Update rms_layernorm.py
This commit is contained in:
parent
32fc5a59a0
commit
2cd4b8debd
1 changed files with 22 additions and 22 deletions
|
|
@ -138,7 +138,7 @@ class Fast_RMS_Layernorm(torch.autograd.Function):
|
|||
def forward(ctx, X, W, eps : float, gemma : bool = False):
|
||||
shape = X.shape
|
||||
dim : int = shape[-1]
|
||||
X : torch.Tensor = X.view(-1, dim)
|
||||
X = X.view(-1, dim)
|
||||
n_rows : int
|
||||
n_cols : int
|
||||
n_rows, n_cols = X.shape
|
||||
|
|
@ -149,27 +149,27 @@ class Fast_RMS_Layernorm(torch.autograd.Function):
|
|||
Y = torch.empty((n_rows, n_cols), dtype = X.dtype, device = "cuda:0")
|
||||
r = torch.empty(n_rows, dtype = torch.float32, device = "cuda:0")
|
||||
|
||||
# if gemma == False:
|
||||
_rms_layernorm_forward[(n_rows,)](
|
||||
Y, Y.stride(0),
|
||||
X, X.stride(0),
|
||||
W, W.stride(0),
|
||||
r, r.stride(0),
|
||||
n_cols = int(n_cols),
|
||||
eps = float(eps),
|
||||
BLOCK_SIZE = BLOCK_SIZE,
|
||||
num_warps = num_warps,
|
||||
)
|
||||
# else:
|
||||
# _gemma_rms_layernorm_forward[(n_rows,)](
|
||||
# Y, Y.stride(0),
|
||||
# X, X.stride(0),
|
||||
# W, W.stride(0),
|
||||
# r, r.stride(0),
|
||||
# n_cols, eps,
|
||||
# BLOCK_SIZE = BLOCK_SIZE,
|
||||
# num_warps = num_warps,
|
||||
# )
|
||||
if not gemma:
|
||||
_rms_layernorm_forward[(n_rows,)](
|
||||
Y, Y.stride(0),
|
||||
X, X.stride(0),
|
||||
W, W.stride(0),
|
||||
r, r.stride(0),
|
||||
n_cols = int(n_cols),
|
||||
eps = float(eps),
|
||||
BLOCK_SIZE = BLOCK_SIZE,
|
||||
num_warps = 16,
|
||||
)
|
||||
else:
|
||||
_gemma_rms_layernorm_forward[(n_rows,)](
|
||||
Y, Y.stride(0),
|
||||
X, X.stride(0),
|
||||
W, W.stride(0),
|
||||
r, r.stride(0),
|
||||
n_cols, eps,
|
||||
BLOCK_SIZE = BLOCK_SIZE,
|
||||
num_warps = num_warps,
|
||||
)
|
||||
ctx.eps = eps
|
||||
ctx.BLOCK_SIZE = BLOCK_SIZE
|
||||
ctx.num_warps = num_warps
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue