MCPcopy Create free account

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

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

↓ 6 callersFunctionnaive_nsa
r""" Args: q (torch.Tensor): Queries of shape `[B, T, HQ, K]` if `head_first=False` else `[B, HQ, T, K]`. k (torch.Ten
examples/deepseek_nsa/reference.py:9
↓ 6 callersMethodrecursive_set
set layers.25.mlp.up_proj to attr
examples/bitnet-1.58b/modeling_bitnet.py:1334
↓ 6 callersFunctionref_program
(A)
testing/python/language/test_tilelang_language_scan.py:61
↓ 6 callersFunctionretrieve_offset
Retrieve per-dimension minima offsets. - Buffer -> [0, 0, ...] - BufferRegion -> [r.min for r in region] - BufferLoad -> indices (or
tilelang/utils/language.py:401
↓ 6 callersFunctionrun_tilelang_grouped_gemm
(batch_sizes_list, K, M, block_M, block_N, block_K, trans_b, num_stages=2, threads=128, profile=False)
examples/grouped_gemm/example_grouped_gemm_fwd.py:129
↓ 6 callersMethodset_function
Sets the function for this template and returns itself. Args: func (PrimFunc): The function to associate with this templ
tilelang/carver/template/base.py:127
↓ 6 callersFunctiontarget_is_cuda
(target: Target)
tilelang/utils/target.py:319
↓ 6 callersMethodwgmma_fence_c
Emit ``warpgroup_fence_operand`` for the accumulator buffer.
tilelang/cuda/intrinsics/macro/wgmma_macro_generator.py:471
↓ 5 callersFunction_bitand_i32
(a: Int32, b: Int32, *, loc=None, ip=None)
tilelang/contrib/cutedsl/warp.py:124
↓ 5 callersFunction_bitor_i32
(a: Int32, b: Int32, *, loc=None, ip=None)
tilelang/contrib/cutedsl/warp.py:129
↓ 5 callersFunction_ceil_div
(x: int, y: int)
examples/deepseek_v4/fp8_fp4_gemm_1d1d_sm100.py:12
↓ 5 callersFunction_collect_init_barrier_calls
(stmt)
testing/python/transform/test_tilelang_transform_lower_shared_barrier.py:36
↓ 5 callersFunction_collect_leading_bind_names
(mod, block_names)
testing/python/transform/test_tilelang_transform_Inject_software_pipeline.py:429
↓ 5 callersFunction_compile_tvm_ffi
(func, pass_configs=None, **kwargs)
testing/python/transform/test_tilelang_transform_producer_consumer_ws.py:422
↓ 5 callersMethod_emit_assign_target
(self, target: ast.expr, rval: ast.expr, annot: ast.expr = None)
tilelang/language/eager/ast.py:322
↓ 5 callersMethod_get_namespace_root
()
tilelang/cache/kernel_cache.py:197
↓ 5 callersFunction_get_target_from_backend
Map backend to target string.
testing/python/cache/test_tilelang_kernel_cache.py:41
↓ 5 callersMethod_legalize_to_buffer_region
Convert Buffer/BufferRegion/BufferLoad to a BufferRegion. - Buffer -> full-region BufferRegion covering entire shape - Buffe
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:870
↓ 5 callersMethod_load_kernel_from_disk
Loads a previously compiled kernel from disk cache. Args: key (str): The hash key identifying the kernel. ta
tilelang/cache/kernel_cache.py:551
↓ 5 callersFunction_log
(msg: str)
cmake/ccache_strip_pep517.py:45
↓ 5 callersMethod_lookup_type
(self, dtype: str | Any)
tilelang/jit/adapter/wrapper.py:783
↓ 5 callersMethod_make_for
Create a new For node with updated body, preserving other attributes.
tilelang/transform/decouple_type_cast.py:347
↓ 5 callersFunction_normalize_index_arg
Normalize warp sizing arguments so both Python ints and PrimExpr values are accepted uniformly.
tilelang/language/builtin.py:20
↓ 5 callersFunction_package_version
Return an installed package version or a readable placeholder.
testing/python/language/test_tilelang_language_pdl.py:171
↓ 5 callersFunction_restore_forced_value
(name, value)
testing/python/components/test_tilelang_env.py:10
↓ 5 callersFunction_run_if_stmt_binding
(func, config=None)
testing/python/transform/test_tilelang_transform_if_stmt_binding.py:7
↓ 5 callersFunction_to_tl_dtype
(torch_dtype: torch.dtype)
tilelang/utils/sparse.py:30
↓ 5 callersMethodassign_pass_configs
(self, pass_configs: dict[str, Any])
tilelang/jit/adapter/wrapper.py:956
↓ 5 callersFunctionbuffer
The buffer declaration function. Parameters ---------- shape : Union[List[PrimExpr], Tuple[PrimExpr], PrimExpr, Integral] The typ
tilelang/language/ast/ir.py:95
↓ 5 callersFunctionceil_div
Perform ceiling division of two integers. Args: x: the dividend. y: the divisor. Returns: The result of the cei
examples/cast/example_group_per_split_token_cast_to_fp8.py:59
↓ 5 callersFunctionceildiv
(a, b)
examples/deepseek_deepgemm/example_deepgemm_fp8_2xAcc.py:83
↓ 5 callersFunctioncheck_shared_mem
(arch: str = "none", tensor_idx: int = 0)
examples/kda/FLA_KDA/fla_utils.py:234
↓ 5 callersFunctioncheck_sm_version
(arch: str)
tilelang/carver/matmul_analysis.py:542
↓ 5 callersFunctiondebug_print_buffer
(M=16, N=16, dtype=T.float16)
testing/python/debug/test_tilelang_debug_print.py:9
↓ 5 callersFunctioneval_op
(op: Operator, left: Any, right: Any)
tilelang/language/eager/ast.py:93
↓ 5 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_sp_macro_generator.py:197
↓ 5 callersMethodfunc
(self)
tilelang/profiler/__init__.py:272
↓ 5 callersFunctiongenerate_qkv
Arguments: q: (batch_size, seqlen_q, nheads, d) k: (batch_size, seqlen_k, nheads_k, d) v: (batch_size, seqlen_k, nheads_k
examples/flash_attention/varlen_utils.py:19
↓ 5 callersFunctionget_arch
(target: str | Target = "cuda")
tilelang/carver/arch/__init__.py:14
↓ 5 callersMethodget_block_info
(self, block: SBlockRV)
tilelang/carver/roller/node.py:43
↓ 5 callersMethodget_host_source
Returns the source code of the host function.
tilelang/jit/kernel.py:460
↓ 5 callersFunctionget_kernel
(reduce_op: str, dtype: str)
testing/python/language/test_tilelang_language_warp_reduce.py:9
↓ 5 callersFunctionget_mxfp_intrin_group
Return metadata for an MXFP decoding intrinsic: function name and C source string. Validates the requested output dtype, source format, and
tilelang/quantize/mxfp.py:160
↓ 5 callersFunctionget_roller_hints_from_func
( func_or_module: tirx.PrimFunc | IRModule, arch: TileDevice, topk: int = 10, tensorcore_only: bool = Fals
tilelang/carver/utils.py:29
↓ 5 callersFunctionget_swizzle_layout
(row_idx, col_idx, row_size, dtype: DataType | str, swizzle_bytes=None)
tilelang/cuda/intrinsics/layout/mma_layout.py:203
↓ 5 callersMethodget_thread_binding
(self)
tilelang/rocm/intrinsics/mfma_macro_generator.py:337
↓ 5 callersFunctionget_var_name
Get the name of a Var, handling different TVM versions.
testing/python/transform/test_tilelang_transform_split_host_device.py:61
↓ 5 callersFunctionimplies
(x, y)
testing/python/arith/test_arith_hard.py:9
↓ 5 callersFunctionis_cuda_arch
(arch: TileDevice)
tilelang/carver/arch/cuda.py:13
↓ 5 callersFunctionis_cutedsl_target
(target: Target)
tilelang/jit/adapter/utils.py:113
↓ 5 callersFunctionis_rdna_arch
(arch: TileDevice)
tilelang/carver/arch/rdna.py:57
↓ 5 callersFunctionkernel
( A: T.Tensor((M, N), dtype), B: T.Tensor((M, N), dtype), C: T.Tensor((M,
testing/python/jit/test_tilelang_jit_nvrtc.py:406
↓ 5 callersFunctionlegalize_pairwise_extents
Right-align and broadcast two extent lists to be mutually compatible. Early-exit rule: - If the number of non-1 dimensions in `src_exten
tilelang/utils/language.py:483
↓ 5 callersFunctionmake_full_bank_swizzled_layout
Args: buffer: BufferLikeType Examples: make_full_bank_swizzled_layout(buffer)
tilelang/layout/swizzle.py:94
↓ 5 callersMethodmake_mfma_store_layout
Create a layout function for storing MFMA results into a fragment buffer. Parameters ---------- local_buf : tirx.Buf
tilelang/rocm/intrinsics/mfma_macro_generator.py:699
↓ 5 callersMethodmap_forward_index
Compute the forward index mapping for a given set of input indices. Parameters ---------- indices : list of PrimExpr
tilelang/layout/layout.py:93
↓ 5 callersMethodmatch
(self, node: ast.AST)
tilelang/autodd.py:308
↓ 5 callersFunctionmatmul
MXFP4 dequantize-GEMM kernel for AMD gfx950. Inputs: A : (M, K) in in_dtype (e.g. bfloat16) B : (N, K//2) uint8 -- pa
examples/dequantize_gemm/example_dequant_gemm_bf16_mxfp4_cdna4.py:63
↓ 5 callersFunctionmatmul
(M, N, K, block_M, block_N, block_K, dtype=T.float16, accum_dtype=T.float32)
testing/python/profiler/test_tilelang_profiler_dynamic_symbolic.py:14
↓ 5 callersMethodmma
Perform simdgroup matrix multiply-accumulate: C += A * B.
tilelang/metal/intrinsics/metal_macro_generator.py:164
↓ 5 callersFunctionmutate
Transform a Python function into an IR (Intermediate Representation) generator. This function takes a regular Python function and performs AS
tilelang/language/eager/ast.py:650
↓ 5 callersFunctionprepare_token_indices
(cu_seqlens: torch.LongTensor)
examples/dsa_sparse_finetune/index.py:80
↓ 5 callersMethodprint_stmt_brief
(self, stmt: Stmt, prefix: str)
tilelang/analysis/ast_printer.py:32
↓ 5 callersMethodpropagate
(self, tile, rstep: dict | None = None, targets=None)
tilelang/carver/roller/node.py:332
↓ 5 callersMethodpropagate_inputs
(self, tile, rstep: dict | None = None)
tilelang/carver/roller/node.py:341
↓ 5 callersFunctionregister_pipeline
Register a lowering pipeline for a backend. The pipeline name should match ``target.kind.name`` (e.g. ``"cuda"``, ``"hip"``, ``"c"``, ``"llvm
tilelang/backend/pass_pipeline/pipeline.py:29
↓ 5 callersFunctionrequires_cuda_compute_version
Mark a test as requiring at least a compute architecture Unit test marked with this decorator will run only if the CUDA compute architecture
tilelang/testing/__init__.py:106
↓ 5 callersFunctionretrive_ptr_from_buffer_region
(buffer_or_load_or_region: BufferLikeType, access_type: str = "r")
tilelang/utils/language.py:288
↓ 5 callersFunctionrun_gemm
( M, N, K, trans_A, trans_B, in_dtype, out_dtype, dtypeAccum, block_M,
maint/gemm/correctness_evaluation.py:92
↓ 5 callersMethodrun_with
(self, cls: type[TaskManager], *args, **kwargs)
tilelang/autodd.py:1235
↓ 5 callersMethodsave_to_disk
Persist autotune result to disk using atomic directory rename. All files are written into a temporary staging directory under the sha
tilelang/autotuner/param.py:385
↓ 5 callersFunctionserial
The serial For statement. Parameters ---------- start : PrimExpr The minimum value of iteration. stop : PrimExpr The
tilelang/language/loop.py:194
↓ 5 callersFunctiontarget_is_gfx950
(target: Target)
tilelang/utils/target.py:359
↓ 5 callersFunctiontorch_convert
Decode a 2D uint8 tensor into a 2D bfloat16 tensor by expanding each byte into two bf16 values using a 4-bit (nibble) encoding. Each input b
examples/dequantize_gemm/dequantize_utils.py:57
↓ 5 callersFunctiontorch_convert_bit_twiddling
This function expects `tensor` to be a 2-D torch.Tensor of dtype `torch.uint8`. Each output element is produced by combining two input bytes and
examples/dequantize_gemm/dequantize_utils.py:4
↓ 5 callersFunctionunwrap_cond
unwrap expr and convert to bool condition
tilelang/language/eager/builder.py:84
↓ 5 callersMethodwrap_device
(self, input: _T)
examples/bitnet-1.58b/vllm_workspace/conftest.py:140
↓ 4 callersFunctionCompute
( acc_s: T.FragmentBuffer([block_M, block_N], accum_dtype), acc_s_cast: T.FragmentBuff
examples/minference/example_vertical_slash_sparse_attn.py:60
↓ 4 callersFunctionLayoutVisual
Apply layout visualization pass if enabled.
tilelang/backend/pass_pipeline/pipeline_utils.py:83
↓ 4 callersFunctionSimplify
Simplify Returns ------- fpass : tvm.transform.Pass The result pass
tilelang/transform/simplify.py:20
↓ 4 callersFunction_align_up
(x: int, y: int)
examples/deepseek_v4/fp8_fp4_gemm_1d1d_sm100.py:192
↓ 4 callersFunction_apply
(func, target=sm100_target)
testing/python/transform/test_tilelang_transform_inject_tcgen05_fence.py:13
↓ 4 callersMethod_as_buffer
(buf_or_region)
tilelang/cuda/intrinsics/macro/tcgen05_macro_generator.py:183
↓ 4 callersFunction_as_range
The range constructor. Parameters ---------- dom : Union[Range, List[PrimExpr]] The domain. Returns ------- res : Ra
tilelang/language/ast/ir.py:510
↓ 4 callersFunction_assert_tl_access_ptr_bases_are_buffer_loads
(stmt)
testing/python/transform/test_tilelang_transform_legalize_safe_memory_access.py:52
↓ 4 callersMethod_assign_a_shared_layout
(self, layout: Layout)
tilelang/cuda/intrinsics/macro/wgmma_macro_generator.py:144
↓ 4 callersMethod_assign_b_shared_layout
(self, layout: Layout)
tilelang/cuda/intrinsics/macro/wgmma_macro_generator.py:148
↓ 4 callersMethod_body_impl
Internal implementation with configurable copy operations
tilelang/contrib/cutedsl/gemm_v1.py:322
↓ 4 callersFunction_collect_call_nodes
(stmt, op_names)
testing/python/transform/test_tilelang_transform_legalize_safe_memory_access.py:29
↓ 4 callersFunction_collect_calls
(stmt, op_name: str | set[str])
testing/python/transform/test_tilelang_transform_lower_shared_barrier.py:20
↓ 4 callersFunction_collect_fence_barrier_init
(stmt)
testing/python/transform/test_tilelang_transform_lower_shared_barrier.py:40
↓ 4 callersFunction_collect_set_max_nreg
(stmt)
testing/python/transform/test_tilelang_transform_inject_set_max_nreg.py:7
↓ 4 callersFunction_collect_tvm_access_ptr_offsets
(func: tvm.tirx.PrimFunc)
testing/python/transform/test_tilelang_transform_flatten_buffer.py:7
↓ 4 callersFunction_compile_and_check
( program, trans_A, trans_B, in_dtype, out_dtype, )
maint/gemm/correctness_evaluation.py:55
↓ 4 callersMethod_compute_shared_memory_usage
Computes the stride map for a given node and TileDict configuration. Parameters ---------- node : PrimFuncNode
tilelang/carver/roller/policy/default.py:452
↓ 4 callersMethod_create_dirs
()
tilelang/cache/kernel_cache.py:182
↓ 4 callersFunction_emit_tmem_fence
Emit tcgen05.wait fence. Called during @cute.jit compilation.
tilelang/contrib/cutedsl/gemm_tcgen05.py:509
↓ 4 callersFunction_ensure_allocator
(config: Config)
maint/scripts/pytest_cuda_scheduler.py:155
↓ 4 callersFunction_env_var_descriptor
(name)
testing/python/components/test_tilelang_env.py:6
↓ 4 callersFunction_find_cast_entry
Find the cast buffer for a given (buffer, indices) pair, or None.
tilelang/transform/decouple_type_cast.py:309
← previousnext →301–400 of 7,254, ranked by callers