chore(core+ui+tests): refactor vLLM plugin and UI updates, add CodeQL integration

- Introduced `qwen_asr.vllm_plugin` for vLLM model registry, improving customization.
- Updated exception handling for `BaseDummyInputsBuilder` imports in `vllm_backend`.
- Fixed `cnn_len` iteration issue in `transformers_backend` by switching to `.tolist()`.
- Enhanced UI callbacks for safe API base URL validation and realtime session ID handling.
- Added new tests for URL and session ID validation, aligning with OpenAI compatibility.
- Integrated `run_codeql.ps1` into scripts and `Taskfile.yml` to enable security analysis.
- Streamlined imports with `TYPE_CHECKING`, optimized logging, and removed unused `logger`.
- Updated DockerHub documentation banner to use new logo.

Author: TheMasterOfDisasters

.gitignore

@@ -23,4 +23,5 @@ venv/
 env/
 
 .idea
+.codeql/
 testbench/results/*.json

Taskfile.yml

@@ -86,6 +86,11 @@ tasks:
     cmds:
       - 'cmd /c ".venv\Scripts\python.exe" -m compileall -q qwen_asr testbench examples'
 
+  codeql:
+    desc: Run CodeQL Python security-and-quality analysis
+    cmds:
+      - powershell -NoProfile -ExecutionPolicy Bypass -File scripts/run_codeql.ps1
+
   deps:
     desc: Regenerate Linux/Python 3.13 requirements.txt from requirements.in
     cmds:

docs/dockerhub.md

@@ -1,5 +1,5 @@
 <p>
-  <img src="https://github.com/Hangry-Labs/Qwen3-ASR-STT/raw/main/hangrylabs/banner.jpg" alt="Hangry Labs Qwen3-ASR-STT banner">
+  <img src="https://github.com/Hangry-Labs/Qwen3-ASR-STT/raw/main/logo.jpg" alt="Hangry Labs Qwen3-ASR-STT banner">
 </p>
 
 # Hangry Labs Qwen3-ASR-STT

pyproject.toml

@@ -49,6 +49,9 @@ Homepage = "https://github.com/hangry-labs/Qwen3-ASR-STT"
 Repository = "https://github.com/hangry-labs/Qwen3-ASR-STT"
 "Upstream Project" = "https://github.com/QwenLM/Qwen3-ASR"
 
+[project.entry-points."vllm.general_plugins"]
+qwen_asr = "qwen_asr.vllm_plugin:register"
+
 [tool.setuptools]
 packages = { find = { where = ["."] , include = ["qwen_asr*"] } }
 include-package-data = true

qwen_asr/__init__.py

@@ -17,9 +17,11 @@
 qwen_asr: Qwen3-ASR package.
 """
 from importlib.metadata import PackageNotFoundError, version
+from typing import TYPE_CHECKING
 
-from .inference.qwen3_asr import Qwen3ASRModel
-from .inference.qwen3_forced_aligner import Qwen3ForcedAligner
+if TYPE_CHECKING:
+    from .inference.qwen3_asr import Qwen3ASRModel
+    from .inference.qwen3_forced_aligner import Qwen3ForcedAligner
 
 from .inference.utils import parse_asr_output
 
@@ -29,3 +31,15 @@
     __version__ = "0.0.0"
 
 __all__ = ["Qwen3ASRModel", "Qwen3ForcedAligner", "parse_asr_output", "__version__"]
+
+
+def __getattr__(name: str):
+    if name == "Qwen3ASRModel":
+        from .inference.qwen3_asr import Qwen3ASRModel
+
+        return Qwen3ASRModel
+    if name == "Qwen3ForcedAligner":
+        from .inference.qwen3_forced_aligner import Qwen3ForcedAligner
+
+        return Qwen3ForcedAligner
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")

qwen_asr/core/__init__.py

@@ -0,0 +1,2 @@
+# coding=utf-8
+

qwen_asr/core/transformers_backend/configuration_qwen3_asr.py

@@ -13,10 +13,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from transformers.configuration_utils import PretrainedConfig
-from transformers.utils import logging
-
-
-logger = logging.get_logger(__name__)
 
 
 class Qwen3ASRAudioEncoderConfig(PretrainedConfig):

qwen_asr/core/transformers_backend/modeling_qwen3_asr.py

@@ -718,7 +718,7 @@ def forward(
         hidden_states = padded_embed[padded_mask_after_cnn]
         cu_chunk_lens = [0]
         window_aftercnn = padded_mask_after_cnn.shape[-1] * (self.n_window_infer // (self.n_window * 2))
-        for cnn_len in aftercnn_lens:
+        for cnn_len in aftercnn_lens.tolist():
             cu_chunk_lens += [window_aftercnn] * (cnn_len // window_aftercnn)
             remainder = cnn_len % window_aftercnn
             if remainder != 0:
@@ -1187,6 +1187,9 @@ def forward(
             (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
         """
 
+        if audio_feature_lengths is None and feature_attention_mask is not None:
+            audio_feature_lengths = torch.sum(feature_attention_mask, dim=1)
+
         if inputs_embeds is None:
             # 1. Extract the input embeddings
             inputs_embeds = self.get_input_embeddings()(input_ids)
@@ -1202,11 +1205,6 @@ def forward(
             audio_mask = self.get_placeholder_mask(input_ids, inputs_embeds=inputs_embeds)
             inputs_embeds = inputs_embeds.masked_scatter(audio_mask, audio_features)
 
-        if feature_attention_mask is not None:
-            audio_feature_lengths = torch.sum(feature_attention_mask, dim=1)
-        else:
-            audio_feature_lengths = None
-
         if attention_mask is not None and position_ids is None:
             if (
                 cache_position is None

qwen_asr/core/vllm_backend/qwen3_asr.py

@@ -106,7 +106,7 @@
 
 try:
     from vllm.multimodal.profiling import BaseDummyInputsBuilder
-except:
+except ImportError:
     from vllm.multimodal.processing import BaseDummyInputsBuilder
 
 logger = init_logger(__name__)
@@ -994,4 +994,4 @@ def get_generation_prompt(
             "prompt_token_ids": prompt_token_ids,
             "multi_modal_data": {"audio": audio},
         }
-        return cast(PromptType, prompt_dict)
+        return cast(PromptType, prompt_dict)

qwen_asr/inference/__init__.py

@@ -0,0 +1,2 @@
+# coding=utf-8
+