fix(glm_asr): dedup result keys for duplicate audio basenames (#3014)

The vLLM ASR result key is derived from the audio file basename, so two
inputs in different directories that share a basename (e.g. spk1/segment.wav
and spk2/segment.wav) both map to the same key. Folding the result list into
the canonical {key: text} mapping then silently drops all but the last
colliding entry, yielding fewer transcripts than inputs with no error.

Add a small _dedup_keys helper that appends a deterministic _N suffix to
later collisions (preserving order and first-occurrence names) and wire it
into generate(), warning once when a collision is resolved. Add unit tests
covering the helper and the end-to-end generate() path.

Author: SuperMarioYL

funasr/models/glm_asr/inference_vllm.py

@@ -110,6 +110,63 @@ def prepare_glmasr_vllm_dir(model_dir: str) -> str:
     return output_dir
 
 
+# Warn only once per process so batch loops do not spam the log.
+_warned_dup_keys = False
+
+
+def _dedup_keys(keys):
+    """Make result keys unique while preserving order and first-occurrence names.
+
+    Each result key is derived from the audio file basename
+    (``os.path.splitext(os.path.basename(path))[0]``), so two inputs that live
+    in different directories but share a basename -- e.g. ``spk1/segment.wav``
+    and ``spk2/segment.wav`` -- both map to ``"segment"``. A downstream
+    ``{r["key"]: r["text"]}`` mapping (the canonical FunASR result shape) would
+    then silently drop all but the last colliding entry, returning fewer
+    transcripts than inputs with no error. Appending a deterministic ``_N``
+    suffix to later collisions keeps every transcript addressable.
+
+    Args:
+        keys: Result keys in input order.
+
+    Returns:
+        A new list of unique keys, same length and order as ``keys``. The first
+        occurrence of each key is preserved unchanged; the n-th repeat becomes
+        ``"<key>_<n-1>"`` (e.g. ``"seg"`` -> ``"seg"``, ``"seg_1"``, ``"seg_2"``).
+    """
+    global _warned_dup_keys
+
+    seen = set()
+    out = []
+    collided = False
+    for key in keys:
+        if key not in seen:
+            seen.add(key)
+            out.append(key)
+            continue
+        # Find the first free "<key>_<n>" so a suffixed key cannot itself clash
+        # with an existing one (e.g. inputs "seg", "seg_1", "seg").
+        collided = True
+        n = 1
+        candidate = f"{key}_{n}"
+        while candidate in seen:
+            n += 1
+            candidate = f"{key}_{n}"
+        seen.add(candidate)
+        out.append(candidate)
+
+    if collided and not _warned_dup_keys:
+        logger.warning(
+            "Duplicate result keys from audio basenames were made unique with "
+            "'_N' suffixes (e.g. two files named 'segment.wav' in different "
+            "directories map to the same key); pass distinct filenames if you "
+            "rely on the basename as the result key."
+        )
+        _warned_dup_keys = True
+
+    return out
+
+
 class GLMASRVLLMEngine:
     """GLM-ASR with vLLM backend.
 
@@ -263,13 +320,18 @@ def generate(self, inputs, prompt="转录以下音频内容", max_new_tokens=500
         t2 = time.perf_counter()
         logger.info(f"vLLM generation: {t2-t1:.3f}s")
 
+        raw_keys = [
+            os.path.splitext(os.path.basename(x))[0] if isinstance(x, str) else f"sample_{i}"
+            for i, x in enumerate(inputs)
+        ]
+        keys = _dedup_keys(raw_keys)
+
         results = []
         for i, output in enumerate(outputs):
             token_ids = list(output.outputs[0].token_ids)
             text = self.tokenizer.decode(token_ids, skip_special_tokens=True)
             text = re.sub(r'\s+', ' ', text).strip()
-            key = os.path.splitext(os.path.basename(inputs[i]))[0] if isinstance(inputs[i], str) else f"sample_{i}"
-            results.append({"key": key, "text": text})
+            results.append({"key": keys[i], "text": text})
 
         return results
 

tests/test_glm_asr_vllm_dedup_keys.py

@@ -0,0 +1,124 @@
+"""Unit tests for GLM-ASR vLLM result-key deduplication.
+
+These tests exercise ``GLMASRVLLMEngine.generate`` without a GPU or a real
+vLLM installation: the vLLM entry points are stubbed in ``sys.modules`` and the
+audio/encoder/engine collaborators are mocked, so only the result-key wiring is
+under test. The pure ``_dedup_keys`` helper is also tested directly.
+"""
+
+import sys
+import types
+import unittest
+from unittest import mock
+
+
+def _install_vllm_stub():
+    """Install a minimal ``vllm`` stub so ``inference_vllm`` imports without GPU."""
+
+    class _SamplingParams:
+        def __init__(self, **kwargs):
+            self.kwargs = kwargs
+
+    class _EmbedsPrompt:
+        def __init__(self, **kwargs):
+            self.kwargs = kwargs
+
+    vllm_mod = types.ModuleType("vllm")
+    vllm_mod.SamplingParams = _SamplingParams
+    vllm_mod.LLM = object
+    inputs_mod = types.ModuleType("vllm.inputs")
+    inputs_mod.EmbedsPrompt = _EmbedsPrompt
+    data_mod = types.ModuleType("vllm.inputs.data")
+    data_mod.EmbedsPrompt = _EmbedsPrompt
+
+    sys.modules["vllm"] = vllm_mod
+    sys.modules["vllm.inputs"] = inputs_mod
+    sys.modules["vllm.inputs.data"] = data_mod
+
+
+class DedupKeysHelperTest(unittest.TestCase):
+    def setUp(self):
+        _install_vllm_stub()
+        from funasr.models.glm_asr import inference_vllm
+
+        self.inference_vllm = inference_vllm
+        # Reset the warn-once flag so warning-related assertions are deterministic.
+        inference_vllm._warned_dup_keys = False
+
+    def test_collision_free_input_unchanged(self):
+        keys = ["a", "b", "c"]
+        self.assertEqual(self.inference_vllm._dedup_keys(keys), ["a", "b", "c"])
+
+    def test_repeated_keys_get_deterministic_suffix(self):
+        keys = ["seg", "seg", "other", "seg"]
+        self.assertEqual(
+            self.inference_vllm._dedup_keys(keys),
+            ["seg", "seg_1", "other", "seg_2"],
+        )
+
+    def test_suffix_does_not_clash_with_existing_key(self):
+        # A naive "seg" -> "seg_1" scheme would re-collide with the literal
+        # "seg_1" input; the result must stay globally unique.
+        keys = ["seg", "seg_1", "seg"]
+        result = self.inference_vllm._dedup_keys(keys)
+        self.assertEqual(result, ["seg", "seg_1", "seg_2"])
+        self.assertEqual(len(set(result)), len(result))
+
+    def test_empty_input(self):
+        self.assertEqual(self.inference_vllm._dedup_keys([]), [])
+
+    def test_sentinel_keys_preserved(self):
+        keys = ["sample_0", "sample_1", "sample_0"]
+        self.assertEqual(
+            self.inference_vllm._dedup_keys(keys),
+            ["sample_0", "sample_1", "sample_0_1"],
+        )
+
+    def test_does_not_mutate_input(self):
+        keys = ["seg", "seg"]
+        self.inference_vllm._dedup_keys(keys)
+        self.assertEqual(keys, ["seg", "seg"])
+
+
+class DedupResultKeysTest(unittest.TestCase):
+    def setUp(self):
+        _install_vllm_stub()
+        from funasr.models.glm_asr.inference_vllm import GLMASRVLLMEngine
+
+        # Build an engine without running __init__ (no model load / GPU needed).
+        engine = GLMASRVLLMEngine.__new__(GLMASRVLLMEngine)
+        engine.device = "cpu"
+        engine._encode_audio = mock.Mock(return_value="audio_embeds")
+        engine._build_prompt_embeds = mock.Mock(
+            return_value=mock.Mock(float=lambda: "embeds")
+        )
+        engine.tokenizer = mock.Mock()
+        engine.tokenizer.decode = mock.Mock(return_value="hello world")
+        self.engine = engine
+
+    def _set_outputs(self, n):
+        token_out = types.SimpleNamespace(token_ids=[1, 2, 3])
+        outs = [types.SimpleNamespace(outputs=[token_out]) for _ in range(n)]
+        self.engine.vllm_engine = mock.Mock()
+        self.engine.vllm_engine.generate = mock.Mock(return_value=outs)
+
+    def test_duplicate_basenames_get_unique_keys(self):
+        self._set_outputs(2)
+        results = self.engine.generate(["spk1/segment.wav", "spk2/segment.wav"])
+        self.assertEqual([r["key"] for r in results], ["segment", "segment_1"])
+        # No transcript is dropped when results are folded into a {key: text} dict.
+        self.assertEqual(len({r["key"] for r in results}), len(results))
+
+    def test_distinct_basenames_unchanged(self):
+        self._set_outputs(2)
+        results = self.engine.generate(["a.wav", "b.wav"])
+        self.assertEqual([r["key"] for r in results], ["a", "b"])
+
+    def test_single_input_key(self):
+        self._set_outputs(1)
+        results = self.engine.generate("only.wav")
+        self.assertEqual(results, [{"key": "only", "text": "hello world"}])
+
+
+if __name__ == "__main__":
+    unittest.main()