MCPcopy Create free account

hub / github.com/tile-ai/tilelang / functions

Functions7,254 in github.com/tile-ai/tilelang

↓ 2 callersFunctionbind_then
(var, value, body)
testing/python/transform/test_tilelang_transform_simplify.py:11
↓ 2 callersFunctionbitonic_sort
( topk_index_shared: T.SharedBuffer([N], dtype=INT32), topk_value_shared: T.SharedBuffer([N],
examples/dsa_sparse_finetune/indexer_topk_reducesum.py:51
↓ 2 callersFunctionblock_sparse_flash_decode_gqa_indice_triton
( q, k_cache, v_cache, cache_seqlens, max_cache_seqlen, max_selected_blocks, block
examples/blocksparse_attention/example_triton_sparse_gqa_decode_varlen_indice.py:170
↓ 2 callersFunctionblock_sparse_flash_decode_gqa_mask_triton
( q, k_cache, v_cache, cache_seqlens, max_cache_seqlen, block_mask, block_size,
examples/blocksparse_attention/example_triton_sparse_gqa_decode_varlen_mask.py:167
↓ 2 callersFunctionblocksparse_flashattn
(batch, heads, seq_len, dim, downsample_len, is_causal)
examples/blocksparse_attention/example_tilelang_block_sparse_attn.py:35
↓ 2 callersMethodbody_rs
GEMM body_rs: A from register, B from shared memory
tilelang/contrib/cutedsl/gemm_v1.py:302
↓ 2 callersFunctionbuild_backend_buffers
()
testing/python/language/test_tilelang_language_ptr.py:185
↓ 2 callersFunctionbwd
( Q, KV, dO, Indices, Lse, Delta, dKV, H, D, D_tail, topk, kv_
examples/deepseek_v32/sparse_mla_bwd.py:77
↓ 2 callersFunctioncal_seq_idx_for_q
(cu_seqlens_qs: torch.LongTensor, cu_seqlens_qe: torch.LongTensor, seq_len: int)
examples/dsa_hisa/tilelang_utils.py:215
↓ 2 callersFunctioncanon_target_host
(target: str | Target, target_host: str | Target | None)
tilelang/engine/lower.py:187
↓ 2 callersFunctionceil_div_int
(xv, yv)
maint/gemm/correctness_evaluation_tcgen05_fp4_unpacked.py:125
↓ 2 callersFunctionceil_to_ue8m0
(t)
maint/gemm/correctness_evaluation_tcgen05_fp4_unpacked.py:131
↓ 2 callersFunctioncheck_cuda_availability
Check if CUDA is available on the system by locating the CUDA path. Returns: bool: True if CUDA is available, False otherwise.
tilelang/utils/target.py:118
↓ 2 callersFunctioncheck_last_trait
(region: list[Range])
tilelang/carver/matmul_analysis.py:352
↓ 2 callersFunctionchunk_gated_delta_rule_fwd_h
( k: torch.Tensor, w: torch.Tensor, u: torch.Tensor, g: torch.Tensor = None, gk: torch.Ten
examples/kda/FLA_KDA/fla_chunk_delta.py:470
↓ 2 callersFunctionchunk_local_cumsum_scalar
( g: torch.Tensor, chunk_size: int, reverse: bool = False, scale: float = None, cu_seqlens
examples/kda/FLA_KDA/cumsum.py:246
↓ 2 callersFunctionclear_let_values
Clear all Python-side bindings recorded for the current thread.
tilelang/language/frame.py:127
↓ 2 callersFunctionclone_data
Recursively goes through data and clones all tensors.
examples/fusedmoe/example_fusedmoe_torch.py:197
↓ 2 callersFunctioncollect_vars_from_expr
(prim_expr)
tilelang/carver/matmul_analysis.py:28
↓ 2 callersFunctioncompare_a
(target, b, s_q, cache_seqlens, h_q, h_kv, d, dv, causal, dtype)
examples/deepseek_mla/benchmark_mla.py:502
↓ 2 callersFunctioncompare_a
(target, b, s_q, cache_seqlens, h_q, h_kv, d, dv, causal, dtype)
examples/deepseek_mla/amd/benchmark_mla_decode_amd_aiter.py:167
↓ 2 callersFunctioncompare_a
(target, b, s_q, cache_seqlens, h_q, h_kv, d, dv, causal, dtype)
examples/deepseek_mla/amd/benchmark_mla_decode_amd_triton.py:386
↓ 2 callersMethodcompile
(self, *args: _P.args, **kwargs: _P.kwargs)
tilelang/autotuner/tuner.py:1324
↓ 2 callersFunctioncompile_cuda
Compile cuda code with NVCC from env. Parameters ---------- code : str The cuda code. target_format : str The target
tilelang/contrib/nvcc.py:48
↓ 2 callersMethodcompile_lib
(self, timeout: float = None)
tilelang/jit/adapter/cutedsl/libgen.py:54
↓ 2 callersFunctioncomplex_expr_1
()
testing/python/arith/test_arith_hard.py:26
↓ 2 callersFunctioncomplex_expr_kernel
Kernel with complex expressions involving symbolic variables.
testing/python/language/test_tilelang_language_subtype.py:155
↓ 2 callersMethodcompute_elements_from_shape
(self, shape: list[int])
tilelang/carver/roller/hint.py:58
↓ 2 callersMethodcompute_tcgen05_a_desc_params
Compute A descriptor parameters from the A shared buffer (SS variant). This is a pure-Python helper -- no TIR code is emitted. Param
tilelang/cuda/intrinsics/macro/tcgen05_macro_generator.py:719
↓ 2 callersMethodcompute_wgmma_a_desc_params
Compute A descriptor parameters from the A shared buffer region (SS variant). This is a pure-Python helper -- no TIR code is emitted.
tilelang/cuda/intrinsics/macro/wgmma_macro_generator.py:358
↓ 2 callersMethodcompute_workload_per_item
(self, output_tile)
tilelang/carver/roller/policy/default.py:190
↓ 2 callersMethodconstruct_dependency_target
(self, targets: tuple[str])
tilelang/carver/roller/shape_inference/tir.py:176
↓ 2 callersFunctionconvert_to_uint16
(x)
examples/deepseek_v32/topk_selector.py:10
↓ 2 callersFunctionconvert_to_uint32
(x)
examples/deepseek_v32/topk_selector.py:17
↓ 2 callersFunctioncopy_kernel
(X, Y)
testing/python/transform/test_tilelang_transform_lower_ldgstg.py:295
↓ 2 callersFunctiondebug_print_msg
(M=16, N=16, msg_only=False)
testing/python/debug/test_tilelang_debug_print.py:98
↓ 2 callersFunctiondecompose_col_major
(index_1d: int, basis: list[int])
tilelang/layout/gemm_sp.py:13
↓ 2 callersMethoddeduplicate
(self, lst)
tilelang/carver/roller/shape_inference/tir.py:65
↓ 2 callersFunctiondefault_compile_options
Build a set of default NVCC compile options for TileLang generated sources. Includes C++ standard and common include paths (TileLang templat
tilelang/contrib/nvcc.py:153
↓ 2 callersFunctiondeprecated_warning
A function to indicate that a method is deprecated
tilelang/utils/deprecated.py:1
↓ 2 callersFunctiondequantize_gemv
( M: int, N: int, K: int, in_dtype: str, out_dtype: str, accum_dtype: str, num_bit
examples/dequantize_gemm/example_dequant_gemv_fp16xint4.py:12
↓ 2 callersFunctiondetermine_torch_fp8_type
(fp8_format: Literal["e4m3", "e5m2"] = "e4m3")
tilelang/utils/target.py:175
↓ 2 callersFunctiondevice_codegen
(device_mod: tvm.IRModule, target: Target)
tilelang/engine/lower.py:230
↓ 2 callersMethoddom_kind
The iteration domain kind of the block, for example, SSSS, SSSR.
tilelang/carver/analysis.py:72
↓ 2 callersFunctionelementwise_add
(A, B, block_M, block_N, in_dtype, out_dtype, threads)
examples/elementwise/example_elementwise_add.py:12
↓ 2 callersMethodenable
Enable kernel caching globally.
tilelang/env.py:187
↓ 2 callersMethodexport_library
Exports the compiled kernel function to a shared library file. Parameters ---------- kernel_file : str T
tilelang/jit/kernel.py:674
↓ 2 callersFunctionexpr_to_zeros
(target: ast.expr)
tilelang/autodd.py:120
↓ 2 callersMethodextent_wrapper
(self, value)
tilelang/carver/roller/node.py:292
↓ 2 callersFunctionextract_smem_offsets
Extract merged shared-memory offsets from generated code.
testing/python/transform/test_tilelang_transform_disable_memory_reuse.py:127
↓ 2 callersMethodextract_thread_binding
is_m_first: True if the thread binding is in the form of (tx, warp_n, warp_m) which represents [warp_size, block_row_warps (split n),
tilelang/cuda/intrinsics/macro/mma_sm70_macro_generator.py:161
↓ 2 callersFunctionfactorize
(n: int)
tilelang/carver/roller/policy/common.py:18
↓ 2 callersFunctionfast_round_scale
(amax, fp8_max_inv)
examples/deepseek_v4/act_quant.py:20
↓ 2 callersFunctionfilter_and_record
Strip kernel-resource-usage remarks from ``output``, parse them, and append the parsed entries to the active recorder (if any). Returns the fi
tilelang/contrib/hip_resource_info.py:77
↓ 2 callersFunctionfind_args_index
(sch: Schedule, buffer: tirx.Buffer)
tilelang/carver/matmul_analysis.py:169
↓ 2 callersFunctionfind_correction_dim
Computes the correction dimension for a given number of rotations in the rotary positional embedding. Args: num_rotation
examples/deepseek_v32/inference/model.py:357
↓ 2 callersFunctionfind_rocm_path
Utility function to find ROCm path Returns ------- path : str Path to ROCm root.
tilelang/contrib/rocm.py:267
↓ 2 callersFunctionflashattn
(batch, heads, seqlen_q, seqlen_kv, dim, is_causal, block_M, block_N)
examples/flash_decoding/example_mha_inference.py:12
↓ 2 callersFunctionflashattn
( batch, heads, k_heads, max_seqlen_kv, total_seqlen_k, dim, has_sink, block_N=128, block_H=64, num_split=
examples/flash_decoding/example_gqa_decode_varlen_logits.py:38
↓ 2 callersFunctionflashattn
(batch, heads, heads_kv, dim, dim_v, block_N, block_H, num_stages, threads)
examples/blocksparse_attention/example_tilelang_sparse_gqa_decode_varlen_indice.py:19
↓ 2 callersFunctionflashattn
(batch, heads, heads_kv, dim, dim_v, block_N, block_H, page_block_size, num_stages, threads, num_pages)
examples/blocksparse_attention/example_tilelang_sparse_gqa_decode_paged.py:21
↓ 2 callersFunctionflashattn
(batch, heads, kv_head_num, seqlen_kv, dim, pe_dim, block_N, block_H, num_split, softmax_scale)
examples/deepseek_mla/example_mla_decode.py:14
↓ 2 callersFunctionflashattn_bwd
(batch, heads, seq_len, dim, is_causal, block_M, block_N)
examples/flash_attention/example_mha_bwd_bhsd.py:158
↓ 2 callersFunctionflashattn_bwd
(batch, heads, seq_len, dim, is_causal, block_M, block_N)
examples/flash_attention/example_mha_bwd_bshd.py:155
↓ 2 callersFunctionflashattn_bwd
( batch, heads, seq_len, dim, window_size=None, # None for full attention sm_scale=No
examples/attention_sink/example_mha_sink_bwd_bhsd.py:203
↓ 2 callersFunctionflashattn_bwd
(batch, heads, seq_len, dim, groups, window_size=None, sm_scale=None, dtype=T.float16)
examples/attention_sink/example_gqa_sink_bwd_bhsd.py:204
↓ 2 callersFunctionflashattn_bwd
GQA backward: K/V/dK/dV use bx // groups; dK/dV use atomic_add.
examples/flash_attention_sm100/gqa_bwd_bshd.py:153
↓ 2 callersFunctionflashattn_bwd
Blackwell MHA backward. Pipeline default (128, 2); ts = (256, 2).
examples/flash_attention_sm100/mha_bwd_bshd.py:148
↓ 2 callersFunctionflashattn_bwd_dsink
(batch, heads, seq_len, block=128, dtype: T.dtype = T.float16)
examples/attention_sink/example_mha_sink_bwd_bhsd.py:314
↓ 2 callersFunctionflashattn_bwd_dsink
(batch, heads, seq_len, block=256, dtype: T.dtype = T.float16)
examples/attention_sink/example_gqa_sink_bwd_bhsd.py:306
↓ 2 callersFunctionflashattn_bwd_preprocess
(batch, heads, seq_len, dim)
examples/flash_attention/example_mha_bwd_bhsd.py:92
↓ 2 callersFunctionflashattn_bwd_preprocess
(batch, heads, seq_len, dim_v)
examples/flash_attention/example_gqa_bwd.py:90
↓ 2 callersFunctionflashattn_bwd_preprocess
(batch, heads, seq_len, dim)
examples/flash_attention/example_mha_bwd_bshd.py:89
↓ 2 callersFunctionflashattn_bwd_preprocess
(batch, heads, total_q, N_CTX, max_seq_len, dim_v)
examples/flash_attention/example_gqa_bwd_tma_reduce_varlen.py:143
↓ 2 callersFunctionflashattn_bwd_preprocess
(batch, heads, seq_len, dim, dtype: T.dtype = T.float16)
examples/attention_sink/example_mha_sink_bwd_bhsd.py:139
↓ 2 callersFunctionflashattn_bwd_preprocess
(batch, heads, seq_len, dim, dtype: T.dtype = T.float16)
examples/attention_sink/example_gqa_sink_bwd_bhsd.py:140
↓ 2 callersFunctionflashattn_bwd_split
(batch, heads, seq_len, dim_qk, dim_v, is_causal, block_M, block_N, threads=256, num_stages=2, groups=1)
examples/flash_attention/example_gqa_bwd.py:248
↓ 2 callersFunctionflashattn_bwd_split
( batch, total_q, total_kv, N_CTX, heads, max_seq_len, dim_qk, dim_v, is_causal, block_M, block_N, threads
examples/flash_attention/example_gqa_bwd_tma_reduce_varlen.py:366
↓ 2 callersFunctionflashattn_fwd
( batch, heads, seq_len, dim, window_size=None, # None for full attention, sm_scale=N
examples/attention_sink/example_mha_sink_bwd_bhsd.py:28
↓ 2 callersFunctionflashattn_fwd
( batch, heads, seq_len, dim, groups=1, window_size=None, # None for full attention
examples/attention_sink/example_gqa_sink_bwd_bhsd.py:26
↓ 2 callersFunctionflashmla_decode
(batch, heads, kv_head_num, seqlen_kv, dim, pe_dim, block_N, block_H, num_split, threads=128)
examples/deepseek_mla/amd/benchmark_mla_decode_amd_tilelang.py:37
↓ 2 callersMethodfootprint
(self, shape, rstep, stride_map: dict | None = None)
tilelang/carver/roller/node.py:447
↓ 2 callersFunctionfp4_act_quant
Block-wise FP4 quantization (bf16 -> float4_e2m1fn, E8M0 scale). Args: x: Input tensor, shape (..., N) with N divisible by block_size, dt
examples/deepseek_v4/act_quant.py:166
↓ 2 callersFunctionfp4_tma_copy_unpacked_smem_load
(M=128, N=256, block_M=64, block_N=128)
testing/python/language/test_tilelang_language_tma_copy.py:288
↓ 2 callersFunctionfp8_act_quant
Block-wise FP8 quantization (bf16 -> fp8_e4m3). Args: x: Input tensor, shape (..., N) with N divisible by block_size, dtype bfloat16.
examples/deepseek_v4/act_quant.py:88
↓ 2 callersMethodfree
(self, block: Block)
tilelang/carver/roller/bestfit.py:53
↓ 2 callersMethodfrom_bit_attention
(cls, bit_attention: BitnetAttention)
examples/bitnet-1.58b/modeling_bitnet.py:471
↓ 2 callersMethodfrom_bit_mlp
(cls, bit_mlp: BitnetMLP)
examples/bitnet-1.58b/modeling_bitnet.py:255
↓ 2 callersMethodfrom_lazy_style
Create template from lazy-style function that returns PrimFunc directly.
tilelang/language/eager/builder.py:1070
↓ 2 callersMethodfrom_quantized
load quantized model from local disk
examples/bitnet-1.58b/modeling_bitnet.py:1386
↓ 2 callersMethodfrom_warp_partition
Determine the warp policy based on the given warp partitioning. Args: m_warp (int): Number of warps in the row dimension
tilelang/tileop/base.py:161
↓ 2 callersFunctionfsymbolic_fuse0
(x)
testing/python/arith/test_arith_iter_affine_map.py:1082
↓ 2 callersFunctionfsymbolic_fuse1
(x)
testing/python/arith/test_arith_iter_affine_map.py:1089
↓ 2 callersFunctionfsymbolic_fuse2
(i)
testing/python/arith/test_arith_iter_affine_map.py:1096
↓ 2 callersMethodfunc
(*inputs: torch.Tensor | Any)
tilelang/jit/adapter/tvm_ffi.py:204
↓ 2 callersFunctiongemm
(M, N, K)
testing/python/carver/test_tilelang_carver_generate_hints.py:12
↓ 2 callersFunctiongemm
TileLang GEMM operator. This is the default synchronous GEMM interface. On Hopper, if the compiler selects WGMMA lowering, TileLang inserts t
tilelang/language/gemm_op.py:149
↓ 2 callersFunctiongemm_clc_persistent_2cta
( A, B, block_M, block_N, store_block_N, block_K, in_dtype, out_dtype, acc
examples/gemm_sm100/gemm_tcgen5mma_ws_clc.py:15
↓ 2 callersFunctiongemm_clc_persistent_2cta_pipelined_clc
( A, B, block_M, block_N, store_block_N, block_K, in_dtype, out_dtype, acc
examples/gemm_sm100/gemm_tcgen5mma_ws_clc.py:181
↓ 2 callersFunctiongen_stride
(shape_ik, order)
tilelang/layout/gemm_sp.py:69
← previousnext →1,001–1,100 of 7,254, ranked by callers