fix(whisper): bound version guard to < 4.8 + guard explicit CUDA devices (#1052)

* fix(whisper): bound ctranslate2 version guard + protect explicit CUDA devices

Addresses two post-merge nits from kwit75 on #1043:

1. Narrow version guard to (4,7) <= ct2 < (4,8) so the restriction
   auto-lifts when ctranslate2 4.8+ ships the cuBLAS 12.8.4 fix,
   rather than trapping 4.8/4.9/5.x indefinitely.

2. Run _check_gpu_compatible() for explicit device='cuda'/'cuda:N' in
   local mode, not just for auto-detect (device=None). The SIGABRT can
   occur regardless of how the CUDA device was selected.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* ci: retrigger CI (Windows HuggingFace network flake)

---------

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: anandray

packages/ai/src/ai/common/models/audio/whisper.py

@@ -89,7 +89,8 @@ def _check_gpu_compatible(cls) -> bool:
             # 1. Version guard: ctranslate2 4.7.x + CUDA 12.8 causes a
             #    tcache_thread_shutdown() SIGABRT during GPU transcription on H200
             #    (heap corruption in cuBLAS 12.8.4). Exit non-zero to force CPU.
-            #    Remove this guard once ctranslate2 ships a fix.
+            #    Upper bound at 4.8 so the guard lifts automatically once
+            #    ctranslate2 ships a fix (expected in 4.8+).
             # 2. StorageView sanity: verify a CUDA StorageView can be created via
             #    the documented from_array() API (no direct (shape,dtype,device)
             #    constructor exists in the Python bindings).
@@ -102,7 +103,7 @@ def _check_gpu_compatible(cls) -> bool:
                 'except (ValueError, AttributeError):\n'
                 '    ct2 = (999, 999)\n'
                 'cuda = torch.version.cuda or ""\n'
-                'if ct2 >= (4, 7) and cuda.startswith("12.8"):\n'
+                'if (4, 7) <= ct2 < (4, 8) and cuda.startswith("12.8"):\n'
                 '    sys.exit(1)\n'
                 't = torch.zeros(1, dtype=torch.float32, device="cuda")\n'
                 'sv = ctranslate2.StorageView.from_array(t)\n'
@@ -213,6 +214,15 @@ def load(
                     device = 'cuda'
                 else:
                     device = 'cpu'
+            elif device != 'cpu' and not WhisperLoader._check_gpu_compatible():
+                # Explicit cuda / cuda:N requested but probe failed — fall back to CPU
+                # so the same SIGABRT protection applies regardless of how the caller
+                # specified the device.
+                logger.warning(
+                    'ctranslate2 CUDA probe failed for explicit device=%r — Whisper will use CPU instead.',
+                    device,
+                )
+                device = 'cpu'
 
             if device == 'cpu':
                 gpu_index = -1