merge: restore WhisperLive production fixes

hybrid_server.py

@@ -859,7 +859,7 @@ print(transcription.text)</code></pre>
                 language = request.form.get('language', None)
                 task = request.form.get('task', 'transcribe')  # 'transcribe' or 'translate'
                 model_size = request.form.get('model', 'base')
-                use_vad = request.form.get('use_vad', 'true').lower() == 'true'
+                use_vad = request.args.get('use_vad', request.form.get('use_vad', 'true')).lower() == 'true'
                 
                 # For now, we'll use the shared transcriber regardless of the requested model size
                 # In the future, we could create different transcriber instances for different models

whisper_live/server.py

@@ -427,7 +427,7 @@ class ServeClientBase(object):
         self.show_prev_out_thresh = 5   # if pause(no output from whisper) show previous output for 5 seconds
         self.add_pause_thresh = 3       # add a blank to segment list as a pause(no speech) for 3 seconds
         self.transcript = []
-        self.send_last_n_segments = 10
+        self.send_last_n_segments = 30
 
         # text formatting
         self.pick_previous_segments = 2
@@ -461,9 +461,9 @@ class ServeClientBase(object):
 
         """
         self.lock.acquire()
-        if self.frames_np is not None and self.frames_np.shape[0] > 45*self.RATE:
+        if self.frames_np is not None and self.frames_np.shape[0] > 90*self.RATE:
-            self.frames_offset += 30.0
+            self.frames_offset += 60.0
-            self.frames_np = self.frames_np[int(30*self.RATE):]
+            self.frames_np = self.frames_np[int(60*self.RATE):]
             # check timestamp offset(should be >= self.frame_offset)
             # this basically means that there is no speech as timestamp offset hasnt updated
             # and is less than frame_offset
@@ -482,7 +482,7 @@ class ServeClientBase(object):
         no valid segment for the last 30 seconds from whisper
         """
         with self.lock:
-            if self.frames_np[int((self.timestamp_offset - self.frames_offset)*self.RATE):].shape[0] > 25 * self.RATE:
+            if self.frames_np[int((self.timestamp_offset - self.frames_offset)*self.RATE):].shape[0] > 60 * self.RATE:
                 duration = self.frames_np.shape[0] / self.RATE
                 self.timestamp_offset = self.frames_offset + duration - 5
 
@@ -807,10 +807,19 @@ class ServeClientFasterWhisper(ServeClientBase):
         self.same_output_threshold = 10
         self.end_time_for_same_output = None
 
-        device = "cuda" if torch.cuda.is_available() else "cpu"
+        # torch.cuda.is_available() fails when torch was compiled against a newer CUDA
+        # than the driver provides. Use ctranslate2's own CUDA probe instead, since
+        # faster_whisper relies on ctranslate2 — not torch — for inference.
+        try:
+            import ctranslate2 as _ct2
+            _cuda_types = _ct2.get_supported_compute_types("cuda")
+            device = "cuda" if _cuda_types else "cpu"
+        except Exception:
+            device = "cpu"
+
         if device == "cuda":
-            major, _ = torch.cuda.get_device_capability(device)
+            # Use int8 to stay within shared GPU memory budget (GPU 1 is shared with TTS/ComfyUI)
-            self.compute_type = "float16" if major >= 7 else "float32"
+            self.compute_type = "int8"
         else:
             self.compute_type = "int8"