disable mlp fusion of fp6 on mtl (#11059)

2024-05-17 10:10:16 +08:00 · 2024-05-17 10:10:16 +08:00 · 3a72e5df8c
commit 3a72e5df8c
parent 192ae35012
1 changed files with 4 additions and 0 deletions
--- a/python/llm/src/ipex_llm/transformers/models/utils.py
+++ b/python/llm/src/ipex_llm/transformers/models/utils.py
@ -358,6 +358,10 @@ def mlp_fusion_check(x, qtype, training):
        return False
    if training or x.requires_grad:
        return False
+    if qtype == FP6:
+        device = get_xpu_device_type(x)
+        if device == "mtl":
+            return False
    return True