↓ 8 callersFunctionprepare_input(
B,
S,
H,
DK,
DV,
chunk_size,
input_dtype,
output_dtype,
accum_dtype,
examples/gdn/example_chunk_delta_bwd.py:32
↓ 8 callersFunctionunpad_input Arguments: hidden_states: (batch, seqlen, ...) attention_mask: (batch, seqlen), bool / int, 1 means valid and 0 means not valid.
examples/flash_attention/bert_padding.py:94
↓ 7 callersFunctionflashattn(
batch,
heads,
seq_q,
seq_kv,
dim,
window_size=None, # None for full attention
s
examples/attention_sink/example_mha_sink_fwd_bhsd.py:25
↓ 7 callersFunctionget_ldmatrix_offset(
matrix: Literal["A", "B"],
row_idx,
col_idx,
stride,
dtype: Literal["float16", "int8", "
tilelang/cuda/intrinsics/layout/utils.py:20
↓ 6 callersFunction_make_nan_reduce_kernel(reduce_fn, M, N, dtype, threads, *, nan_propagate)
testing/python/language/test_tilelang_language_reduce.py:351
↓ 6 callersFunctionassert_simdgroup_store_correctness(M, N, K, block_M, block_N, block_K, dtype=T.float16, accum_dtype=T.float32, atol=1e-2)
testing/python/metal/test_metal_simdgroup_store.py:40
↓ 6 callersMethodget_tcgen5_mma_metaQuery the FFI for TCGEN5MMA atom metadata (atom_m, atom_n, atom_k, enable_ws, enable_2cta), and record them in `self.meta`.
tilelang/cuda/intrinsics/macro/tcgen05_macro_generator.py:572
↓ 6 callersFunctionkernel(
logits: T.Tensor((n_tok, n_exp), T.float32),
out_max: T.Tensor((n_tok,), T.float32),
testing/python/amd/test_tilelang_hip_codegen.py:58