diff --git a/tests/models/quantization/test_gpt_oss.py b/tests/models/quantization/test_gpt_oss.py index 1f5e48cb0c..783f1773d2 100644 --- a/tests/models/quantization/test_gpt_oss.py +++ b/tests/models/quantization/test_gpt_oss.py @@ -104,11 +104,6 @@ def test_gpt_oss_attention_quantization( model_args = EvaluationConfig(model_name).get_model_args(tp_size) - # Emulation backend on MI300, MI250 is opt-in - # following https://github.com/vllm-project/vllm/pull/45896 - if not on_gfx950(): - model_args["moe_backend"] = "emulation" - extra_run_kwargs = { "gen_kwargs": {"max_gen_toks": 8000}, "apply_chat_template": True, diff --git a/tests/quantization/test_quark.py b/tests/quantization/test_quark.py index 38b66552f4..c1cb18f8e2 100644 --- a/tests/quantization/test_quark.py +++ b/tests/quantization/test_quark.py @@ -221,10 +221,6 @@ class AccuracyTestConfig: if model_max_len is not None: model_args["max_model_len"] = model_max_len - # Emulation backend on MI300, MI250 is opt-in following https://github.com/vllm-project/vllm/pull/45896 - if not on_gfx950(): - model_args["moe_backend"] = "emulation" - return model_args diff --git a/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py b/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py index 56a7a6482d..b1b41ded11 100644 --- a/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py +++ b/vllm/model_executor/layers/fused_moe/oracle/mxfp4.py @@ -319,6 +319,7 @@ def _get_priority_backends_for_gpt_oss() -> list[Mxfp4MoeBackend]: Mxfp4MoeBackend.MARLIN, Mxfp4MoeBackend.BATCHED_MARLIN, Mxfp4MoeBackend.XPU, + Mxfp4MoeBackend.EMULATION, ] return _AVAILABLE_BACKENDS @@ -554,8 +555,6 @@ def select_mxfp4_moe_backend( f"weight_key=kMxfp4Static, activation_key={activation_key}. " "Native backends require specific hardware. " "Set `VLLM_LOGGING_LEVEL=DEBUG` to see detailed unsupported reasons. " - "To use the emulation backend for research/debugging, pass " - "--moe-backend emulation." ) return Mxfp4MoeBackend.NONE, None