fix(serve): hide incomplete GGUF shard sets from model list

src/serve.rs

@@ -78,25 +78,34 @@ pub fn list_models(dir: &Path) -> Vec<(String, PathBuf, u64)> {
     out
 }
 
-/// The models among files given as (name, bytes): .gguf files, a split model once by its first shard with the
-/// bytes of all its shards, vision projectors (`mmproj*`) left out.
+/// The models among files given as (name, bytes): .gguf files, complete split models once by their first shard
+/// with the bytes of their shards, vision projectors (`mmproj*`) left out.
 pub fn models(files: &[(String, u64)]) -> Vec<(String, u64)> {
-    let parts = |stem: &str| files.iter().filter(|(m, _)| shard(m).is_some_and(|(s, _)| s == stem)).map(|x| x.1).sum();
+    let sizes: std::collections::HashMap<&str, u64> = files.iter().map(|(name, bytes)| (name.as_str(), *bytes)).collect();
     files.iter().filter(|(n, _)| n.ends_with(".gguf") && !n.rsplit('/').next().unwrap_or(n).starts_with("mmproj"))
-        .filter_map(|(n, b)| match shard(n) {
-            Some((stem, 1)) => Some((n.clone(), parts(stem))),
+        .filter_map(|(n, b)| match shard_parts(n) {
+            Some((stem, 1, total)) if total > 0 && total as usize <= files.len() => {
+                (1..=total).try_fold(0u64, |bytes, i| {
+                    sizes.get(format!("{stem}-{i:05}-of-{total:05}.gguf").as_str())
+                        .and_then(|part| bytes.checked_add(*part))
+                }).map(|bytes| (n.clone(), bytes))
+            }
             Some(_) => None,
             None => Some((n.clone(), *b)),
         }).collect()
 }
 
 /// "dir/name-00002-of-00003.gguf" gives ("dir/name", 2).
 pub fn shard(n: &str) -> Option<(&str, u32)> {
+    shard_parts(n).map(|(stem, i, _)| (stem, i))
+}
+
+fn shard_parts(n: &str) -> Option<(&str, u32, u32)> {
     let (head, total) = n.strip_suffix(".gguf")?.rsplit_once("-of-")?;
     let (stem, i) = head.rsplit_once('-')?;
     let digits = |s: &str| s.len() == 5 && s.bytes().all(|b| b.is_ascii_digit());
     if !digits(total) || !digits(i) { return None; }
-    Some((stem, i.parse().ok()?))
+    Some((stem, i.parse().ok()?, total.parse().ok()?))
 }
 
 /// The parts of an OpenAI chat completion request that strata uses. Sampling fields left out fall back to the

tests/serve.rs

@@ -1,5 +1,5 @@
 use rusting_strata::chat::{Hold, Json, Msg, Template};
-use rusting_strata::serve::{chat_request, host_ok, key_ok, list_models, messages_body, read_request};
+use rusting_strata::serve::{Keys, Who, chat_request, host_ok, key_ok, list_models, messages_body, models, read_request};
 
 /// Every request in tests/fixtures/chat renders byte-identical to llama.cpp's Jinja rendering of the model's own
 /// template (tools/gen_chat_fixtures.py); a request llama.cpp rejects must be rejected too.
@@ -234,6 +234,26 @@ fn lists_models_once_per_split() {
     assert_eq!(l, [("a.gguf".into(), "a.gguf".into(), 3), ("org/repo/big-00001-of-00002.gguf".into(), "org/repo/big-00001-of-00002.gguf".into(), 12)]);
 }
 
+#[test]
+fn lists_only_complete_split_models() {
+    let files = vec![
+        ("model-00001-of-00003.gguf".into(), 10),
+        ("model-00002-of-00003.gguf".into(), 20),
+        ("model-00001-of-00002.gguf".into(), 100),
+        ("model-00002-of-00002.gguf".into(), 20),
+        ("plain.gguf".into(), 7),
+    ];
+    assert_eq!(models(&files), vec![("model-00001-of-00002.gguf".into(), 120), ("plain.gguf".into(), 7)]);
+
+    let mut complete = files;
+    complete.push(("model-00003-of-00003.gguf".into(), 30));
+    assert_eq!(models(&complete), vec![
+        ("model-00001-of-00003.gguf".into(), 60),
+        ("model-00001-of-00002.gguf".into(), 120),
+        ("plain.gguf".into(), 7),
+    ]);
+}
+
 #[test]
 fn download_names_stay_in_the_models_folder() {
     use rusting_strata::hf::local_name;