From 6e2c176e1ffd99382d03bc31a454834ca2532f7f Mon Sep 17 00:00:00 2001 From: Michael Goin Date: Sun, 13 Jul 2025 02:34:40 +0900 Subject: [PATCH] [Bugfix] Restrict Machete to only run on Hopper (#20830) Signed-off-by: mgoin --- .../layers/quantization/kernels/mixed_precision/machete.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/vllm/model_executor/layers/quantization/kernels/mixed_precision/machete.py b/vllm/model_executor/layers/quantization/kernels/mixed_precision/machete.py index 851fd155465d4..ed81b02bc4a10 100644 --- a/vllm/model_executor/layers/quantization/kernels/mixed_precision/machete.py +++ b/vllm/model_executor/layers/quantization/kernels/mixed_precision/machete.py @@ -32,6 +32,9 @@ class MacheteLinearKernel(MPLinearKernel): if not current_platform.is_cuda(): return False, "Machete only supported on CUDA" + if not current_platform.is_device_capability(90): + return False, "Machete requires compute capability of 90 (Hopper)" + if c.has_g_idx and\ c.partition_weight_shape[0] != c.full_weight_shape[0]: return False, "Act reordering currently not supported by Machete, "\