diff --git a/examples/transformers/README.md b/examples/transformers/README.md
index d1c75be2..dc55d7b5 100644
--- a/examples/transformers/README.md
+++ b/examples/transformers/README.md
@@ -54,6 +54,42 @@ per file. `keywords` and `prompt` are hints, not enforced vocabulary. Missing EO
 or empty text produces a nonzero exit after printing the diagnostic result.
 EOS is a generation boundary, not proof that every word was recognized.
 
+## CUDA: an isolated, tested recipe
+
+Use a separate environment for this GPU recipe; do not install the CPU
+requirements into it. It was functionally tested on Linux x86-64, Python 3.12,
+an NVIDIA H100 80 GB and driver **550.127.08**, with PyTorch/torchaudio
+**2.11.0+cu128** and Transformers **5.17.0**. Other GPU/driver combinations need
+their own validation.
+
+```bash
+python3.12 -m venv .venv-native-gpu
+. .venv-native-gpu/bin/activate
+python -m pip install --index-url https://download.pytorch.org/whl/cu128 'torch==2.11.0+cu128' 'torchaudio==2.11.0+cu128'
+python -m pip install -r examples/transformers/requirements-gpu.txt
+python -m pip check
+python examples/transformers/transcribe.py --device cuda --dtype float32
+python examples/transformers/transcribe.py chinese.wav english.wav --language zh en --device cuda --dtype bfloat16
+```
+
+The last command uses your local files. Without device flags, the CLI still uses
+CPU float32. CUDA requests fail if CUDA is unavailable; they never silently fall
+back to CPU. BF16 also requires device support. Model weights and processor
+tensors are moved to the selected device without converting integer token IDs
+to a floating dtype. JSON output records the actual device, dtype and versions.
+
+On 2026-09-10, both float32 and BF16 passed English, Chinese, Chinese with
+keywords, and padded Chinese/English batch inference using the pinned public
+samples. Outputs were nonempty and reached EOS. The Chinese sample still
+transcribed `开放时间` as `开饭时间`, including with the keyword hint; these checks
+do not establish accuracy or keyword benefit. They are not throughput, minimum
+VRAM or concurrency benchmarks. Float16 and other accelerators were not tested.
+
+Transformers emitted an attention-implementation warning in this environment.
+Successful inference does not verify the attention kernel of every component;
+this recipe makes no Flash Attention or all-SDPA performance claim. The online
+demo and notebook linked above do not imply a verified hosted GPU environment.
+
 ## Select a backend, not just a suffix
 
 | Need | Artifact and entry point |
@@ -66,8 +102,8 @@ EOS is a generation boundary, not proof that every word was recognized.
 The native `-hf` export produces transcription text. It does not add word
 timestamps, speaker identities, a streaming protocol or an HTTP server. The
 31-language MLT checkpoint is a separate model, not an alternate name for this
-zh/en/ja export. GPU dtype, attention kernels and throughput require separate
-hardware validation; the CPU sample is not a GPU benchmark.
+zh/en/ja export. The CUDA recipe above covers the stated functional cases only;
+attention kernels and serving throughput require separate hardware evaluation.
 
 ## Verification
 
diff --git a/examples/transformers/requirements-gpu.txt b/examples/transformers/requirements-gpu.txt
new file mode 100644
index 00000000..15c64e6d
--- /dev/null
+++ b/examples/transformers/requirements-gpu.txt
@@ -0,0 +1,8 @@
+transformers==5.17.0
+torch==2.11.0+cu128
+torchaudio==2.11.0+cu128
+numpy==1.26.4
+librosa==0.11.0
+soundfile==0.13.1
+huggingface-hub==1.30.0
+tokenizers==0.23.2
diff --git a/examples/transformers/transcribe.py b/examples/transformers/transcribe.py
index d61389ec..a431bf2e 100644
--- a/examples/transformers/transcribe.py
+++ b/examples/transformers/transcribe.py
@@ -1,4 +1,4 @@
-"""Native Transformers CPU example for short, authorized recordings."""
+"""Native Transformers example for short, authorized recordings; CPU by default."""
 import argparse
 import json
 from pathlib import Path
@@ -38,6 +38,17 @@ def languages_for_batch(languages, count):
     return languages * count if len(languages) == 1 else languages
 
 
+def validate_runtime(device, dtype, *, cuda_available=False, bf16_supported=False):
+    if device not in ("cpu", "cuda") or dtype not in ("float32", "bfloat16"):
+        raise ValueError("Choose cpu/cuda and float32/bfloat16 explicitly")
+    if device == "cpu" and dtype != "float32":
+        raise ValueError("This example supports float32 on CPU; use CUDA for bfloat16")
+    if device == "cuda" and not cuda_available:
+        raise ValueError("CUDA was requested but is unavailable; no CPU fallback")
+    if device == "cuda" and dtype == "bfloat16" and not bf16_supported:
+        raise ValueError("The selected CUDA device does not support bfloat16")
+
+
 def load_audio(path):
     import soundfile as sf
 
@@ -55,10 +66,17 @@ def main():
     parser.add_argument("--keywords", nargs="*", default=None)
     parser.add_argument("--prompt", default=None)
     parser.add_argument("--max-new-tokens", type=int, default=256)
+    parser.add_argument("--device", choices=["cpu", "cuda"], default="cpu")
+    parser.add_argument("--dtype", choices=["float32", "bfloat16"], default="float32")
     args = parser.parse_args()
     if not 1 <= args.max_new_tokens <= 1024:
         parser.error("--max-new-tokens must be between 1 and 1024")
     languages = languages_for_batch(args.language, len(args.audio) or 1)
+    import torch
+
+    cuda_available = args.device == "cuda" and torch.cuda.is_available()
+    validate_runtime(args.device, args.dtype, cuda_available=cuda_available,
+                     bf16_supported=cuda_available and torch.cuda.is_bf16_supported())
     if not args.audio:
         from huggingface_hub import hf_hub_download
 
@@ -67,7 +85,6 @@ def main():
     if sum(len(item) for item in audio) > MAX_SECONDS * SAMPLE_RATE:
         raise ValueError("Keep total batch audio within 60 seconds for this example")
 
-    import torch
     import transformers
     from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
 
@@ -75,13 +92,13 @@ def main():
     processor = AutoProcessor.from_pretrained(MODEL_ID, revision=REVISION, trust_remote_code=False, token=False)
     model = AutoModelForSpeechSeq2Seq.from_pretrained(
         MODEL_ID, revision=REVISION, trust_remote_code=False, token=False,
-        dtype=torch.float32,
-    ).to("cpu").eval()
+        dtype=getattr(torch, args.dtype),
+    ).to(args.device).eval()
     inputs = processor.apply_transcription_request(
         audio=audio, language=languages, keywords=args.keywords, prompt=args.prompt,
         processor_kwargs={"return_tensors": "pt", "audio_kwargs": {"sampling_rate": SAMPLE_RATE},
                           "text_kwargs": {"padding": True}},
-    )
+    ).to(args.device)
     with torch.inference_mode():
         generated = model.generate(**inputs, max_new_tokens=args.max_new_tokens, do_sample=False)
     new_tokens = generated[:, inputs.input_ids.shape[1]:]
@@ -93,7 +110,8 @@ def main():
         results.append({"file": str(path), "language": language, "text": text,
                         "reached_eos": any(token in eos_ids for token in tokens)})
     print(json.dumps({"model": MODEL_ID, "revision": REVISION, "transformers": transformers.__version__,
-                      "device": "cpu", "results": results}, ensure_ascii=False, indent=2))
+                      "torch": torch.__version__, "device": str(next(model.parameters()).device),
+                      "dtype": str(model.dtype), "results": results}, ensure_ascii=False, indent=2))
     if not all(row["reached_eos"] and row["text"].strip() for row in results):
         raise SystemExit("Incomplete generation: inspect empty text or missing EOS; do not treat this as a complete transcript")
 
diff --git a/tests/test_transformers_quickstart.py b/tests/test_transformers_quickstart.py
index 74c130e5..f04f8fb8 100644
--- a/tests/test_transformers_quickstart.py
+++ b/tests/test_transformers_quickstart.py
@@ -16,6 +16,34 @@
 
 
 class NativeExampleTests(unittest.TestCase):
+    def test_runtime_choices_fail_closed_without_silent_cpu_fallback(self):
+        validate = getattr(native, 'validate_runtime', None)
+        self.assertTrue(callable(validate), 'Explicit runtime validation is missing')
+        validate('cpu', 'float32')
+        validate('cuda', 'float32', cuda_available=True)
+        validate('cuda', 'bfloat16', cuda_available=True, bf16_supported=True)
+        for device, dtype, available, bf16 in [
+            ('cuda', 'float32', False, False),
+            ('cuda', 'bfloat16', True, False),
+            ('cpu', 'bfloat16', False, False),
+            ('auto', 'float32', False, False),
+            ('cuda', 'float16', True, True),
+        ]:
+            with self.subTest(device=device, dtype=dtype), self.assertRaises(ValueError):
+                validate(device, dtype, cuda_available=available, bf16_supported=bf16)
+
+    def test_gpu_recipe_keeps_cpu_requirements_separate(self):
+        gpu = ROOT / 'examples/transformers/requirements-gpu.txt'
+        self.assertTrue(gpu.is_file())
+        requirements = gpu.read_text()
+        self.assertIn('torch==2.11.0+cu128', requirements)
+        self.assertIn('torchaudio==2.11.0+cu128', requirements)
+        self.assertIn('transformers==5.17.0', requirements)
+        self.assertIn('torch==2.10.0', (ROOT / 'examples/transformers/requirements.txt').read_text())
+        guide = (ROOT / 'examples/transformers/README.md').read_text()
+        for marker in ('--device cuda --dtype bfloat16', 'requirements-gpu.txt', 'H100', '550.127.08'):
+            self.assertIn(marker, guide)
+
     def test_fixed_native_artifact(self):
         self.assertEqual(native.MODEL_ID, "FunAudioLLM/Fun-ASR-Nano-2512-hf")
         self.assertEqual(native.REVISION, "d93b302ee7fd505e1b3576120fc142fc6f7820e1")