merge: CR-4 — rename milady model-tooling scripts to eliza1, consolidate HF bundle naming, plugin/dir cleanup

Author: lalalune

.github/workflows/local-inference-bench.yml

@@ -32,7 +32,7 @@ on:
       - ".github/workflows/local-inference-bench.yml"
       - "plugins/plugin-local-embedding/**"
       - "packages/app-core/scripts/build-llama-cpp-dflash.mjs"
-      - "packages/training/scripts/publish_milady_model.py"
+      - "packages/training/scripts/publish_eliza1_model.py"
       - "packages/training/scripts/sync_catalog_from_hf.py"
       - "scripts/verify-phone-download.mjs"
   workflow_dispatch:
@@ -486,7 +486,7 @@ jobs:
 
   publish-models-nightly:
     # Weekly nightly publish to the elizaos HF org. Walks W5-Pipeline's
-    # output, calls publish_milady_model.py for each fused-kernel GGUF,
+    # output, calls publish_eliza1_model.py for each fused-kernel GGUF,
     # then runs sync_catalog_from_hf.py to produce the catalog diff and
     # opens a PR with the update.
     #
@@ -502,7 +502,7 @@ jobs:
     env:
       # W5-Pipeline drops finished checkpoints under this path. The
       # nightly publish step expects each subdirectory to be a complete
-      # `<base>-milady-optimized` (or `-milady-drafter`) bundle —
+      # `<base>-optimized` (or `-drafter`) bundle —
       # i.e. a single .gguf + manifest.json + README.md.
       W5_PIPELINE_OUTPUT_ROOT: ${{ vars.W5_PIPELINE_OUTPUT_ROOT || '/mnt/w5-pipeline-output' }}
       HF_TOKEN: ${{ secrets.MILADY_HF_TOKEN }}
@@ -568,15 +568,15 @@ jobs:
           FAILURES=0
           for dir in "$W5_PIPELINE_OUTPUT_ROOT"/*/; do
             name=$(basename "$dir")
-            # Bundle naming: <base>-milady-optimized or <base>-milady-drafter.
+            # Bundle naming: <base>-optimized or <base>-drafter.
             # Anything else is rejected — these repos are for fused builds only.
-            if [[ "$name" != *-milady-optimized && "$name" != *-milady-drafter ]]; then
+            if [[ "$name" != *-optimized && "$name" != *-drafter ]]; then
               echo "::warning ::skipping $name (not a fused-kernel bundle name)"
               continue
             fi
             repo_id="elizaos/$name"
             echo "::group::publish $repo_id"
-            if ! python packages/training/scripts/publish_milady_model.py \
+            if ! python packages/training/scripts/publish_eliza1_model.py \
                 --model-dir "$dir" \
                 --repo-id "$repo_id"; then
               echo "::error ::publish failed for $name"

docs/training/architecture.md

@@ -20,7 +20,7 @@ this file, this file is the intended source of truth.
 `packages/training/` is the only thing in the repo that actually produces the
 `eliza-1` weights that ship with Milady and get downloaded onto phones.
 
-- Entry point: `packages/training/scripts/optimize_for_milady.py` (full recipe
+- Entry point: `packages/training/scripts/optimize_for_eliza1.py` (full recipe
   in [`optimization-pipeline.md`](optimization-pipeline.md)).
 - Base model: a Qwen3 causal-LM. The **smallest target is `eliza-1-0_6b`,
   derived from `Qwen/Qwen3-0.6B`** (older docs called this `eliza-1-lite-0_6b`
@@ -31,7 +31,7 @@ this file, this file is the intended source of truth.
   spec-decode CLI surface, all of which require the
   `elizaOS/llama.cpp` v0.4.0-milady fork at runtime.
 - Publishes to `elizaos/eliza-1-<tier>` on HuggingFace, then
-  `emit_milady_catalog.py` wires the new repo into
+  `emit_eliza1_catalog.py` wires the new repo into
   `packages/app-core/src/services/local-inference/catalog.ts`.
 - It does **not** mount any HTTP routes. It is a CLI / cron pipeline.
 

packages/training/docs/training/gguf-to-runtime.md

@@ -4,8 +4,8 @@ This is the handoff between this package (the offline training /
 quantization / GGUF pipeline) and the Milady runtime
 (`packages/app-core`, `packages/shared`). It assumes you already have a
 freshly produced `eliza-1-<tier>.gguf` (e.g. from
-`scripts/optimize_for_milady.py`, which also writes a
-`gguf/milady_manifest.json` next to it).
+`scripts/optimize_for_eliza1.py`, which also writes a
+`gguf/eliza1_manifest.json` next to it).
 
 There are two ways to get the runtime to use that file as the
 `TEXT_LARGE` model: **point at a local file** (fast, for testing the
@@ -56,7 +56,7 @@ A catalog entry is **not required** to load a local file — the catalog
 (`runtime.kvCache`, `runtime.dflash`, `runtime.optimizations`). Without a
 catalog entry the loader falls back to plain defaults; supply the
 fork-only KV cache types (`qjl1_256` / `tbq3_0` / `q4_polar`) via
-per-load overrides if you need them and you're on the milady llama.cpp
+per-load overrides if you need them and you're on the elizaOS/llama.cpp
 fork.
 
 ## B. The machinery that picks it up
@@ -138,20 +138,20 @@ manifest:
 
 ```bash
 # Print the entry + where it goes (recommended):
-uv run python scripts/emit_milady_catalog.py \
-    --manifest checkpoints/eliza-1-0_6b/gguf/milady_manifest.json
+uv run python scripts/emit_eliza1_catalog.py \
+    --manifest checkpoints/eliza-1-0_6b/gguf/eliza1_manifest.json
 
 # Or produce a unified diff against the canonical shared catalog:
-uv run python scripts/emit_milady_catalog.py \
-    --manifest checkpoints/eliza-1-0_6b/gguf/milady_manifest.json \
+uv run python scripts/emit_eliza1_catalog.py \
+    --manifest checkpoints/eliza-1-0_6b/gguf/eliza1_manifest.json \
     --catalog packages/shared/src/local-inference/catalog.ts \
     --output reports/training/catalog-eliza-1-0_6b.diff
 ```
 
 The canonical catalog is **`packages/shared/src/local-inference/catalog.ts`**
 (`@elizaos/shared/local-inference/catalog`). The
 `packages/app-core/src/services/local-inference/catalog.ts` path is a
-re-export shim — do not edit it. `emit_milady_catalog.py` does not
+re-export shim — do not edit it. `emit_eliza1_catalog.py` does not
 rewrite the file; it prints a labeled patch fragment and names the file
 to apply it to. If you are introducing a **new** tier id (not just
 refreshing `ggufFile` / `hfRepo` on an existing tier), you must also add

packages/training/scripts/HF_PUBLISHING.md

@@ -122,7 +122,7 @@ tiers require it even for the 64k default.
 ### Recording Metal verification on a hardware host
 
 ```bash
-# On an Apple Silicon Mac with the milady checkout:
+# On an Apple Silicon Mac with the elizaOS/llama.cpp checkout:
 cd packages/inference/verify
 make metal
 ./metal_verify > metal_verify.txt
@@ -426,7 +426,7 @@ explicit.
 - **URL:** https://huggingface.co/elizaos
 - **Owner:** Eliza core team. Add new members via HF org settings.
 - **Visibility:** repos are public by default once the GGUF is real. Use
-  `--no-public` on `publish_milady_model.py` to create a private repo
+  `--no-public` on `publish_eliza1_model.py` to create a private repo
   for staging.
 
 ### One-time org setup
@@ -439,7 +439,7 @@ this once:
    `elizaos` org if it does not already exist.
 3. Invite the publishing service account so CI can push.
 
-If the org doesn't exist yet, `publish_milady_model.py` errors out
+If the org doesn't exist yet, `publish_eliza1_model.py` errors out
 explicitly with the URL above — it does not silently create the org.
 
 ### Token requirements
@@ -511,7 +511,7 @@ file expects. Schema:
     "kvV": "TBQ4_0",
     "speculativeDecode": "DFlash",
     "kernels": ["q4_polar", "qjl1_256", "tbq3_0", "tbq4_0", "dflash"],
-    "requiresFork": "elizaOS/llama.cpp@v0.1.0-milady"
+    "requiresFork": "elizaOS/llama.cpp@v1.0.0-eliza"
   },
   "drafter": {
     "repo": "elizaos/eliza-1-1_7b",
@@ -523,7 +523,7 @@ file expects. Schema:
     "publishedAt": "2026-05-10T00:00:00Z",
     "trainedFrom": "elizaos/eliza-1-9b",
     "trainingPipeline": "elizaos/eliza-1-pipeline",
-    "buildScript": "packages/training/scripts/publish_milady_model.py"
+    "buildScript": "packages/training/scripts/publish_eliza1_model.py"
   }
 }
 ```
@@ -537,7 +537,7 @@ catalog sync script can walk either side and reconstruct pairings.
 ```bash
 # Dry-run — refuses to push anything, prints the manifest and what
 # would upload. No HF_TOKEN required.
-uv run python scripts/publish_milady_model.py \
+uv run python scripts/publish_eliza1_model.py \
     --model-dir /path/to/eliza-1-1_7b \
     --repo-id elizaos/eliza-1-1_7b \
     --dry-run
@@ -547,7 +547,7 @@ uv run python scripts/publish_milady_model.py \
 # `published.json` next to the GGUF with the canonical URL + sha256
 # + size; subsequent runs skip re-upload when the sha matches the
 # existing remote LFS pointer.
-HF_TOKEN=hf_xxx uv run python scripts/publish_milady_model.py \
+HF_TOKEN=hf_xxx uv run python scripts/publish_eliza1_model.py \
     --model-dir /path/to/eliza-1-1_7b \
     --repo-id elizaos/eliza-1-1_7b
 ```

…/training/scripts/emit_milady_catalog.py …/training/scripts/emit_eliza1_catalog.pypackages/training/scripts/emit_milady_catalog.py renamed to packages/training/scripts/emit_eliza1_catalog.py packages/training/scripts/emit_milady_catalog.py renamed to packages/training/scripts/emit_eliza1_catalog.py

@@ -1,7 +1,7 @@
 """Emit a MODEL_CATALOG entry for a freshly-produced eliza-1-<tier> GGUF.
 
-After ``optimize_for_milady.py`` produces a GGUF + ``milady_manifest.json``
-the Milady runtime can only pick the model up once it has a catalog
+After ``optimize_for_eliza1.py`` produces a GGUF + ``eliza1_manifest.json``
+the Eliza-1 runtime can only pick the model up once it has a catalog
 entry. The canonical catalog (``MODEL_CATALOG``, ``ELIZA_1_TIER_IDS``,
 ``DEFAULT_ELIGIBLE_MODEL_IDS``, the HuggingFace URL builders) lives in:
 
@@ -28,12 +28,12 @@
 Usage::
 
     # Print the entry + where to put it (recommended):
-    uv run python scripts/emit_milady_catalog.py \\
-        --manifest checkpoints/eliza-1-0_6b/gguf/milady_manifest.json
+    uv run python scripts/emit_eliza1_catalog.py \\
+        --manifest checkpoints/eliza-1-0_6b/gguf/eliza1_manifest.json
 
     # Also produce a unified diff against the canonical shared catalog:
-    uv run python scripts/emit_milady_catalog.py \\
-        --manifest checkpoints/eliza-1-0_6b/gguf/milady_manifest.json \\
+    uv run python scripts/emit_eliza1_catalog.py \\
+        --manifest checkpoints/eliza-1-0_6b/gguf/eliza1_manifest.json \\
         --catalog packages/shared/src/local-inference/catalog.ts \\
         --output reports/training/catalog-eliza-1-0_6b.diff
 
@@ -67,7 +67,7 @@
     level=logging.INFO,
     format="%(asctime)s [%(levelname)s] %(message)s",
 )
-log = logging.getLogger("emit_milady_catalog")
+log = logging.getLogger("emit_eliza1_catalog")
 
 
 # The canonical catalog this script targets. Both the server
@@ -128,7 +128,7 @@
 
 
 @dataclass(frozen=True)
-class MiladyCatalogEntry:
+class Eliza1CatalogEntry:
     id: str
     display_name: str
     hf_repo: str
@@ -155,7 +155,7 @@ def to_ts_literal(self) -> str:
             "      kvCache: {\n"
             f'        typeK: "{self.cache_type_k}",\n'
             f'        typeV: "{self.cache_type_v}",\n'
-            '        requiresFork: "milady-llama-cpp",\n'
+            '        requiresFork: "buun-llama-cpp",\n'
             "      },\n"
         )
         if self.spec_type:
@@ -205,13 +205,13 @@ def _slug_from_repo(hf_repo: str) -> str:
     return last.lower()
 
 
-def build_catalog_entry(manifest: dict[str, object]) -> MiladyCatalogEntry:
+def build_catalog_entry(manifest: dict[str, object]) -> Eliza1CatalogEntry:
     base_model = str(manifest.get("base_model", ""))
     base_meta = KNOWN_BASE_MODELS.get(base_model)
     if base_meta is None:
         raise SystemExit(
             f"manifest's base_model {base_model!r} is not in KNOWN_BASE_MODELS; "
-            "add it to packages/training/scripts/emit_milady_catalog.py"
+            "add it to packages/training/scripts/emit_eliza1_catalog.py"
         )
 
     target_repo = str(manifest.get("target_repo") or "")
@@ -248,7 +248,7 @@ def build_catalog_entry(manifest: dict[str, object]) -> MiladyCatalogEntry:
                 ) or None
 
     slug = _slug_from_repo(target_repo)
-    return MiladyCatalogEntry(
+    return Eliza1CatalogEntry(
         id=slug,
         display_name=slug,
         hf_repo=target_repo,
@@ -291,7 +291,7 @@ def _find_model_catalog_close(text: str) -> int:
     return close
 
 
-def emit_diff(catalog_path: Path, new_entry: MiladyCatalogEntry) -> str:
+def emit_diff(catalog_path: Path, new_entry: Eliza1CatalogEntry) -> str:
     """Build a unified diff that inserts ``new_entry`` at the end of MODEL_CATALOG."""
     if not catalog_path.exists():
         raise SystemExit(f"catalog file does not exist: {catalog_path}")
@@ -314,7 +314,7 @@ def emit_diff(catalog_path: Path, new_entry: MiladyCatalogEntry) -> str:
     return "".join(diff_lines)
 
 
-def _entry_with_header(entry: MiladyCatalogEntry, catalog_hint: str) -> str:
+def _entry_with_header(entry: Eliza1CatalogEntry, catalog_hint: str) -> str:
     return (
         f"// Add this entry to the `MODEL_CATALOG` array in:\n"
         f"//   {catalog_hint}\n"
@@ -334,7 +334,7 @@ def main(argv: list[str] | None = None) -> int:
         "--manifest",
         type=Path,
         required=True,
-        help="Path to milady_manifest.json from optimize_for_milady.py.",
+        help="Path to eliza1_manifest.json from optimize_for_eliza1.py.",
     )
     ap.add_argument(
         "--catalog",