glm52).unwrap();
let tokens =
model.render_prompt("You are a helpful assistant", prompt, ReasoningMode::Direct);
assert_eq!(tokens.len(), 102);
@@ -3382,15 +3382,13 @@ mod tests {
use crate::settings::EngineSpeculativeSettings;
use std::sync::atomic::AtomicBool;
let path = std::env::var("DS4_GLM_MODEL").unwrap_or_else(|_| {
"../ds4/models/GLM-5.2-UD-IQ2_XXS_RoutedIQ2XXS_blk78Q2K.gguf".into()
});
if !Path::new(&path).is_file() {
eprintln!("skipping unavailable GLM fixture: {path}");
let path = installed_glm_path();
if !path.is_file() {
eprintln!("skipping unavailable GLM fixture: {}", path.display());
return;
}
let run = |enabled| {
let model = Model::open_main(Path::new(&path), ModelChoice::Glm52).unwrap();
let model = Model::open_main(&path, ModelChoice::Glm52).unwrap();
let tokens = model.tokenize("Write one short greeting.");
let mut executor = GlmExecutor::open_profile(
model,