Sync from GitHub f50c848

app.py

@@ -53,7 +53,7 @@ import warnings
 
 import gradio as gr
 from fastapi.responses import FileResponse, HTMLResponse, Response
-from huggingface_hub import snapshot_download
+from huggingface_hub import HfApi, snapshot_download
 
 from core.constants import (
     DEFAULT_TOP_K,
@@ -127,6 +127,9 @@ def sync_library() -> None:
         snapshot_download(
             LIBRARY_DATASET_ID, repo_type="dataset", local_dir=PREINDEXED_DIR
         )
+        remote_files = HfApi().list_repo_files(
+            LIBRARY_DATASET_ID, repo_type="dataset"
+        )
         log.info("library synced from %s", LIBRARY_DATASET_ID)
     except Exception as e:
         log.warning("library dataset not synced (%s): %s", LIBRARY_DATASET_ID, e)
@@ -141,6 +144,25 @@ def sync_library() -> None:
         if os.path.isdir(path) and os.path.isfile(os.path.join(path, "index.json")):
             print(f"Pruning stale pre-migration doc dir: {entry}")
             shutil.rmtree(path, ignore_errors=True)
+    # snapshot_download only adds/updates; it never removes a local doc dir that
+    # was deleted from the dataset. Prune any method/doc dir absent remotely so
+    # deletions in the library propagate to the Space on the next sync.
+    remote_docs = {
+        (parts[0], parts[1])
+        for f in remote_files
+        if len(parts := f.split("/")) >= 3 and parts[0] in (VISUAL_SUBDIR, PARSED_SUBDIR)
+    }
+    for method in (VISUAL_SUBDIR, PARSED_SUBDIR):
+        method_dir = os.path.join(PREINDEXED_DIR, method)
+        if not os.path.isdir(method_dir):
+            continue
+        for doc_id in os.listdir(method_dir):
+            doc_dir = os.path.join(method_dir, doc_id)
+            if doc_id.startswith(".") or not os.path.isdir(doc_dir):
+                continue
+            if (method, doc_id) not in remote_docs:
+                print(f"Pruning doc removed from library: {method}/{doc_id}")
+                shutil.rmtree(doc_dir, ignore_errors=True)
 
 
 sync_library()
@@ -381,6 +403,7 @@ _VENDOR_MEDIA_TYPES = {
     ".mjs": "text/javascript",
     ".wasm": "application/wasm",
     ".onnx": "application/octet-stream",
+    ".json": "application/json",  # Moonshine config/tokenizer (self-hosted ASR)
 }
 
 

frontend/index.html

@@ -763,20 +763,20 @@ function repairGuy() {
       window.ort.env.wasm.wasmPaths = '/vendor/onnx/';
       await this.loadScript('/vendor/vad/bundle.min.js');
 
-      // 2) Moonshine ASR (transformers.js from CDN; the tiny model is fetched
-      //    from the HF hub once, then cached in the browser). Pinned to a fixed
-      //    v4 so the import can't silently float to a newer default.
-      //    dtype MUST be a single string, NOT a per-module object: it's applied
-      //    to every session unconditionally. v4's default for this model picks
+      // 2) Moonshine ASR — transformers.js library from CDN (pinned), but the
+      //    MODEL is SELF-HOSTED under /vendor/moonshine-tiny-ONNX/ and remote
+      //    fetching is OFF. This is deliberate: left to the hub, v4 defaults to
       //    the 4-bit (q4) merged decoder, whose MatMulNBits export is malformed
       //    for onnxruntime-web and dies at session creation ("Missing required
-      //    scale … embed_tokens"). A per-module object doesn't help here —
-      //    Moonshine's encoder session key is `model` (not `encoder_model`), so
-      //    a partial object is ignored and the decoder falls back to q4. The q8
-      //    graph is verified MatMulNBits-free. Bump to 'fp32' for best accuracy
-      //    on part numbers (≈4x larger first download).
-      const { pipeline } = await import('https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.2.0');
-      this.asr = await pipeline('automatic-speech-recognition', 'onnx-community/moonshine-tiny-ONNX', { dtype: 'q8', device: 'wasm' });
+      //    scale … embed_tokens") — and that bad file then sticks in the browser
+      //    cache. We ship ONLY the q8 files (verified MatMulNBits-free), so there
+      //    is no q4 to pick, no hub round-trip, and no cached-q4 ghost. dtype is
+      //    a single string (applied to every session); device defaults to wasm.
+      const { pipeline, env } = await import('https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.2.0');
+      env.allowRemoteModels = false;   // never touch the hub — local files only
+      env.allowLocalModels = true;
+      env.localModelPath = '/vendor/'; // model id is appended → /vendor/moonshine-tiny-ONNX/…
+      this.asr = await pipeline('automatic-speech-recognition', 'moonshine-tiny-ONNX', { dtype: 'q8', device: 'wasm' });
 
       // 3) the mic: Silero VAD segments speech; onSpeechEnd is our send trigger
       this.vad = await window.vad.MicVAD.new({

frontend/vendor/moonshine-tiny-ONNX/config.json

@@ -0,0 +1,34 @@
+{
+  "architectures": [
+    "MoonshineForConditionalGeneration"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 1,
+  "decoder_hidden_act": "silu",
+  "decoder_num_attention_heads": 8,
+  "decoder_num_hidden_layers": 6,
+  "decoder_num_key_value_heads": 8,
+  "decoder_start_token_id": 1,
+  "encoder_hidden_act": "gelu",
+  "encoder_num_attention_heads": 8,
+  "encoder_num_hidden_layers": 6,
+  "encoder_num_key_value_heads": 8,
+  "eos_token_id": 2,
+  "hidden_size": 288,
+  "initializer_range": 0.02,
+  "intermediate_size": 1152,
+  "is_encoder_decoder": true,
+  "max_position_embeddings": 512,
+  "model_type": "moonshine",
+  "num_attention_heads": 8,
+  "num_hidden_layers": 6,
+  "num_key_value_heads": 8,
+  "partial_rotary_factor": 0.9,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "torch_dtype": "float32",
+  "transformers_version": "4.48.0.dev0",
+  "use_cache": true,
+  "vocab_size": 32768
+}

frontend/vendor/moonshine-tiny-ONNX/generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "decoder_start_token_id": 1,
+  "eos_token_id": 2,
+  "transformers_version": "4.47.0.dev0"
+}

frontend/vendor/moonshine-tiny-ONNX/onnx/decoder_model_merged_quantized.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eed87831c3a6103534aae7d47a5d485025c659a1323901513961c39fe8a1a367
+size 20243286

frontend/vendor/moonshine-tiny-ONNX/onnx/encoder_model_quantized.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6fc4b7bc5af75c0591fd157a1f3829b533d18e9769a888fd95a62e470dd4f4a
+size 7937661

frontend/vendor/moonshine-tiny-ONNX/preprocessor_config.json

@@ -0,0 +1,5 @@
+{
+  "feature_extractor_type": "MoonshineFeatureExtractor",
+  "processor_class": "MoonshineProcessor",
+  "sampling_rate": 16000
+}

frontend/vendor/moonshine-tiny-ONNX/special_tokens_map.json

@@ -0,0 +1 @@
+{}