Fix input dtype mismatches; drop MIGraphXExecutionProvider

- Drop MIGraphXExecutionProvider — symbol mismatch with apt migraphx,
  ROCMExecutionProvider handles GPU execution fine without it
- Add _ort_type_to_np() helper to read expected dtypes from session metadata
- prepare_voice: cast audio to session's declared input dtype (float32, not float16)
- _generate: read embed_dtype and kv_dtype from LM session metadata

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

engine.py

@@ -43,8 +43,8 @@ def _get_providers(device: str) -> list:
     import onnxruntime as ort
     available = set(ort.get_available_providers())
     providers = []
-    if "MIGraphXExecutionProvider" in available:
-        providers.append("MIGraphXExecutionProvider")
+    # MIGraphXExecutionProvider excluded — symbol mismatch between onnxruntime-rocm
+    # and apt migraphx; ROCMExecutionProvider covers GPU execution adequately.
     if "ROCMExecutionProvider" in available:
         providers.append(("ROCMExecutionProvider", {"device_id": 0}))
     providers.append("CPUExecutionProvider")
@@ -52,6 +52,22 @@ def _get_providers(device: str) -> list:
     return providers
 
 
+def _ort_type_to_np(ort_type: str):
+    """Convert ORT type string (e.g. 'tensor(float16)') to numpy dtype."""
+    mapping = {
+        "tensor(float)": np.float32,
+        "tensor(float16)": np.float16,
+        "tensor(double)": np.float64,
+        "tensor(int64)": np.int64,
+        "tensor(int32)": np.int32,
+        "tensor(int16)": np.int16,
+        "tensor(int8)": np.int8,
+        "tensor(uint8)": np.uint8,
+        "tensor(bool)": np.bool_,
+    }
+    return mapping.get(ort_type, np.float32)
+
+
 def load_model() -> bool:
     global _sessions, _tokenizer
 
@@ -111,10 +127,13 @@ def prepare_voice(audio_prompt_path: str) -> None:
     logger.info(f"Preparing voice conditionals for '{audio_prompt_path}'")
     try:
         audio, _ = librosa.load(audio_prompt_path, sr=SAMPLE_RATE, mono=True)
-        audio = audio[np.newaxis, :].astype(np.float16)  # [1, T]
 
         session = _sessions["speech_encoder"]
-        input_name = session.get_inputs()[0].name
+        input_meta = session.get_inputs()[0]
+        input_dtype = _ort_type_to_np(input_meta.type)
+        audio = audio[np.newaxis, :].astype(input_dtype)  # [1, T]
+
+        input_name = input_meta.name
         output_names = [o.name for o in session.get_outputs()]
         outputs = session.run(None, {input_name: audio})
 
@@ -171,12 +190,18 @@ def _generate(input_ids: np.ndarray, cond: Dict[str, np.ndarray]) -> np.ndarray:
 
     embed_input_name = embed_sess.get_inputs()[0].name
 
-    # Discover KV cache slot names from session metadata
+    # Discover KV cache slot names and dtypes from session metadata
     past_names = [i.name for i in lm_sess.get_inputs() if "past_key_values" in i.name]
     present_names = [o.name for o in lm_sess.get_outputs() if "present" in o.name]
     lm_out_names = [o.name for o in lm_sess.get_outputs()]
 
-    kv_dtype = np.float16
+    # Read expected embed dtype from language_model's inputs_embeds slot
+    embeds_meta = next((i for i in lm_sess.get_inputs() if "embeds" in i.name), None)
+    embed_dtype = _ort_type_to_np(embeds_meta.type) if embeds_meta else np.float32
+
+    # Read KV cache dtype from first past_key_values slot
+    kv_meta = next((i for i in lm_sess.get_inputs() if "past_key_values" in i.name), None)
+    kv_dtype = _ort_type_to_np(kv_meta.type) if kv_meta else np.float16
 
     # Embed full text sequence
     text_embeds = embed_sess.run(None, {embed_input_name: input_ids})[0]  # [1, seq, hidden]
@@ -201,7 +226,7 @@ def _generate(input_ids: np.ndarray, cond: Dict[str, np.ndarray]) -> np.ndarray:
 
     for _ in range(MAX_NEW_TOKENS):
         feed = {
-            "inputs_embeds": inputs_embeds.astype(kv_dtype),
+            "inputs_embeds": inputs_embeds.astype(embed_dtype),
             "attention_mask": attention_mask,
             "position_ids": position_ids,
         }