[OCP MX ] Add back emulation to available OCP MX backends list (#46629)
Signed-off-by: Felix Marty <Felix.Marty@amd.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>
fxmarty-amd committed
Jun 28, 2026 at 19:43 UTC
03c6d01c3028cd567ddddc54e1d8414a24dbe501
3 files changed
+1
-11
tests/models/quantization/test_gpt_oss.py
-5
@@ -104,11 +104,6 @@ def test_gpt_oss_attention_quantization(
104
105
model_args = EvaluationConfig(model_name).get_model_args(tp_size)
106
107
- # Emulation backend on MI300, MI250 is opt-in
108
- # following https://github.com/vllm-project/vllm/pull/45896
109
- if not on_gfx950():
110
- model_args["moe_backend"] = "emulation"
111
-
107
extra_run_kwargs = {
108
"gen_kwargs": {"max_gen_toks": 8000},
109
"apply_chat_template": True,
tests/quantization/test_quark.py
-4
@@ -221,10 +221,6 @@ class AccuracyTestConfig:
221
if model_max_len is not None:
222
model_args["max_model_len"] = model_max_len
223
224
- # Emulation backend on MI300, MI250 is opt-in following https://github.com/vllm-project/vllm/pull/45896
225
- if not on_gfx950():
226
- model_args["moe_backend"] = "emulation"
227
-
224
return model_args
225
226
vllm/model_executor/layers/fused_moe/oracle/mxfp4.py
+1
-2
@@ -319,6 +319,7 @@ def _get_priority_backends_for_gpt_oss() -> list[Mxfp4MoeBackend]:
319
Mxfp4MoeBackend.MARLIN,
320
Mxfp4MoeBackend.BATCHED_MARLIN,
321
Mxfp4MoeBackend.XPU,
322
+ Mxfp4MoeBackend.EMULATION,
323
]
324
return _AVAILABLE_BACKENDS
325
@@ -554,8 +555,6 @@ def select_mxfp4_moe_backend(
555
f"weight_key=kMxfp4Static, activation_key={activation_key}. "
556
"Native backends require specific hardware. "
557
"Set `VLLM_LOGGING_LEVEL=DEBUG` to see detailed unsupported reasons. "
557
- "To use the emulation backend for research/debugging, pass "
558
- "--moe-backend emulation."
558
)
559
560
return Mxfp4MoeBackend.NONE, None