Merge branch 'develop' of https://github.com/elizaOS/eliza into develop

Author: lalalune

.github/workflows/lifeops-bench-multi-tier.yml

@@ -295,7 +295,7 @@ jobs:
         id: dflash-cache
         uses: actions/cache@v4
         with:
-          path: ~/.cache/eliza-dflash/milady-llama-cpp/build/bin/llama-server
+          path: ~/.cache/eliza-dflash/eliza-llama-cpp/build/bin/llama-server
           key: dflash-llama-cpp-${{ runner.os }}-v1
 
       - name: Build dflash fork (best-effort)

.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:
@@ -379,7 +379,7 @@ jobs:
 
       - name: Cross-build windows-x64-cpu
         env:
-          # Override the dflash fork to elizaOS/llama.cpp v0.1.0-milady so
+          # Override the dflash fork to elizaOS/llama.cpp v1.0.0-eliza so
           # CI exercises the symbols downstream consumers actually expect.
           # Operators that want to test against the spiritbuun upstream can
           # leave this unset.
@@ -388,7 +388,7 @@ jobs:
           set -euo pipefail
           node packages/app-core/scripts/build-llama-cpp-dflash.mjs \
             --target windows-x64-cpu \
-            --ref v0.1.0-milady \
+            --ref v1.0.0-eliza \
             --cache-dir "$RUNNER_TEMP/llama-cpp-cross-cache" \
             --out-dir "$GITHUB_WORKSPACE/build-output/windows-x64-cpu"
 
@@ -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"

.gitmodules

@@ -1,14 +1,16 @@
 [submodule "packages/inference/llama.cpp"]
+	# The single canonical llama.cpp checkout for the whole repo. This is the
+	# elizaOS/llama.cpp fork (@ v1.0.0-eliza, commit 08032d57): the unified
+	# fork with the milady kernels (Q4_POLAR / QJL1_256 / TBQ4_0 / TBQ3_0
+	# GGML types + Metal/Vulkan/CUDA kernels) and DFlash spec-decode. The
+	# host build (build-llama-cpp-dflash.mjs) + AOSP cross-compile
+	# (aosp/compile-libllama.mjs) default to this submodule; bun's postinstall
+	# (scripts/ensure-llama-cpp-submodule.mjs) initializes it. The fork is
+	# itself a llama.cpp fork, so it carries convert_hf_to_gguf.py /
+	# llama-quantize / llama-cli too — the training pipeline's plain Q4_K_M
+	# GGUF path uses the fork's tooling (there is no separate "stock upstream"
+	# submodule). build/ is gitignored by llama.cpp's own .gitignore so only
+	# the gitlink (commit SHA) is tracked.
 	path = packages/inference/llama.cpp
 	url = https://github.com/elizaOS/llama.cpp.git
 	branch = eliza/main
-[submodule "packages/training/vendor/llama.cpp"]
-	# Stock upstream llama.cpp pinned to a release tag (b6650). Used by the
-	# training pipeline's plain GGUF Q4_K_M path (convert_hf_to_gguf.py +
-	# llama-quantize + llama-cli). The Milady fork — Q4_POLAR/QJL1_256/TBQ
-	# GGML types — is the *other* submodule (packages/inference/llama.cpp).
-	# scripts/vendor_llama_cpp.sh inits + builds this; build/ is gitignored
-	# by llama.cpp's own .gitignore so only the gitlink (commit SHA) is tracked.
-	path = packages/training/vendor/llama.cpp
-	url = https://github.com/ggml-org/llama.cpp.git
-	shallow = true