[BugFix] Fix silu kernel (#661)

Co-authored-by: sufubao <[email protected]>

lightllm/models/llama/triton_kernel/silu_and_mul.py

@@ -17,8 +17,8 @@ def _silu_and_mul_kernel(
     BLOCK_M: tl.constexpr,
     BLOCK_N: tl.constexpr,
 ):
-    stride_input_m = stride_input_m.to(tl.int64)
-    stride_output_m = stride_output_m.to(tl.int64)
+    stride_input_m = tl.cast(stride_input_m, dtype=tl.int64)
+    stride_output_m = tl.cast(stride_output_m, dtype=tl.int64)
 
     tid = tl.program_id(0)
     input_m_offsets = tid * BLOCK_M + tl.arange(0, BLOCK_M)
@@ -53,7 +53,7 @@ def _silu_and_mul_kernel(
     )
 
 
-def silu_and_mul_fwd(input, output):
+def silu_and_mul_fwd(input: torch.Tensor, output):
     stride_input_m = input.stride(0)
     stride_input_n = input.stride(1)
     stride_output_m = output.stride(0)
@@ -88,13 +88,13 @@ def torch_silu_and_mul(input: torch.Tensor):
 def test_silu_and_mul(M, N, dtype, device="cuda"):
     # create data
     X = torch.randn((M, N), dtype=dtype, device=device)
-
+    y_tri = torch.empty((M, N // 2), dtype=dtype, device=device)
     # run
-    y_tri = silu_and_mul_fwd(X)
+    silu_and_mul_fwd(X, y_tri)
     y_ref = torch_silu_and_mul(X)
 
     # compare
     print("type:", y_tri.dtype, y_ref.dtype)
     print("max delta:", torch.max(torch.abs(y_tri - y_ref)))
-    assert torch.allclose(y_tri, y_ref, atol=1e-6, rtol=0)
+    assert torch.allclose(y_tri, y_ref, atol=1e-5, rtol=0)
     return