submission 68960

- import triton
- import triton.language as tl
+ #!POPCORN leaderboard vectoradd_v2
+ #!POPCORN gpus B200 H100 A100 L4
+ from task import input_t, output_t
+ import torch
   
   
- @triton.autotune(
-     configs=[
-         # ------- Tier 1: good for L4 / small tensors -------
-         triton.Config(kwargs={"BLOCK_SIZE": 256}, num_warps=2, num_stages=1),
-         triton.Config(kwargs={"BLOCK_SIZE": 512}, num_warps=4, num_stages=1),
-         # ------- Tier 2: A100 general sweet spot -------
-         triton.Config(kwargs={"BLOCK_SIZE": 1024}, num_warps=4, num_stages=2),
-         triton.Config(kwargs={"BLOCK_SIZE": 2048}, num_warps=8, num_stages=2),
-         # ------- Tier 3: H100 / Blackwell optimized -------
-         #   Large blocks + pipelined loads
-         triton.Config(kwargs={"BLOCK_SIZE": 4096}, num_warps=8, num_stages=3),
-         triton.Config(kwargs={"BLOCK_SIZE": 8192}, num_warps=8, num_stages=4),
-     ],
-     key=["n_elements"],  # autotune per problem size
- )
- @triton.jit
- def _kernel(x_ptr, y_ptr, output_ptr, n_elements, BLOCK_SIZE: tl.constexpr):
-     pid = tl.program_id(axis=0)
-     block_start = pid * BLOCK_SIZE
-     offsets = block_start + tl.arange(0, BLOCK_SIZE)
-     mask = offsets < n_elements
+ def add(x, y):
+     return x + y
   
-     x = tl.load(x_ptr + offsets, mask=mask)
-     y = tl.load(y_ptr + offsets, mask=mask)
   
-     tl.store(output_ptr + offsets, x + y, mask=mask)
+ add_ = torch.compile(add)
   
   
- def custom_kernel(data):
-     x, y, output = data
-     n_elements = output.numel()
-     grid = lambda meta: (triton.cdiv(n_elements, meta["BLOCK_SIZE"]),)
-     _kernel[grid](x, y, output, n_elements)
-     return output
+ def custom_kernel(data: input_t) -> output_t:
+     d_in1, d_in2, d_out = data
+     return add_(d_in1, d_in2)