[CI][ROCm] Shorten the blocking CosyVoice test path (#7748)

Signed-off-by: andyluo7 <andy.luo@amd.com>

.buildkite/amd/test-amd-nightly.yml

@@ -64,3 +64,21 @@ steps:
         commands:
           - export VLLM_ROCM_USE_AITER=1
           - pytest -s -v tests/e2e/online_serving/test_qwen3_omni_expansion.py::test_text_to_text_audio_001 -m "full_model and rocm and MI325 and cards_2" --run-level "full_model"
+
+  - group: ":speaker: CosyVoice3 MI300 Nightly"
+    steps:
+      # Exercise all sync/async and Chinese/English cases with the platform's
+      # default attention backend. On ROCm this retains end-to-end AITER
+      # coverage without putting its cold JIT build on the READY critical path.
+      - label: "CosyVoice3-TTS E2E Test"
+        agent_pool: mi300_1
+        depends_on: amd-build
+        mirror_hardwares: [amdproduction]
+        grade: NonBlocking
+        timeout_in_minutes: 90
+        retry:
+          automatic:
+            - exit_status: 134
+              limit: 1
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py -m "slow" --run-level "core_model"

.buildkite/amd/test-amd-ready.yml

@@ -328,7 +328,10 @@ steps:
         - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
         - timeout 30m pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m "core_model" --run-level "core_model"
 
-    - label: "CosyVoice3-TTS E2E Test"
+    # Keep a fast model/server integration check in the blocking lane. The
+    # full default-backend suite remains in the nightly pipeline because a
+    # cold ROCm AITER MHA build can add more than 25 minutes to this job.
+    - label: "CosyVoice3-TTS E2E Smoke (SDPA)"
       agent_pool: mi300_1
       depends_on: amd-build
       mirror_hardwares: [amdproduction]
@@ -340,7 +343,8 @@ steps:
           - exit_status: 134
             limit: 1
       commands:
-        - pytest -s -v tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py -m "slow" --run-level "core_model"
+        - export DIFFUSION_ATTENTION_BACKEND=TORCH_SDPA
+        - pytest -s -v tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py::test_voice_clone_zh_002 -m "slow" --run-level "core_model"
 
 # TODO: Bagel test on ROCm is very unstable. @tjtanaa
 # Need to debug before reneable numerical changes across large PRs

tests/buildkite/test_amd_pipeline.py

@@ -67,11 +67,31 @@ def test_ready_diffusion_cpu_suite_is_sharded() -> None:
     assert "--shard-id=$$BUILDKITE_PARALLEL_JOB" in pytest_command
 
 
-def test_cosyvoice_gpu_abort_gets_one_fresh_job_retry() -> None:
-    step = _find_step("CosyVoice3-TTS E2E Test", AMD_READY_PIPELINE)
+def test_cosyvoice_ready_smoke_uses_sdpa() -> None:
+    step = _find_step("CosyVoice3-TTS E2E Smoke (SDPA)", AMD_READY_PIPELINE)
 
+    assert step["grade"] == "Blocking"
     assert step["retry"] == {"automatic": [{"exit_status": 134, "limit": 1}]}
+    assert "export DIFFUSION_ATTENTION_BACKEND=TORCH_SDPA" in step["commands"]
 
+    pytest_command = next(command for command in step["commands"] if "pytest" in command)
+    assert "tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py::test_voice_clone_zh_002" in pytest_command
+
+
+def test_cosyvoice_full_default_backend_suite_runs_nightly() -> None:
+    step = _find_step("CosyVoice3-TTS E2E Test", AMD_NIGHTLY_PIPELINE)
+
+    assert step["grade"] == "NonBlocking"
+    assert step["timeout_in_minutes"] == 90
+    assert step["retry"] == {"automatic": [{"exit_status": 134, "limit": 1}]}
+    assert all("DIFFUSION_ATTENTION_BACKEND" not in command for command in step["commands"])
+
+    pytest_command = next(command for command in step["commands"] if "pytest" in command)
+    assert "tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py" in pytest_command
+    assert "::" not in pytest_command
+
+
+def test_amd_template_preserves_step_retry_policy() -> None:
     template = AMD_TEMPLATE.read_text(encoding="utf-8")
     # Both grouped and top-level AMD steps must preserve an explicit retry
     # policy when the source suite is rendered into the uploaded pipeline.

tests/model_executor/models/cosyvoice3/test_cosyvoice3_components.py

@@ -260,6 +260,37 @@ class TestDiTAttention:
         assert attention.to_k.out_features == 512
         assert attention.to_v.out_features == 512
 
+    def test_embedded_dit_honors_diffusion_attention_backend(self, monkeypatch):
+        """The LLM-hosted code2wav DiT must consume the diffusion env override."""
+        from vllm_omni.diffusion.config import get_current_diffusion_config_or_none
+        from vllm_omni.model_executor.models.cosyvoice3.cosyvoice3_code2wav import _build_dit_estimator
+
+        monkeypatch.setenv("DIFFUSION_ATTENTION_BACKEND", "TORCH_SDPA")
+
+        estimator = _build_dit_estimator(
+            {
+                "dim": 32,
+                "depth": 1,
+                "heads": 2,
+                "dim_head": 16,
+                "dropout": 0.0,
+                "ff_mult": 2,
+                "mel_dim": 8,
+                "mu_dim": 8,
+                "spk_dim": 4,
+                "out_channels": 8,
+                "static_chunk_size": 4,
+                "num_decoding_left_chunks": 1,
+            }
+        )
+
+        attention = estimator.transformer_blocks[0].attn.attn
+        assert attention.backend_pref == "TORCH_SDPA"
+        assert attention.backend_explicit is True
+        assert attention.attn_backend is not None
+        assert attention.attn_backend.get_name() == "SDPA"
+        assert get_current_diffusion_config_or_none() is None
+
     @pytest.mark.core_model
     @pytest.mark.cpu
     def test_forward_passes_2d_mask_without_casting_qkv(self):

tests/model_executor/models/cosyvoice3/test_cosyvoice3_incremental_hift.py

@@ -39,7 +39,10 @@ CHUNK_LEN = 24
 TOTAL_MEL = 96
 SPM = 480  # samples per mel frame for the small test HiFT (8*5*3*4), matching the real model
 
-LONG_TOTAL_MEL = 400  # long enough that the 64-frame window actually truncates
+WINDOW_LEN = 64
+# Six chunks exercise three updates after history exceeds the bounded window,
+# without spending minutes recomputing an unnecessarily long synthetic signal.
+LONG_TOTAL_MEL = 144
 PCM16_LSB = 1.0 / 32767.0  # smallest representable PCM16 difference
 
 # Worst observed deviation is 1.97e-7 (~2x float32 eps); atol=1e-6 keeps 5x headroom.
@@ -189,16 +192,19 @@ def test_incremental_hift_bounded_window_is_close(config):
 def test_incremental_hift_matches_streaming_reference_within_tolerance(config):
     """Windowed output matches the full-cumulative reference within float32 rounding.
 
-    The load-bearing correctness test: 400 mel frames far exceeds the 64-frame
-    window, so truncation genuinely exercises the phase/noise carry. Float64
-    agrees to ~1e-16, confirming the residual here is rounding, not truncation.
+    The load-bearing correctness test crosses the 64-frame window repeatedly,
+    so truncation genuinely exercises the phase/noise carry. Float64 agrees to
+    ~1e-16, confirming the residual here is rounding, not truncation.
     """
     hift = _make_hift(config)
-    model = _make_model(hift, window_len=64)  # the real _hift_window_len
+    model = _make_model(hift, window_len=WINDOW_LEN)  # the real _hift_window_len
     chunks = _chunks(total_mel=LONG_TOTAL_MEL)
 
     trim = int(hift.f0_predictor.condnet[0].causal_padding)
-    assert LONG_TOTAL_MEL > 64 + trim + CHUNK_LEN  # window must actually truncate
+    truncating_steps = sum(
+        history_len > WINDOW_LEN + trim for history_len in range(CHUNK_LEN, LONG_TOTAL_MEL, CHUNK_LEN)
+    )
+    assert truncating_steps >= 3  # exercise repeated truncation, not only the boundary
 
     full = _full_reference(model, chunks)
     incr = _incremental(model, chunks)

vllm_omni/model_executor/models/cosyvoice3/cosyvoice3_code2wav.py

@@ -10,7 +10,8 @@ This module contains the code2wav (token-to-waveform) stage which uses:
 """
 
 from collections import Counter
-from typing import TypedDict, cast
+from collections.abc import Mapping
+from typing import Any, TypedDict, cast
 
 import numpy as np
 import torch
@@ -18,6 +19,8 @@ import torch.nn as nn
 from omegaconf import DictConfig
 from vllm.logger import init_logger
 
+from vllm_omni.diffusion.config import get_current_diffusion_config_or_none, set_current_diffusion_config
+from vllm_omni.diffusion.data import OmniDiffusionConfig
 from vllm_omni.diffusion.models.cosyvoice3_audio.cosyvoice3_dit import DiT
 from vllm_omni.model_executor.models.cosyvoice3.code2wav_core.cfm import (
     CausalConditionalCFM,
@@ -37,6 +40,23 @@ from vllm_omni.transformers_utils.configs.cosyvoice3 import CosyVoice3Config
 logger = init_logger(__name__)
 
 
+def _build_dit_estimator(estimator_config: Mapping[str, Any]) -> DiT:
+    """Build CosyVoice's embedded DiT with diffusion backend configuration.
+
+    The code2wav model runs inside an LLM-generation worker, so it does not go
+    through the diffusion model loader that normally installs an
+    ``OmniDiffusionConfig`` during model construction. Install a local default
+    config when needed so ``DIFFUSION_ATTENTION_BACKEND`` and the platform
+    default are resolved by the same path as standalone diffusion models.
+    """
+    if get_current_diffusion_config_or_none() is not None:
+        return DiT(**estimator_config)
+
+    diffusion_config = OmniDiffusionConfig()
+    with set_current_diffusion_config(diffusion_config):
+        return DiT(**estimator_config)
+
+
 class StreamingFlowItem(TypedDict, total=False):
     """One entry of the batched-streaming item list passed to forward_streaming_batch."""
 
@@ -71,7 +91,7 @@ class CosyVoice3Code2Wav(nn.Module):
         cfm_params = DictConfig(decoder_cfg["cfm_params"])
 
         # DiT estimator using diffusion attention (Flash/Sage/SDPA backends)
-        estimator = DiT(**decoder_cfg["estimator"])
+        estimator = _build_dit_estimator(decoder_cfg["estimator"])
 
         decoder = CausalConditionalCFM(
             in_channels=decoder_cfg["in_channels"],