A high-throughput and memory-efficient inference and serving engine for LLMs
[bugfix] skip conch kernel for g_idx reordering (#45072)
Signed-off-by: Divakar Verma <divakar.verma@amd.com>
Divakar Verma committed
Jun 10, 2026 at 11:04 UTC
166d14e9bf18e57e1e6833f9dd645ad6c4928f18
1 file changed
+6
vllm/model_executor/kernels/linear/mixed_precision/conch.py
+6
@@ -43,6 +43,12 @@ class ConchLinearKernel(MPLinearKernel):
43
)
44
return False, error_msg
45
46
+ if c.has_g_idx:
47
+ return (
48
+ False,
49
+ "Activation reordering (g_idx) is not supported by ConchLinearKernel",
50
+ )
51
+
52
if find_spec("conch") is None:
53
error_msg = (
54
"conch-triton-kernels is not installed, please "