fix(whisper): harden runner initialization

python/tensorrt_model_connect/families/whisper/debug_runner.py

@@ -317,17 +317,22 @@ def generate(self, input_ids: list[int], max_new_tokens: int) -> list[dict[str,
     def __del__(self):
         if cudart is None:
             return
-        if not hasattr(self, "_d_token_id"):
-            return
-        bufs = [self._d_token_id, self._d_position_id, self._d_mask, self._d_logits,
-                self._d_mel, self._d_enc_out]
-        bufs.extend(self._d_cache_k)
-        bufs.extend(self._d_cache_v)
-        bufs.extend(self._d_present_k)
-        bufs.extend(self._d_present_v)
-        bufs.extend(self._d_cross_k)
-        bufs.extend(self._d_cross_v)
-        for d_ptr in bufs:
-            cudart.cudaFree(d_ptr)
-        if hasattr(self, "stream"):
-            cudart.cudaStreamDestroy(self.stream)
+        scalar_buffers = (
+            "_d_token_id", "_d_position_id", "_d_mask", "_d_logits", "_d_mel", "_d_enc_out",
+        )
+        list_buffers = (
+            "_d_cache_k", "_d_cache_v", "_d_present_k", "_d_present_v", "_d_cross_k", "_d_cross_v",
+        )
+        for name in scalar_buffers:
+            d_ptr = getattr(self, name, None)
+            if d_ptr is not None:
+                cudart.cudaFree(d_ptr)
+                setattr(self, name, None)
+        for name in list_buffers:
+            for d_ptr in getattr(self, name, ()):
+                cudart.cudaFree(d_ptr)
+            setattr(self, name, [])
+        stream = getattr(self, "stream", None)
+        if stream is not None:
+            cudart.cudaStreamDestroy(stream)
+            self.stream = None

python/tensorrt_model_connect/families/whisper/decoder_tp_builder.py

@@ -14,18 +14,19 @@
 
 from __future__ import annotations
 
-from typing import TYPE_CHECKING
+import importlib
 import sys
+from typing import TYPE_CHECKING
 
 import numpy as np
 from tensorrt_model_connect import trt_compat
 
-from . import graph_ops
 from ...parallel_config import (
     add_all_reduce_sum,
     normalize_parallel_config,
 )
 
+graph_ops = importlib.import_module(f"{__package__}.graph_ops")
 trt = trt_compat.get_trt()
 
 if TYPE_CHECKING:

src/runtime/models/whisper/MODEL.toml

@@ -9,5 +9,5 @@ runtime_tests = [
   "test_whisper_mel_spectrogram|test_whisper_mel_spectrogram.cpp|_|whisper_mel_spectrogram.cpp|_",
   "test_whisper_host_plan|test_whisper_host_plan.cpp|_|_|_",
   "test_whisper_decode_policy|test_whisper_decode_policy.cpp|_|_|_",
-  "test_whisper_pipeline|test_whisper_pipeline.cpp|trtmc_model_whisper|_|REQUIRES_TRT",
+  "test_whisper_pipeline|test_whisper_pipeline.cpp|trtmc_model_whisper|_|REQUIRES_TRT,REQUIRES_GPU",
 ]

tests/e2e/models/whisper/test_whisper_builder_engine.py

@@ -34,6 +34,8 @@
 
 from __future__ import annotations
 
+import importlib
+
 import numpy as np
 import pytest
 
@@ -118,9 +120,8 @@ def rand(*shape: int) -> np.ndarray:
 
 
 def _whisper_tp_builder_module():
-    return pytest.importorskip(
-        "tensorrt_model_connect.families.whisper.decoder_tp_builder",
-        reason="TensorRT is required for Whisper TP builder tests",
+    return importlib.import_module(
+        "tensorrt_model_connect.families.whisper.decoder_tp_builder"
     )