Revert "[Perf][Qwen3-TTS] Fuse QKV and gate_up projections in code predictor …"

This reverts commit 6fdc4ca62d.
This commit is contained in:
Gao Han authored and GitHub committed 2026-08-04 22:25:12 +08:00
1 parent 6fdc4ca62d
commit 5cf2681aeb
8 files changed
+36 -638

No files matched your search

-15
View File
@@ -414,21 +414,6 @@ def test_qwen3_tts_code_predictor_forward_uses_projected_embedding_and_sampling(
assert sample_calls == [{0: generator}]
def test_310p_attention_forward_uses_fused_qkv(monkeypatch: pytest.MonkeyPatch) -> None:
"""The fused code predictor dropped q_proj/k_proj/v_proj in favor of a single
``qkv_proj`` (+ ``_split_qkv``). The 310P attention override must follow
suit, otherwise its NPU flash-attention path raises AttributeError on 310P.
The NPU path itself cannot run in CPU CI, so guard the contract statically."""
import inspect
module, _ = _load_qwen3_tts_patch(monkeypatch)
source = inspect.getsource(module._Qwen3CodePredictorAttention310P.forward)
assert "qkv_proj" in source and "_split_qkv" in source
assert "self.q_proj" not in source
assert "self.k_proj" not in source
assert "self.v_proj" not in source
def test_qwen3_tts_talker_patch_uses_fp16_runtime_dtype(monkeypatch: pytest.MonkeyPatch) -> None:
module, _ = _load_qwen3_tts_patch(monkeypatch)
talker = module._Qwen3TTSTalker310P(vllm_config=object())