chore[vla]: bump @qvac/vla-ggml to 0.1.0 for initial release

packages/vla-ggml/CHANGELOG.md

@@ -1,45 +1,21 @@
 # Changelog
 
-## 0.2.0
-
-- `VlaModel.run()` now returns `{ actions, stats }` instead of a raw
-  `Float32Array`. The `stats` object carries per-stage wall-clock timings
-  (`vision_ms`, `smollm2_compute_ms`, `smollm2_total_ms`, `ode_ms`,
-  `total_ms`) captured during inference.
-- Integration test: added tolerance-based assertion against a committed
-  PyTorch reference output (`test/integration/assets/pt_actions_libero_fixed.json`)
-  and wired the shared performance reporter
-  (`scripts/test-utils/performance-reporter.js`, `addonType: 'vla'`).
-- Input validation hardened: `model.run()` now rejects mismatched
-  `imgWidth` / `imgHeight` (must equal `hparams.visionImageSize`) at both
-  the JS and C++ layers, instead of letting the conv2d/reshape mismatch
-  trip a `GGML_ASSERT` that hard-aborts the worker. Equivalent guards
-  added for `n_images`, `lang_len`, and `state_dim`.
-- GGUF load hardened: rejects malformed files with out-of-range hparams,
-  mismatched `text_num_layers` / `expert_num_layers`, or any missing
-  required tensor, instead of null-dereffing on the first inference.
-  Per-tensor mmap-bounds check rejects crafted GGUFs whose tensor
-  `(offset, nbytes)` would point outside the mapped region.
-- Inference path returns a clean `false` (and surfaces an `Error` event
-  to JS) when any ggml graph allocation fails, instead of computing
-  against a partially-initialised graph.
-- `_hasActiveResponse` clearing migrated to `.finally()` on the response
-  promise so a worker abort can't wedge subsequent `run()` calls with
-  `JOB_ALREADY_RUNNING`. Mirrors the pattern used by
-  `qvac-lib-infer-llamacpp-llm`.
-- `smolvla_load_model` split from one 600-line function into a ~125-line
-  orchestrator plus 9 file-static helpers; RAII guards for `FILE*` /
-  `gguf_context` / fd close out every error branch automatically.
-- Bumped `qvac-fabric` to `>=8189.0.2` and
-  `qvac-lib-inference-addon-cpp` to `>=1.1.7#1`. The new addon-cpp port
-  renamed its include namespace from `qvac-lib-inference-addon-cpp/*` to
-  `inference-addon-cpp/*`; lint-cpp moved its config from
-  `share/qvac-lint-cpp/` to `share/lint-cpp/`. Source/CMakeLists updated
-  to match.
-
 ## 0.1.0
 
 - Initial release of `@qvac/vla-ggml`. Ports the SmolVLA vision-language-action
   model to ggml with Vulkan / Metal / OpenCL / CPU backends. Bundles the
   full SigLIP vision encoder, SmolLM2 text tower, action expert, and
   10-step flow-matching ODE in a single Bare addon.
+- `VlaModel.run()` returns `{ actions, stats }` where `stats` carries
+  per-stage wall-clock timings (`vision_ms`, `smollm2_compute_ms`,
+  `smollm2_total_ms`, `ode_ms`, `total_ms`).
+- Input validation: `model.run()` rejects mismatched `imgWidth` /
+  `imgHeight` (must equal `hparams.visionImageSize`), `n_images`,
+  `lang_len`, and `state_dim` at both the JS and C++ layers.
+- GGUF load rejects malformed files with out-of-range hparams,
+  mismatched `text_num_layers` / `expert_num_layers`, missing required
+  tensors, or per-tensor `(offset, nbytes)` pointing outside the mmap
+  region.
+- Integration test includes a tolerance-based assertion against a
+  committed PyTorch reference output and wires the shared performance
+  reporter (`addonType: 'vla'`).

packages/vla-ggml/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@qvac/vla-ggml",
-  "version": "0.2.0",
+  "version": "0.1.0",
   "description": "SmolVLA vision-language-action inference addon for QVAC (ggml backend)",
   "addon": true,
   "engines": {