[AMD][OCP MX][CI] Fix tests to not dispatch on UNFUSED_TRITON backend on MI300, improve w_mxfp4_a_fp8 emulation support (#46142)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
This commit is contained in:
fxmarty-amd
2026-06-23 14:25:27 -04:00
committed by GitHub
parent ceae5bcbda
commit e368415daa
7 changed files with 74 additions and 17 deletions
+12
View File
@@ -30,6 +30,14 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
)
from vllm.platforms import current_platform
if current_platform.is_rocm():
from vllm.platforms.rocm import on_gfx950
else:
def on_gfx950() -> bool:
return False
from .reference_mxfp4 import dq_mxfp4_torch, qdq_mxfp4_torch
# Minimum amd-quark version for MXFP4/OCP_MX tests (single source of truth).
@@ -213,6 +221,10 @@ 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