MCPcopy Create free account
hub / github.com/tile-ai/tilelang / kernel

Function kernel

testing/python/amd/test_tilelang_hip_codegen.py:58–72  ·  view source on GitHub ↗
(
            logits: T.Tensor((n_tok, n_exp), T.float32),
            out_max: T.Tensor((n_tok,), T.float32),
            out_sum: T.Tensor((n_tok,), T.float32),
        )

Source from the content-addressed store, hash-verified

56 def gate_reduce(n_tok: int, n_exp: int):
57 @T.prim_func
58 def kernel(
59 logits: T.Tensor((n_tok, n_exp), T.float32),
60 out_max: T.Tensor((n_tok,), T.float32),
61 out_sum: T.Tensor((n_tok,), T.float32),
62 ) -> None:
63 with T.Kernel(n_tok, threads=32) as pid:
64 lf = T.alloc_fragment(n_exp, T.float32)
65 T.copy(logits[pid, 0], lf)
66 mx = T.alloc_fragment(1, T.float32)
67 T.reduce_max(lf, mx, dim=0)
68 sm = T.alloc_fragment(1, T.float32)
69 T.reduce_sum(lf, sm, dim=0)
70 if T.get_thread_binding() == 0:
71 out_max[pid] = mx[0]
72 out_sum[pid] = sm[0]
73
74 return kernel
75

Calls 1

get_thread_bindingMethod · 0.45

Tested by

no test coverage detected

Used in the wild real call sites across dependent graphs

searching dependent graphs…