fix(deepgemm): add nsa init in deepgemm init
This commit is contained in:
@@ -2529,6 +2529,14 @@ class ModelRunner(ModelRunnerKVCacheMixin):
|
||||
k,
|
||||
1, # num_groups=1 for standard GEMM
|
||||
)
|
||||
if module.weight.dim() == 2 and module.weight.dtype == torch.bfloat16:
|
||||
n, k = module.weight.shape
|
||||
_maybe_compile_deep_gemm_one_type_all(
|
||||
DeepGemmKernelType.GEMM_NT_BF16BF16F32,
|
||||
n,
|
||||
k,
|
||||
1,
|
||||
)
|
||||
# deep_gemm.preload_kernels()
|
||||
|
||||
def init_threads_binding(self):
|
||||
|
||||
Reference in New Issue
Block a user