MiniMax-M2.5 - Support dp attention, dp reduce scatter, FP4 all gather, AR fusion in prepare_attn (#20067)

This commit is contained in:
Trevor Morris
2026-04-10 12:41:27 -07:00
committed by GitHub
parent a937ec31be
commit 7dbd0dd9f0
3 changed files with 39 additions and 6 deletions
@@ -29,6 +29,10 @@ class TestMiniMaxM25(unittest.TestCase):
"--mem-fraction-static=0.85",
"--reasoning-parser=minimax-append-think",
]
dp_attn_args = base_args + [
"--enable-dp-attention",
"--dp=8",
]
variants = [
ModelLaunchSettings(
@@ -37,6 +41,12 @@ class TestMiniMaxM25(unittest.TestCase):
extra_args=base_args,
variant="TP8+EP8",
),
ModelLaunchSettings(
MINIMAX_M25_MODEL_PATH,
tp_size=8,
extra_args=dp_attn_args,
variant="TP8+DP8+EP8+DPAttn",
),
]
run_combined_tests(