↓ 1 callersFunctionprepare_output(
B,
S,
H,
DV,
chunk_size,
d_type,
)
examples/kda/chunk_bwd_gla_dA.py:27
↓ 1 callersFunctionprepare_output(
B,
S,
H,
DV,
chunk_size,
output_dtype,
)
examples/kda/chunk_bwd_dv.py:31
↓ 1 callersFunctionprepare_wy_repr_bwd(
k: torch.Tensor,
v: torch.Tensor,
beta: torch.Tensor,
gk: torch.Tensor,
A: torch.Tensor,
examples/kda/FLA_KDA/fla_wy_fast.py:257
↓ 1 callersFunctionpreprocess(
O,
dO,
block_ND=32,
num_stages=5,
dtype=T.bfloat16,
accum_dtype=T.float32,
)
examples/dsa_sparse_finetune/sparse_mla_bwd.py:11
↓ 1 callersFunctionref_attention Compute reference attention output and weights. Args: q: [b, q_heads, head_size] k, v: [b, kv_heads, max_seqlen, head_size]
examples/flash_decoding/example_gqa_decode_varlen_logits.py:133
↓ 1 callersFunctionref_pool_mqa_fp8Spec: for each (m, n), logits[m, n] = sum_h ReLU(q[m,h] . k[n] * k_scale[n]) * w[m,h]. Computes the full dense [M, Nb] grid — caller is responsibl
examples/dsa_hisa/pool_mqa_fp8.py:139
↓ 1 callersFunctionref_program(A, B, BlockMask, block_M, block_N, block_K)
examples/blocksparse_gemm/example_blocksparse_gemm.py:41
↓ 1 callersFunctionref_program(Q, K, V, padding_mask, is_causal, groups=1)
examples/flash_attention/example_gqa_bwd_tma_reduce_varlen.py:593
↓ 1 callersFunctionref_program(q, q_pe, kv, k_pe, glse, Output_partial)
examples/deepseek_mla/amd/benchmark_mla_decode_amd_tilelang.py:205
↓ 1 callersFunctionref_program_torch(query, key, value, block_indices, cache_seqlens, max_cache_seqlen, num_blocks, block_size)
examples/blocksparse_attention/example_tilelang_sparse_gqa_decode_varlen_indice.py:267
↓ 1 callersFunctionref_program_torch(query, key, value, block_mask, cache_seqlens, max_cache_seqlen, num_blocks, block_size)
examples/blocksparse_attention/example_triton_sparse_gqa_decode_varlen_mask.py:271
↓ 1 callersFunctionref_program_torch(query, key, value, block_mask, cache_seqlens, max_cache_seqlen, num_blocks, block_size)
examples/blocksparse_attention/example_tilelang_sparse_gqa_decode_varlen_mask.py:257
↓ 1 callersFunctionref_program_torch(query, key, value, block_indices, cache_seqlens, max_cache_seqlen, num_blocks, block_size)
examples/blocksparse_attention/example_triton_sparse_gqa_decode_varlen_indice.py:275
↓ 1 callersFunctionref_sparse_mla_bwd_interface(q, kv, o, do, indices, lse, sm_scale=None, is_casual=True)
examples/deepseek_v32/sparse_mla_bwd.py:270
↓ 1 callersFunctionref_sparse_mla_bwd_interface(q, kv, o, do, indices, lse, offsets, sm_scale=None, is_casual=True)
examples/dsa_sparse_finetune/sparse_mla_bwd.py:281
↓ 1 callersFunctionref_sparse_mla_fwd_interface(q, kv, indices, q_start_index_s, kv_stride=4, sm_scale=None, is_casual=True)
examples/deepseek_v32/sparse_mla_fwd_pipelined.py:382
↓ 1 callersFunctionref_sparse_mla_fwd_interface(q, kv, indices, q_start_index_s, kv_stride=1, sm_scale=None, is_casual=True)
examples/deepseek_v32/sparse_mla_fwd_seesaw.py:551