Upload k3_bootstrap.sh with huggingface_hub

k3_bootstrap.sh

@@ -305,7 +305,18 @@ c = json.loads(p.read_text()); old = c.get("architectures")
 c["architectures"] = ["Qwen3DSparkModel"]; p.write_text(json.dumps(c, indent=2))
 print(f"architectures {old} -> {c['architectures']}")
 PY
+      # The converter needs the TARGET model's tokenizer to write the draft's
+      # vocabulary ("DFlash draft model requires --target-model-dir"). Only the
+      # tokenizer files are fetched -- five small files, not the 2.8 TB of
+      # weights.
+      mkdir -p /models/k3-tok
+      for f in config.json tokenizer_config.json tiktoken.model \
+               tokenization_kimi.py encoding_k3.py; do
+        hf download moonshotai/Kimi-K3 "$f" --local-dir /models/k3-tok >/dev/null 2>&1
+      done
+      log "target tokenizer files: $(ls /models/k3-tok 2>/dev/null | wc -l)/5"
       python3 "$SRC/convert_hf_to_gguf.py" /models/dspark-src \
+          --target-model-dir /models/k3-tok \
           --outfile "$DRAFT_DIR/$DS_FIXED" --outtype bf16 2>&1 | tail -12
       if [ -s "$DRAFT_DIR/$DS_FIXED" ]; then
         log "DSpark draft built: $(du -h "$DRAFT_DIR/$DS_FIXED" | cut -f1)"