↓ 2 callersFunctionpreprocess(
O,
dO,
block_ND=32,
num_stages=5,
dtype=T.bfloat16,
accum_dtype=T.float32,
)
examples/deepseek_v32/sparse_mla_bwd.py:9
↓ 2 callersFunctionref_program_fa(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:301
↓ 2 callersFunctionref_program_fa(query, key, value, block_indices, cache_seqlens, max_cache_seqlen, num_blocks, block_size)
examples/blocksparse_attention/example_tilelang_sparse_gqa_decode_varlen_mask.py:292
↓ 2 callersFunctionref_sparse_mla_fwd_interface(Q, KV, Indices, offsets, sm_scale=None, is_casual=True)
examples/dsa_sparse_finetune/sparse_mla_fwd.py:214
↓ 2 callersFunctionrepeat_kv This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, he
examples/flash_decoding/example_gqa_decode_varlen_logits.py:11
↓ 2 callersFunctionrun_autotune(M, N, K, M_value=None, N_value=None, K_value=None)
testing/python/autotune/test_tilelang_autotune_with_inputs.py:103
↓ 2 callersFunctionrun_clamp(
N,
block_N,
dtype,
min=None,
max=None,
)
testing/python/language/test_tilelang_language_clamp.py:29
↓ 2 callersFunctionrun_cython_dynamic_shape(
M, N, K, trans_A, trans_B, in_dtype, out_dtype, dtypeAccum, block_M, block_N, block_K, num_stages=3, num
testing/python/jit/test_tilelang_jit_gemm_cython.py:259
↓ 2 callersFunctionrun_fp8_fp4_gemm_1d1d(
a_fp4,
b_fp8,
sfa,
sfb,
out_dtype=T.bfloat16,
)
examples/deepseek_v4/fp8_fp4_gemm_1d1d_sm100.py:339
↓ 2 callersFunctionrun_gemm(
M,
N,
K,
trans_A,
trans_B,
in_dtype,
out_dtype,
dtypeAccum,
block_M,
maint/gemm/correctness_evaluation_sm70.py:93
↓ 2 callersFunctionrun_gemm(
M,
N,
K,
in_dtype,
out_dtype,
dtypeAccum,
block_M,
block_N,
block_K,
examples/dequantize_gemm/example_dequant_gemm_fine_grained.py:93
↓ 2 callersFunctionrun_gemm(
M,
N,
K,
trans_A,
trans_B,
in_dtype,
out_dtype,
dtypeAccum,
block_M,
testing/python/components/test_tilelang_pass_config_disable_warp_specialized.py:52
↓ 2 callersFunctionrun_gemm_rs(
M,
N,
K,
trans_A,
trans_B,
in_dtype,
out_dtype,
dtypeAccum,
block_M,
maint/gemm/correctness_evaluation_sm70.py:176
↓ 2 callersFunctionrun_gemm_ss(
M,
N,
K,
trans_A,
trans_B,
in_dtype,
out_dtype,
dtypeAccum,
block_M,
testing/python/tilelibrary/test_tilelang_tilelibrary_gemm_sp.py:64
↓ 2 callersFunctionrun_grouped_mxfp8_blockscaled_gemm(
a,
b,
sfa_flat,
sfb_flat,
offsets,
max_M_per_E,
transpose_B=True,
persistent
examples/blockscaled_gemm_sm100/grouped_gemm_mxfp8_blockscaled_1d1d.py:536
↓ 2 callersFunctionrun_tilelang_copy_with_stride(M=1024, N=1024, NN=2048, block_M=128, block_N=128, dtype=T.float16)
testing/python/language/test_tilelang_language_copy.py:61
↓ 2 callersFunctionrun_tilelang_grouped_gemm_ptr(
batch_sizes_list,
K,
N,
block_M,
block_N,
block_K,
num_stages=2,
threads=128
examples/grouped_gemm/example_grouped_gemm_fwd_ptr.py:119
↓ 2 callersMethodset_kernel_parameters(self, k_parameters: tuple[tuple[Any, ...], tuple[tuple[str, Any], ...]], f_parameters: dict[str, Any])
tilelang/autotuner/tuner.py:412
↓ 2 callersFunctionshuffle_weight(
x: torch.Tensor,
layout=(16, 32),
k_pack=1,
is_transpose=False,
)
testing/python/amd/test_tilelang_gemm_mfma_preshuffle.py:251
↓ 2 callersFunctionsparse_attn_fwd(
batch,
heads,
seq_len,
seq_len_kv,
dim,
topk,
sm_scale=None,
block_N=64,
examples/deepseek_v4/sparse_attn_fwd_sm90.py:27
↓ 2 callersFunctionsparse_mla_bwd(q, kv, o, do, indices, lse, sm_scale=None, is_casual=True, return_kernel=False, delta=None)
examples/deepseek_v32/sparse_mla_bwd.py:244
↓ 2 callersFunctionsparse_mla_fwd(
Q,
KV,
Indices,
heads,
dim,
tail_dim,
topk,
kv_group=1,
sm_scale=None,
examples/deepseek_v32/sparse_mla_fwd.py:15
↓ 2 callersFunctionsparse_mla_fwd(
Q,
KV,
Indices,
q_start_index_s,
heads,
dim,
tail_dim,
topk,
kv_stride,
examples/deepseek_v32/sparse_mla_fwd_pipelined.py:13