[Bugfix][CPU] Fix DeepSeek-R1 (FP8 MLA + MoE) correctness on CPU backend (#56773)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
This commit is contained in:
@@ -261,6 +261,8 @@ class CPUFp8BlockScaledMMKernel(Fp8BlockScaledMMLinearKernel):
|
||||
return True, None
|
||||
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
if getattr(layer, "_cpu_skip_gemm_dispatch", False):
|
||||
return
|
||||
# Skip the base class process (FP8 padding / fnuz normalization)
|
||||
# which is GPU-oriented. Instead, VNNI-prepack weights for AMX.
|
||||
params = self._get_layer_params(layer)
|
||||
|
||||
Reference in New Issue
Block a user