[Model] MiniCPM-o 4.5: enable fused CFM body by default on CUDA (#8557)

Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>

docs/design/minicpm_o45_mrv2_performance.md

@@ -23,3 +23,14 @@ performance or interruption correctness.
 The shared asynchronous output snapshot and batched Talker preprocessing also
 apply to V1 when async chunking/scheduling is enabled. Default V1 behavior must
 therefore be included in end-to-end regression validation.
+
+CFM graph execution defaults to the fused DiT body on CUDA when TF32 is
+allowed. NVIDIA SM80+ uses tiled TF32 attention; older CUDA devices use the
+fused body's SDPA fallback. Unsupported DiT layouts or non-FP32 attention
+caches retain the original body. Other platforms retain their existing default.
+Set `cfm_fused_body: false` in connector `extra` to restore the original body.
+Disabling TF32 with `token2wav_allow_tf32: false` or
+`MINICPMO_CODE2WAV_TF32=off` also disables the default fusion choice; an explicit
+`cfm_fused_body: true` overrides that choice and may use TF32 attention.
+Slot pooling and row-offset merging remain opt-in. Fusion changes floating-point
+rounding and does not promise identical waveforms.

tests/model_executor/models/minicpmo_4_5/test_code2wav_batching.py

@@ -172,6 +172,36 @@ def _config(minimum: int = 1, initial: int = 0, *, runtime_prompt_cache_size: in
     )
 
 
+@pytest.mark.parametrize(
+    ("cuda", "extra", "env", "expected"),
+    [
+        (True, {}, None, True),
+        (False, {}, None, False),
+        (True, {"cfm_fused_body": False}, None, False),
+        (True, {"token2wav_allow_tf32": False}, None, False),
+        (True, {"code2wav_allow_tf32": False}, None, False),
+        (True, {}, "off", False),
+        (True, {"token2wav_allow_tf32": False}, "tf32", True),
+        (True, {"cfm_fused_body": True, "token2wav_allow_tf32": False}, None, True),
+    ],
+)
+def test_cfm_fused_body_default_respects_platform_and_precision(monkeypatch, cuda, extra, env, expected):
+    from vllm_omni.platforms import current_omni_platform
+
+    monkeypatch.setattr(current_omni_platform, "is_cuda", lambda: cuda)
+    monkeypatch.delenv("MINICPMO_CODE2WAV_TF32", raising=False)
+    if env is not None:
+        monkeypatch.setenv("MINICPMO_CODE2WAV_TF32", env)
+    config = _config()
+    config.model_config.stage_connector_config["extra"].update(extra)
+
+    model = MiniCPMO45Code2Wav(vllm_config=config)
+
+    assert model._cfm_graph_config["fused_body"] is expected
+    assert model._cfm_graph_config["slot_pool"] is False
+    assert model._cfm_graph_config["row_offset_merge"] is False
+
+
 @pytest.mark.parametrize(
     ("extra", "max_num_seqs", "micro"),
     [

vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_code2wav.py

@@ -288,6 +288,8 @@ class MiniCPMO45Code2Wav(nn.Module):
         self._runtime_prompt_dir = tempfile.TemporaryDirectory(
             prefix="minicpmo45-runtime-prompts-",
         )
+        from vllm_omni.platforms import current_omni_platform
+
         extra = self._extra_config()
         self._runtime_prompt_cache_size = int(extra.get("token2wav_runtime_prompt_cache_size", 4))
         if self._runtime_prompt_cache_size < 0:
@@ -342,7 +344,11 @@ class MiniCPMO45Code2Wav(nn.Module):
             "max_graph_batch": max_graph_batch,
             "micro_batch_size": micro_batch_size,
             "pad_max_rows": extra.get("whole_euler_pad_max_rows"),
-            "fused_body": bool(extra.get("cfm_fused_body", False)),
+            # The fused body uses tiled TF32 attention on NVIDIA SM80+;
+            # older CUDA devices retain its SDPA fallback. Respect TF32 opt-out.
+            "fused_body": bool(
+                extra.get("cfm_fused_body", current_omni_platform.is_cuda() and _tf32_mode(extra) != "off")
+            ),
             "slot_pool": bool(extra.get("cfm_slot_pool", False)),
             "row_offset_merge": extra.get("cfm_row_offset_merge", False) is True,
         }