multi-GPU: translator.py

src/translator.py

@@ -15,7 +15,7 @@ import sentencepiece as spm
 from huggingface_hub import snapshot_download
 
 from chunker import split_chunks
-from hardware import HardwareProfile, detect_hardware_profile
+from hardware import HardwareProfile, detect_hardware_profile, resolve_gpu_indices
 from token_chunker import source_token_ids, split_for_translation
 
 ROOT = Path(__file__).resolve().parent.parent
@@ -560,16 +560,33 @@ class HachimiTranslator:
         elif env_compute_type and ct2_device == "cuda":
             attempts.append(("cpu", "int8_float32"))
 
+        # Multi-GPU: nếu có >1 GPU (vd Kaggle T4x2) → CT2 chia batch ra các GPU
+        # (~1.69× compute đo thực). Chỉ áp cho device cuda; CPU bỏ qua device_index.
+        try:
+            cuda_count = ctranslate2.get_cuda_device_count()
+        except Exception:
+            cuda_count = 1
+        gpu_indices = resolve_gpu_indices(
+            cuda_count, os.environ.get("HACHIMIMT_GPU_INDICES")
+        )
+
         translator = None
         last_error: Exception | None = None
         for device, compute_type in attempts:
             try:
-                translator = ctranslate2.Translator(
-                    str(model_path / config.ct2_subdir),
+                kwargs = dict(
                     device=device,
                     compute_type=compute_type,
                     intra_threads=self._ct2_threads,
-                    inter_threads=self._ct2_inter_threads,
+                )
+                if device == "cuda" and len(gpu_indices) > 1:
+                    # 1 worker/GPU (CT2 khuyến nghị) + gán nhiều device.
+                    kwargs["device_index"] = gpu_indices
+                    kwargs["inter_threads"] = len(gpu_indices)
+                else:
+                    kwargs["inter_threads"] = self._ct2_inter_threads
+                translator = ctranslate2.Translator(
+                    str(model_path / config.ct2_subdir), **kwargs
                 )
                 self._ct2_compute_type = compute_type
                 break