MCPcopy Create free account

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

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

↓ 2 callersMethodtask_generator
(self)
tilelang/autodd.py:725
↓ 2 callersMethodtcgen05_ss_atom
Emit a single TCGEN05MMA SS instruction for atom ``(inst_m_idx, inst_n_idx, ki)``. Must be called after descriptor initialization and before
tilelang/cuda/intrinsics/macro/tcgen05_macro_generator.py:841
↓ 2 callersMethodtcgen05mma
Emit a TCGEN5MMA operation, dispatching to SS or TS variant based on A's memory scope. If *A_buf* resides in tensor memory (``shared.tmem``),
tilelang/cuda/intrinsics/macro/tcgen05_macro_generator.py:202
↓ 2 callersMethodtcgen05mma_blockscaled
Emit a block-scaled TCGEN5MMA (SS variant with TMEM scale factors). Uses ``tcgen05.mma.cta_group::1|2.kind::mxf8f6f4.block_scale`` PTX instru
tilelang/cuda/intrinsics/macro/tcgen05_macro_generator.py:321
↓ 2 callersFunctiontest
(n: int, h: int)
examples/deepseek_mhc/example_mhc_post.py:100
↓ 2 callersFunctiontest_gemm_fp8
(M, N, K, dtype)
examples/gemm_fp8/example_tilelang_gemm_fp8_2xAcc.py:57
↓ 2 callersFunctiontest_gemm_fp8
(M, N, K, dtype)
examples/gemm_fp8/example_tilelang_gemm_fp8.py:38
↓ 2 callersFunctiontilelang_chunk_fwd_o
( # task config B, S, H, DK, DV, input_dtype, output_dtype, accum_dtype,
examples/gdn/example_chunk_o.py:59
↓ 2 callersFunctiontilelang_chunk_gated_delta_rule_bwd_dhu
( # task config B, S, H, DK, DV, input_dtype, output_dtype, accum_dtype,
examples/gdn/example_chunk_delta_bwd.py:187
↓ 2 callersFunctiontilelang_chunk_gated_delta_rule_fwd_h
( # task config B, S, H, DK, DV, input_dtype, output_dtype, accum_dtype,
examples/gdn/example_chunk_delta_h.py:102
↓ 2 callersFunctiontilelang_chunk_local_cumsum_scalar
( # task config B, S, H, chunk_size=64, is_varlen=False, head_first=False, rev
examples/gdn/example_cumsum.py:24
↓ 2 callersFunctiontilelang_chunk_o_bwd_dqkwg
( # task config B, S, H, DK, DV, input_dtype, output_dtype, accum_dtype,
examples/gdn/example_chunk_o_bwd.py:114
↓ 2 callersFunctiontilelang_chunk_scaled_dot_kkt_fwd
( # task config B, S, H, DK, chunk_size=64, input_dtype=T.bfloat16, output_dty
examples/gdn/example_chunk_scaled_dot_kkt.py:53
↓ 2 callersFunctiontilelang_op
TileLang operation interface.
maint/precision/compare_ops.py:242
↓ 2 callersFunctiontilelang_recompute_w_u_fwd
( # task config B, S, H, DK, DV, input_dtype, output_dtype, gate_dtype,
examples/gdn/example_wy_fast.py:48
↓ 2 callersFunctiontilelang_wy_fast_bwd
( # task config B, S, H, DK, DV, input_dtype, output_dtype, accum_dtype,
examples/gdn/example_wy_fast_bwd_split.py:99
↓ 2 callersFunctiontilelang_wy_fast_bwd_split
( # task config B, S, H, DK, DV, input_dtype, output_dtype, accum_dtype,
examples/gdn/example_wy_fast_bwd_split.py:251
↓ 2 callersFunctiontl_fused_chunk_bwd
(Q, K, V, dO)
examples/linear_attention/example_linear_attn_bwd.py:123
↓ 2 callersFunctiontl_fused_chunk_bwd_kernel
( B, S, H, DK, DV, dtype: T.dtype = T.float16, scale: float = None, )
examples/linear_attention/example_linear_attn_bwd.py:13
↓ 2 callersFunctiontl_fused_chunk_fwd
(q, k, v)
examples/linear_attention/example_linear_attn_fwd.py:85
↓ 2 callersFunctiontl_fused_chunk_fwd_kernel
( B, S, H, DK, DV, dtype: T.dtype = T.float16, scale: float = None, )
examples/linear_attention/example_linear_attn_fwd.py:16
↓ 2 callersFunctiontl_gemm
( A, B, scales_a, scales_b, block_N, in_dtype, out_dtype, accum_dtype, )
examples/deepseek_deepgemm/example_deepgemm_fp8_2xAcc.py:12
↓ 2 callersFunctiontl_matmul_streamk
( A, B, C, streamk_tiles, block_M, block_N, block_K, trans_A, trans_B,
examples/gemm_streamk/example_tilelang_gemm_streamk.py:58
↓ 2 callersFunctiontma_store_wait_read_option
(read)
testing/python/language/test_tilelang_language_tma_store.py:129
↓ 2 callersMethodtok_encode
(self, string: str, add_special_tokens=True)
examples/bitnet-1.58b/eval_utils.py:100
↓ 2 callersMethodtop
Returns the item on the top of the stack without removing it, or None if the stack is empty.
tilelang/language/kernel.py:51
↓ 2 callersMethodtop
Return the top item of the stack without removing it. Returns: The top frame object from the stack Raises: I
tilelang/language/frame.py:81
↓ 2 callersFunctiontopo_sort_dfs
(node, visited, topo_order)
tilelang/carver/roller/node.py:560
↓ 2 callersFunctiontorch_chunk_gated_delta_rule_bwd_dhu
( Q: torch.Tensor, K: torch.Tensor, W: torch.Tensor, G: torch.Tensor, h0: torch.Tensor,
examples/gdn/example_chunk_delta_bwd.py:108
↓ 2 callersFunctiontorch_convert
(tensor)
examples/dequantize_gemm/example_dequant_gemm_fp4_hopper.py:32
↓ 2 callersFunctiontriton_program
(Q, K, V, Sinks, window_size: Optional[int] = None)
examples/attention_sink/benchmark_gqa_sink_fwd.py:98
↓ 2 callersFunctiontry_inline
Try to inline as many blocks as possible, and return the remaining blocks. Parameters ---------- sch : tirx.Schedule The TIR sche
tilelang/carver/common_schedules.py:92
↓ 2 callersFunctionunpack_sf_u32_1d
(packed_sf, mn, sf_k_blocks)
examples/blockscaled_gemm_sm100/gemm_mxfp8_blockscaled_1d1d.py:516
↓ 2 callersFunctionunpack_sf_u32_1d
(packed_sf: torch.Tensor, mn: int, sf_k_blocks: int)
examples/deepseek_v4/fp8_fp4_gemm_1d1d_sm100.py:212
↓ 2 callersFunctionunpack_sf_u32_rows
(packed_sf, sf_k_blocks)
examples/blockscaled_gemm_sm100/grouped_gemm_mxfp8_blockscaled_1d1d.py:445
↓ 2 callersMethodupdate_lib_code
(self, lib_code: str)
tilelang/jit/adapter/libgen.py:49
↓ 2 callersMethodupdate_lib_code
Update the library code with the given code string.
tilelang/jit/adapter/cutedsl/wrapper.py:1620
↓ 2 callersMethodupdate_tags
(self, tags: dict)
tilelang/carver/roller/node.py:107
↓ 2 callersMethodupdate_tuner_result
Updates the tuning results for this kernel. Parameters ---------- latency : float The measured latency o
tilelang/jit/kernel.py:573
↓ 2 callersFunctionvec_add
Build one vector-add candidate for the autotuner.
testing/python/autotune/test_tilelang_autotune_cutedsl_cache.py:80
↓ 2 callersFunctionwarp_shfl
(local, buf, thread_elem, warp_size, round)
examples/hadamard_transform/example_hadamard.py:17
↓ 2 callersFunctionweight_quant
(weight, num_bits=1)
examples/bitnet-1.58b/utils_quant.py:16
↓ 2 callersMethodwgmma
( self, A_region: BufferRegion, B_region: BufferRegion, C_region: BufferRegion, clear_accum: PrimExpr
tilelang/cuda/intrinsics/macro/wgmma_macro_generator.py:203
↓ 2 callersMethodwgmma_fence_a
Emit ``warpgroup_fence_operand`` for the A fragment buffer.
tilelang/cuda/intrinsics/macro/wgmma_macro_generator.py:461
↓ 2 callersMethodwgmma_ss
( self, A_region: BufferRegion, E_region: BufferRegion, B_region: BufferRegion
tilelang/cuda/intrinsics/macro/wgmma_sp_macro_generator.py:117
↓ 2 callersMethodwgmma_ss_atom
Emit a single WGMMA SS instruction for atom ``(inst_m_idx, inst_n_idx, ki)``. Must be called between fence/arrive and commit/wait sequences.
tilelang/cuda/intrinsics/macro/wgmma_macro_generator.py:609
↓ 2 callersMethodwith_source
(self, source: str)
tilelang/autodd.py:549
↓ 1 callersFunctionAtomicAdd
Perform atomic addition on a pointer. Supports float16, float32, int32, and int64 types. Returns the old value before addition (atomicrmw sem
tilelang/contrib/cutedsl/atomic.py:87
↓ 1 callersFunctionAtomicMax
Perform atomic maximum operation. For integers, uses nvvm.atomicrmw with MAX. For floats, uses CAS loop since PTX doesn't have atomic max for
tilelang/contrib/cutedsl/atomic.py:282
↓ 1 callersFunctionAtomicMin
Perform atomic minimum operation. For integers, uses nvvm.atomicrmw with MIN. For floats, uses CAS loop since PTX doesn't have atomic min for
tilelang/contrib/cutedsl/atomic.py:365
↓ 1 callersFunctionLowerAccessPtr
Lower TileLang frontend `tl.access_ptr` to `tir.builtin.tvm_access_ptr`. Returns ------- fpass : tvm.transform.Pass The result pa
tilelang/transform/__init__.py:195
↓ 1 callersFunctionLowerPTXAsyncCopy
Lower eligible global->shared copies into PTX `cp.async` on CUDA. When enabled (pass config `tl.enable_async_copy`, default True), this pass
tilelang/transform/__init__.py:265
↓ 1 callersFunctionMMA0
( K: T.Tensor(shape, dtype), Q_shared: T.SharedBuffer([block_M, dim], dtype),
benchmark/blocksparse_attention/benchmark_tilelang_block_sparse_fmha.py:49
↓ 1 callersFunctionMMA0
( K: T.Tensor(shape, dtype), Q_shared: T.Tensor([block_M, dim], dtype), K_
testing/python/transform/test_tilelang_transform_config_index_bitwidth.py:28
↓ 1 callersFunctionMMA1
( V: T.Tensor(shape, dtype), V_shared: T.SharedBuffer([block_M, dim], dtype),
benchmark/blocksparse_attention/benchmark_tilelang_block_sparse_fmha.py:68
↓ 1 callersFunctionMMA1
( V: T.Tensor(shape, dtype), V_shared: T.Tensor([block_M, dim], dtype), ac
testing/python/transform/test_tilelang_transform_config_index_bitwidth.py:47
↓ 1 callersFunctionMarkHostMetalContext
()
tilelang/metal/transform/mark_host_metal_context.py:50
↓ 1 callersFunctionProducerConsumerWarpSpecialized
Producer-consumer warp specialization at the tile-op level. This pass runs before LayoutInference and LowerTileOp. It rewrites eligible pipel
tilelang/cuda/transform/__init__.py:6
↓ 1 callersFunctionRescale
( acc_o: T.FragmentBuffer([block_M, dim], accum_dtype), scores_scale: T.FragmentBuffer
benchmark/blocksparse_attention/benchmark_tilelang_block_sparse_fmha.py:113
↓ 1 callersFunctionRescale
( acc_o: T.Tensor([block_M, dim], accum_dtype), scores_scale: T.Tensor([block_M], accu
testing/python/transform/test_tilelang_transform_config_index_bitwidth.py:92
↓ 1 callersFunctionSoftmax
( acc_s: T.FragmentBuffer([block_M, block_N], accum_dtype), acc_s_cast: T.FragmentBuff
benchmark/blocksparse_attention/benchmark_tilelang_block_sparse_fmha.py:81
↓ 1 callersFunctionSoftmax
( acc_s: T.Tensor([block_M, block_N], accum_dtype), acc_s_cast: T.Tensor([block_M, blo
testing/python/transform/test_tilelang_transform_config_index_bitwidth.py:60
↓ 1 callersFunction__dtype_is_float4_e2m1_unpacked__
8-bit FP4 E2M1 unpacked shared-memory storage. Only for tcgen05 ``kind::f8f6f4`` and block-scaled ``kind::mxf8f6f4`` mixed-precision paths. I
tilelang/language/dtypes.py:272
↓ 1 callersFunction__dtype_is_float4_e2m1fn__
Packed 4-bit FP4 E2M1 (CUTLASS float_e2m1_t). Use for pure FP4 workloads on tcgen05 mxf4 / mxf4nvf4 (packed SMEM, half the footprint). Global
tilelang/language/dtypes.py:262
↓ 1 callersMethod__exit__
(self, exc_type, exc_value, traceback)
tilelang/language/eager/builder.py:116
↓ 1 callersMethod__init__
(self, devices: Iterable[str])
maint/scripts/pytest_cuda_scheduler.py:48
↓ 1 callersMethod__init__
(self, *kargs, weight_bits=1, input_bits=8, **kwargs)
examples/bitnet-1.58b/utils_quant.py:214
↓ 1 callersMethod__init__
( self, model_name: str, dtype: str = "half", *, model_kwargs: Optiona
examples/bitnet-1.58b/vllm_workspace/conftest.py:146
↓ 1 callersMethod__init__
( self, a_dtype: str = T.float16, b_dtype: str = T.float16, accum_dtype: str =
tilelang/rocm/intrinsics/mfma_macro_generator.py:70
↓ 1 callersMethod__init__
(self, var: Var)
tilelang/analysis/fragment_loop_checker.py:14
↓ 1 callersMethod__init__
( self, a_dtype: str = T.float16, b_dtype: str = T.float16, accum_dtype: str =
tilelang/cuda/intrinsics/macro/mma_macro_generator.py:72
↓ 1 callersMethod__new__
(cls, *args)
tilelang/contrib/cutedsl/gemm_v1.py:234
↓ 1 callersMethod_analysis_funcinfo
(self)
tilelang/carver/roller/node.py:246
↓ 1 callersMethod_analyze_copy
Analyze memory copy operations (e.g., tl.copy). Args: call: A TVM Call node representing the copy operation.
tilelang/tools/Analyzer.py:58
↓ 1 callersMethod_analyze_gemm
Analyze matrix multiplication (GEMM) operations (e.g., tl.gemm). Args: call: A TVM Call node representing the GEMM operat
tilelang/tools/Analyzer.py:90
↓ 1 callersFunction_apply_plan_update
(func: tvm.tirx.PrimFunc)
testing/python/transform/test_tilelang_transform_plan_update_buffer_allocation_location.py:8
↓ 1 callersFunction_assert_fp4_packed_tma_descriptor
(host_source, desc_name)
testing/python/language/test_tilelang_language_tma_copy.py:346
↓ 1 callersFunction_assert_fp4_unpacked_tma_descriptor
(host_source, desc_name, *, expect_swizzle=None)
testing/python/language/test_tilelang_language_tma_copy.py:368
↓ 1 callersMethod_assign_block_size
Assigns a block size to a given PrimFuncNode based on the TileDict configuration and the specified block size. Parameters --
tilelang/carver/roller/policy/default.py:673
↓ 1 callersMethod_assign_placeholder_node
(self)
tilelang/carver/roller/node.py:217
↓ 1 callersMethod_assign_reduce_step
Assigns an optimal reduction step for the given PrimFuncNode. Parameters ---------- node : PrimFuncNode
tilelang/carver/roller/policy/default.py:269
↓ 1 callersFunction_ast_match
(temp: ast.AST, node: ast.expr, placeholders: set[str])
tilelang/autodd.py:219
↓ 1 callersFunction_async_raise_in_thread
Asynchronously raise ``exc_type`` in the thread identified by ``tid``. Uses CPython's :c:func:`PyThreadState_SetAsyncExc` — the same primitive th
tilelang/autotuner/tuner.py:65
↓ 1 callersFunction_bench_with_cuda_events
Benchmark using CUDA events for timing.
tilelang/profiler/bench.py:142
↓ 1 callersFunction_bench_with_cudagraph
Benchmark using CUDA graph for minimal launch overhead. This implementation follows triton.testing.do_bench_cudagraph. It captures the kernel
tilelang/profiler/bench.py:222
↓ 1 callersFunction_blockscaled_gemm_ref
(a, b, sfa_packed, sfb_packed, *, sf_granularity_k=128, transpose_B=False)
maint/gemm/correctness_evaluation_tcgen05_fp4_unpacked.py:267
↓ 1 callersFunction_buggy_kernel
(S: T.Tensor((8), T.bfloat16), D: T.Tensor((4, 64), T.bfloat16))
testing/python/issue/test_tilelang_issue_1744.py:10
↓ 1 callersFunction_build_autotuner
( M: int, N: int, K: int, with_roller: bool, profile_backend: str, execution_backend:
examples/gemm/example_gemm_advanced_autotune.py:161
↓ 1 callersMethod_build_kernel
( self, func: Callable | None, host_kernel_source: CachedTextSource, device_ke
tilelang/cache/kernel_cache.py:804
↓ 1 callersFunction_byte_offset
(access_ptr)
testing/python/issue/test_tilelang_issue_sm120_tma_smem_alignment.py:143
↓ 1 callersMethod_call_packed_name
Return the packed function name when a TIR argument names one.
tilelang/jit/adapter/cutedsl/wrapper.py:838
↓ 1 callersFunction_can_be_tensorized
(sch: Schedule, block: SBlockRV)
tilelang/carver/matmul_analysis.py:533
↓ 1 callersMethod_can_implement_layout
(self, node: PrimFuncNode, td: TileDict)
tilelang/carver/roller/policy/tensorcore.py:237
↓ 1 callersFunction_candidate_clang_cl_paths
(compiler_env: dict[str, str] | None)
tilelang/contrib/msvc.py:47
↓ 1 callersFunction_candidate_nvidia_dirs
()
cmake/find_pip_cuda.py:87
↓ 1 callersFunction_canon_backend
(name: str | None)
tilelang/jit/execution_backend.py:14
↓ 1 callersFunction_ceildiv_kernel
(a: int, b: int)
testing/python/language/test_tilelang_language_ceildiv.py:7
↓ 1 callersFunction_ceildiv_kernel_dyn
(b: int)
testing/python/language/test_tilelang_language_ceildiv.py:31
↓ 1 callersFunction_check
(original, transformed)
testing/python/transform/test_tilelang_transform_cluster_planning.py:10
↓ 1 callersFunction_check_is_cdna
()
tilelang/testing/__init__.py:38
← previousnext →1,401–1,500 of 7,254, ranked by callers