MCPcopy Create free account

hub / github.com/ByteDance-Seed/Triton-distributed / functions

Functions5,103 in github.com/ByteDance-Seed/Triton-distributed

⨍ Functions 5,103◇ Types & classes 521↳ Endpoints 149
↓ 2 callersMethod__init__
(self)
python/little_kernel/core/ir_base.py:57
↓ 2 callersMethod__str__
(self)
python/triton_dist/language/simt_ops.py:152
↓ 2 callersFunction__syncthreads
(_semantic=None)
python/triton_dist/language/extra/cuda/language_extra.py:109
↓ 2 callersFunction_as_index
(config: triton.Config)
python/triton_dist/tools/tune/tune_gemm.py:288
↓ 2 callersFunction_attn_fwd_inner
(acc, l_i, m_i, q, # desc_k, desc_v, # dtype: tl.constexpr, start_m,
python/triton_dist/mega_triton_kernel/test/triton_impl_utils.py:32
↓ 2 callersFunction_attn_fwd_with_qkv_desc
( desc_q, desc_k, desc_v, start_m, offset_o, out_ptr, sm_scale, # Z, H_Q,
python/triton_dist/mega_triton_kernel/test/triton_impl_utils.py:82
↓ 2 callersFunction_build_gemm_tiling
Build sorted per-m-tile grouped-GEMM metadata from per-expert token counts. Returns expert_ids / split_size_cum / tile_num / tile_num_cum (each i
python/triton_dist/kernels/amd/ep_all2all_fused.py:516
↓ 2 callersFunction_build_gemm_tiling_device
Device-kernel version of ``_build_gemm_tiling`` (same dict layout).
python/triton_dist/kernels/amd/ep_all2all_fused.py:750
↓ 2 callersFunction_codegen_flash_attn_impl
(task: FlashAttnTask, NUM_KV_HEADS, qkv_pack=False)
python/triton_dist/mega_triton_kernel/tasks/flash_attn.py:69
↓ 2 callersMethod_collect_nodes_with_attrs
Collect all nodes with custom attributes and intrin info.
python/little_kernel/core/passes/utils/enhanced_unparse.py:71
↓ 2 callersFunction_compute_pid
(pid, NUM_SMS, NUM_XCDS, GROUP_SIZE_M)
python/triton_dist/kernels/amd/gemm.py:48
↓ 2 callersFunction_compute_pid
(tile_id, num_pid_in_group, num_pid_m, GROUP_SIZE_M)
python/triton_dist/kernels/amd/gemm_allreduce.py:94
↓ 2 callersFunction_compute_pid
(tile_id, num_pid_in_group, num_pid_m, GROUP_SIZE_M, NUM_SMS)
python/triton_dist/kernels/nvidia/gemm_allreduce.py:373
↓ 2 callersFunction_compute_pid
(tile_id, num_pid_in_group, num_pid_m, GROUP_SIZE_M, NUM_GEMM_SMS)
python/triton_dist/kernels/nvidia/sp_ulysess_o_all2all_gemm.py:111
↓ 2 callersFunction_copy_if_changed
(dest: Path, src: Path)
python/triton_dist/tools/compile_aot.py:141
↓ 2 callersFunction_cp_engine_copy_data
(dst_ptr, src_ptr, cp_size, stream)
python/triton_dist/kernels/nvidia/sp_ag_attention_inter_node.py:216
↓ 2 callersFunction_create_data
(numel, dtype=torch.float32)
python/triton_dist/test/nvidia/test_allreduce.py:71
↓ 2 callersMethod_emit_struct_definitions
Auto-generate C++ structs for all tracked StructType instances. Note: Enum and special struct definitions are now handled vi
python/little_kernel/codegen/codegen_base.py:187
↓ 2 callersMethod_ensure_context
Ensure this runtime's CUDA context is the current one.
python/little_kernel/runtime/cuda_runtime.py:328
↓ 2 callersFunction_flash_attn_forward_inner
( acc, l_i, m_i, q, global_offset_q, # K_block_ptr, V_block_ptr, # start_m,
python/triton_dist/kernels/nvidia/sp_ag_attention_intra_node.py:188
↓ 2 callersFunction_flash_attn_forward_inner
( acc, l_i, m_i, q, global_offset_q, # K_block_ptr, V_block_ptr, # start_m,
python/triton_dist/kernels/nvidia/sp_ag_attention_inter_node.py:260
↓ 2 callersFunction_force_update_symlink_recursive
(dest_dir, src_dir)
python/setup.py:894
↓ 2 callersMethod_forward_push_3d
(self, symm_buffer: torch.Tensor, use_ll_protocol: bool = False)
python/triton_dist/layers/nvidia/low_latency_allgather_layer.py:80
↓ 2 callersMethod_full_key
(self, key)
python/triton_dist/tune.py:419
↓ 2 callersMethod_generate_method
Generate C++ method from Python method AST.
python/little_kernel/codegen/special_struct/struct_converter.py:345
↓ 2 callersFunction_generate_random_choosed_experts
(ntokens, topk, nexperts, generator: torch.Generator = None)
python/triton_dist/test/nvidia/test_moe_utils.py:39
↓ 2 callersFunction_get_algo_info
(kernel_name, orig_kernel_name)
python/triton_dist/tools/compile_aot.py:327
↓ 2 callersMethod_get_attr_comment
Get the comment string for a node's custom attributes.
python/little_kernel/core/passes/utils/enhanced_unparse.py:190
↓ 2 callersMethod_get_cache_key
(self)
python/little_kernel/runtime/kernel.py:72
↓ 2 callersMethod_get_cache_path
(self, key)
python/triton_dist/tune.py:422
↓ 2 callersFunction_get_mori_shmem_libdevice
()
python/triton_dist/utils.py:704
↓ 2 callersFunction_get_mxshmem_libdevice
()
python/triton_dist/utils.py:665
↓ 2 callersMethod_get_operator_str
Map Python AST operators to C++ operators.
python/little_kernel/codegen/visitors/expression_codegen.py:211
↓ 2 callersFunction_get_rocshmem_libdevice
()
python/triton_dist/utils.py:650
↓ 2 callersMethod_handle_builtin_with_codegen
Handle builtin function with codegen function.
python/little_kernel/codegen/visitors/call_codegen.py:308
↓ 2 callersMethod_handle_empty_like_call
Handle `ll.empty` / `ll.zeros` calls, extract shape, dtype, and scope, then update memory analysis.
python/little_kernel/core/passes/mem_analysis.py:127
↓ 2 callersMethod_handle_llkernel_call
Handle LLKernel recursive calls.
python/little_kernel/codegen/visitors/call_codegen.py:330
↓ 2 callersMethod_hash
(self, key)
python/triton_dist/tune.py:416
↓ 2 callersMethod_init_torch_backend
Lazy initialization of PyTorch native backend
python/triton_dist/layers/nvidia/pp_block.py:152
↓ 2 callersFunction_is_cta_master
()
python/triton_dist/kernels/amd/common_ops.py:43
↓ 2 callersFunction_is_cuda_platform
Checks if 'nvidia-smi' is available on the system's PATH.
python/setup.py:57
↓ 2 callersFunction_is_gpu_master
()
python/triton_dist/kernels/nvidia/common_ops.py:53
↓ 2 callersMethod_lltype_to_cpp
Convert an LLType instance to a valid C++ type string.
python/little_kernel/codegen/codegen_base.py:180
↓ 2 callersFunction_make_data
(token_num)
python/triton_dist/test/amd/test_ep_ll_a2a.py:130
↓ 2 callersFunction_make_data
(M)
python/triton_dist/test/amd/test_gemm_rs_intra_node.py:156
↓ 2 callersFunction_make_data
(M, N, K, pg: torch.distributed.ProcessGroup)
python/triton_dist/test/amd/test_gemm_ar_intra_node.py:63
↓ 2 callersFunction_make_data
(token_num)
python/triton_dist/test/amd/test_ep_a2a.py:328
↓ 2 callersFunction_make_data
(M, N, K, has_bias, tp_group: torch.distributed.ProcessGroup)
python/triton_dist/test/amd/test_ag_gemm_intra_node.py:263
↓ 2 callersFunction_make_data
(token_num)
python/triton_dist/test/nvidia/test_ep_ll_a2a.py:126
↓ 2 callersFunction_make_data
(M)
python/triton_dist/test/nvidia/test_gemm_rs.py:208
↓ 2 callersFunction_make_data
(token_num)
python/triton_dist/test/nvidia/test_ep_a2a.py:437
↓ 2 callersFunction_make_data
(cur_local_seq_len)
python/triton_dist/test/nvidia/test_ulysses_sp_dispatch.py:193
↓ 2 callersFunction_make_data
(M)
python/triton_dist/test/nvidia/test_gemm_ar.py:40
↓ 2 callersFunction_mapping
(x, delta)
python/triton_dist/profiler_utils.py:73
↓ 2 callersMethod_materialize
Called by allocator to set the actual tensor.
python/triton_dist/utils.py:1160
↓ 2 callersFunction_multimem_st_impl
(ptr, val0, suffix: core.constexpr, _semantic=None)
python/triton_dist/language/extra/cuda/language_extra.py:260
↓ 2 callersFunction_path_to_binary
(binary: str)
python/triton_dist/nv_utils.py:402
↓ 2 callersFunction_pretty_duration
(duration_ms)
python/triton_dist/kernels/amd/gemm.py:778
↓ 2 callersMethod_process_single_index
Helper to process a single index element.
python/little_kernel/codegen/visitors/expression_codegen.py:286
↓ 2 callersFunction_ptx_suffix_to_tl_type
(suffix: core.constexpr, _semantic=None)
python/triton_dist/kernels/nvidia/memory_ops.py:55
↓ 2 callersFunction_reduce_scatter_fn
()
python/triton_dist/test/nvidia/test_reduce_scatter.py:77
↓ 2 callersFunction_run
(cmds)
python/triton_dist/tune.py:104
↓ 2 callersFunction_run_dist_triton
(A: torch.Tensor, B: torch.Tensor, ctx, straggler_option=None)
python/triton_dist/test/stress/stress_test_ag_gemm.py:92
↓ 2 callersFunction_store_v4_impl
(ptr, val0, val1, val2, val3, suffix: core.constexpr, scope="gpu", semantic="relaxed", _sem
python/triton_dist/language/extra/cuda/language_extra.py:201
↓ 2 callersFunction_team_sync_impl
(team, SCOPE_SUFFIX: core.constexpr, _semantic=None)
python/triton_dist/language/extra/cuda/libnvshmem_device.py:330
↓ 2 callersMethod_try_resolve_from_struct_stub
Try to resolve method return type from struct stub registry.
python/little_kernel/core/passes/utils/method_resolver.py:59
↓ 2 callersMethod_try_resolve_from_struct_stub
Try to resolve method return type from struct stub registry.
python/little_kernel/core/passes/utils/type_inference/method_resolver.py:102
↓ 2 callersMethod_unparse_with_comments
Unparse with comments inserted at appropriate positions.
python/little_kernel/core/passes/utils/enhanced_unparse.py:392
↓ 2 callersMethod_update_tasks
(self, tasks: List[TaskBase], do_not_update_dependency=False)
python/triton_dist/mega_triton_kernel/models/model_builder.py:152
↓ 2 callersFunction_verify_and_check_bitwise
(torch_outs: List[torch.Tensor], triton_dist_outs: List[torch.Tensor], atol, rtol)
python/triton_dist/test/nvidia/test_llm_ulysess_gemm_all2all_intra_node.py:135
↓ 2 callersFunction_verify_and_check_bitwise
(torch_outs: List[torch.Tensor], triton_dist_outs: List[torch.Tensor], atol, rtol)
python/triton_dist/test/nvidia/test_ulysses_sp_infer_qkv_proj_a2a.py:48
↓ 2 callersFunction_verify_and_check_bitwise
(torch_outs: List[torch.Tensor], triton_dist_outs: List[torch.Tensor], atol, rtol)
python/triton_dist/test/nvidia/test_llm_ulysess_pre_attn_all2all_intra_node.py:165
↓ 2 callersMethodact
(t: torch.Tensor, quant: bool)
python/triton_dist/test/nvidia/test_ep_moe_inference.py:456
↓ 2 callersFunctionadd_continuous_kernel
(lhs_ptr, rhs_ptr, out_ptr, N, BLOCK_SIZE: tl.constexpr)
python/triton_dist/kernels/nvidia/reduce_scatter.py:182
↓ 2 callersFunctionadd_parent_references
Recursively traverse the AST and add a `parent` attribute to each node, pointing to its parent node in the AST hierarchy.
python/little_kernel/core/passes/utils/add_parent_reference.py:30
↓ 2 callersFunctionag_gemm_intra_node
allgather gemm for intra-node Allgather global matrix A and do matmul with local matrix B, produces local matrix C Args: a (torch.Te
python/triton_dist/kernels/metax/allgather_gemm.py:1128
↓ 2 callersFunctionag_gemm_torch
(A: torch.Tensor, B: torch.Tensor, tp_group: torch.distributed.ProcessGroup)
python/triton_dist/test/nvidia/test_ag_gemm.py:67
↓ 2 callersFunctionag_group_gemm
(a: torch.Tensor, b: torch.Tensor, ctx: MoEAllGatherGroupGEMMTensorParallelContext, full_topk_ids)
python/triton_dist/kernels/nvidia/allgather_group_gemm.py:402
↓ 2 callersFunctionalign_to
(value, alignment)
python/triton_dist/test/nvidia/test_fast_allgather.py:127
↓ 2 callersFunctionall_to_all_post_process
( ctx: AllToAllContext, input_splits: torch.Tensor, recv_buffer: torch.Tensor, scale_buffer: O
python/triton_dist/kernels/amd/low_latency_all_to_all.py:277
↓ 2 callersFunctionall_to_all_single_gemm
All-to-All Single GEMM operation with separated communication and computation Uses copy engine (cudaMemcpyAsync) for communication and s
python/triton_dist/kernels/nvidia/all_to_all_single_gemm.py:298
↓ 2 callersFunctionall_zeros
(data_sizes)
python/triton_dist/mega_triton_kernel/core/utils.py:30
↓ 2 callersFunctionallgather_strided_chunked_pull_kernel
( pid, npid, symm_ptr, # (M, N), M = M_per_rank * num_ranks M_per_rank, N, stride_m,
python/triton_dist/kernels/amd/allgather.py:185
↓ 2 callersFunctionaot_runtime_path
()
python/triton_dist/tools/compile_aot.py:716
↓ 2 callersFunctionappend_result_immediately
Write single result immediately to CSV with forced flushing
python/triton_dist/benchmark/bench_tp_attn.py:78
↓ 2 callersFunctionapply_rotary_pos_emb
Applies Rotary Position Embedding to the input tensor with shape [bsz, seq, head, head_dim]. Args: x: Input tensor of shape (batch_s
python/triton_dist/layers/amd/tp_attn.py:128
↓ 2 callersMethodapply_rotary_pos_emb
Applies Rotary Position Embedding inplace.
python/triton_dist/layers/amd/tp_attn.py:228
↓ 2 callersFunctionapply_triton340_inductor_patch
Call this function Before your program to apply the necessary patches for compatibility with Triton 3.4.0 in the context of PyTorch 2.7.1 (`t
python/triton_dist/tools/monkey_inductor.py:348
↓ 2 callersFunctionbench
(kernel, M, N, K, iters=20, label='')
python/little_kernel/design/sm90_bf16_gemm.py:330
↓ 2 callersFunctionbench
()
python/little_kernel/design/test_flashcomm_compute.py:206
↓ 2 callersFunctionbisect_right_kernel_aligned
( sorted_values_ptr, # Pointer to sorted input array (1D) target_values, N: tl.constexpr, )
python/triton_dist/kernels/nvidia/common_ops.py:340
↓ 2 callersFunctionbuild_mori_shmem
()
python/setup.py:504
↓ 2 callersFunctionbuild_shmem
()
python/setup.py:517
↓ 2 callersMethodbuild_tasks
(cls, device_prop: 'DeviceProp', layer_id: int, dependency: TaskDependency, io_tensors: Li
python/triton_dist/mega_triton_kernel/core/builder.py:85
↓ 2 callersFunctioncalc_gather_index
( scatter_index: torch.Tensor, row_start: int, row_end: int, BLOCK_SIZE: int = 1024, )
python/triton_dist/test/nvidia/test_ep_a2a.py:92
↓ 2 callersFunctioncalculate_pcie_bandwidth_gbps
Calculate PCIe bandwidth for a given generation and number of lanes. Returns (per_direction_gbs, bidirectional_gbs) Args: genera
python/triton_dist/nv_utils.py:222
↓ 2 callersFunctioncalculate_settings
(n)
python/triton_dist/kernels/nvidia/swiglu.py:33
↓ 2 callersFunctioncall_ag
()
python/triton_dist/kernels/amd/allgather_gemm.py:1070
↓ 2 callersFunctioncall_ag
(use_pull)
python/triton_dist/kernels/metax/allgather_gemm.py:856
↓ 2 callersFunctioncdiv
(n, m)
python/triton_dist/kernels/nvidia/gemm_rs_threadblock_swizzle.py:247
← previousnext →601–700 of 5,103, ranked by callers