fix(ci): pin legacy Qwen integration environment

Bundle the empty dllm compatibility namespace and keep the Mac integration workflow on Transformers 4.x, while leaving the Gemma/K3 production dependency range unchanged.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: 2 people

.github/workflows/integration.yaml

@@ -90,6 +90,12 @@ jobs:
           # requirements.txt rather than an editable `-e .` (which errors with
           # "does not appear to be a Python project").
           python3 -m pip install -r requirements.txt
+          # The integration suite exercises the legacy dllm-hub Qwen proposer,
+          # whose remote modeling file depends on the Transformers 4.x
+          # decoder_layer.attention_type API. Keep this runner in the dedicated
+          # legacy range; K3/Gemma production paths use requirements.txt's
+          # unbounded Transformers 5.x-compatible environment.
+          python3 -m pip install 'transformers>=4.45,<5.0'
           python3 -m pip install pytest pytest-asyncio pytest-timeout coverage
 
       - name: Run integration suite

dllm/__init__.py

@@ -0,0 +1,7 @@
+"""Compatibility namespace for legacy dllm-hub remote model imports.
+
+The legacy Qwen diffusion checkpoint declares ``dllm`` in its Transformers
+auto-map source even though Kakeya does not import runtime symbols from that
+package. Keeping this empty namespace on PYTHONPATH satisfies the static
+dependency check consistently in CI and source checkouts.
+"""

docs/quickstart.md

@@ -78,15 +78,11 @@ Same steps as Mac, plus a `torch` build matching your CUDA toolkit.
 ### Common pitfalls
 
 - **`ModuleNotFoundError: dllm`**: the legacy diffusion proposer (v0.2)
-  references a `dllm` package that v0.3 doesn't need but transformers'
-  static imports flag. Fix:
-  ```bash
-  python3 -c "import site, os; \
-      p = os.path.join(site.getusersitepackages(), 'dllm'); \
-      os.makedirs(p, exist_ok=True); \
-      open(os.path.join(p, '__init__.py'), 'a').close()"
-  ```
-  `setup_mac.sh` and `setup_cuda.sh` do this automatically.
+  declares a `dllm` package that Kakeya does not call at runtime but
+  Transformers validates statically. Current source checkouts include the empty
+  compatibility namespace at `dllm/__init__.py`; ensure the repository root is
+  on `PYTHONPATH`. Older tags can use `setup_mac.sh` / `setup_cuda.sh`, which
+  create the same namespace in site-packages.
 - **Connection refused to `huggingface.co`**: set `HF_ENDPOINT` before
   setup (see above) or pre-warm offline.