MCPcopy Create free account

hub / github.com/RL-Align/RL-Kernel / functions

Functions1,290 in github.com/RL-Align/RL-Kernel

⨍ Functions 1,290◇ Types & classes 193↳ Endpoints 3
↓ 1 callersMethod_current_gpu_uuid
(self)
rl_engine/executors/bridge.py:2626
↓ 1 callersFunction_deep_merge
(base: dict[str, Any], override: Mapping[str, Any])
rl_engine/executors/deepspeed_trainer.py:686
↓ 1 callersMethod_detect_backend
(self)
rl_engine/kernels/sampling.py:17
↓ 1 callersFunction_dlpack_ctypes
()
rl_engine/executors/bridge.py:889
↓ 1 callersFunction_dtype_default_atol
(dtype: torch.dtype, reference_mode: str)
tests/linear_logp_tp.py:44
↓ 1 callersFunction_dtype_default_rtol
(dtype: torch.dtype, reference_mode: str)
tests/linear_logp_tp.py:52
↓ 1 callersFunction_dtype_name
(dtype: torch.dtype)
rl_engine/kernels/gtest/op_checks.py:426
↓ 1 callersFunction_dtype_to_dlpack
(dtype: torch.dtype)
rl_engine/executors/bridge.py:865
↓ 1 callersFunction_embedding_layer
(model: torch.nn.Module)
rl_engine/executors/deepspeed_trainer.py:451
↓ 1 callersFunction_enough_gpu_memory
(num_bytes: int)
tests/test_attention.py:447
↓ 1 callersMethod_ensure_context
(self)
rl_engine/executors/bridge.py:1143
↓ 1 callersMethod_error_detail
(self, code: int)
rl_engine/executors/bridge.py:1158
↓ 1 callersFunction_expand_prompts
(prompts: Sequence[str | Mapping[str, Any]], num_generations: int)
rl_engine/executors/vllm_sampler.py:179
↓ 1 callersMethod_export_zero3_full_state_model
(self)
rl_engine/executors/deepspeed_trainer.py:248
↓ 1 callersFunction_extract_logits
(raw_outputs: Any)
rl_engine/executors/stateless_executor.py:529
↓ 1 callersFunction_extract_logits
(raw_outputs: Any)
rl_engine/executors/paged_kv_baseline.py:379
↓ 1 callersFunction_extract_named_tensor
(raw_outputs: Any, names: tuple[str, ...])
rl_engine/executors/stateless_executor.py:544
↓ 1 callersFunction_fallback_op
Portable op for inputs the SM90 forward cannot take (fp32/fp16, or a hidden dim not divisible by the kernel's K slice). Prefers Triton, else nativ
rl_engine/kernels/ops/cuda/loss/linear_logp.py:45
↓ 1 callersFunction_first_initialize_result
(init_result: Any)
rl_engine/executors/deepspeed_trainer.py:394
↓ 1 callersMethod_flat_row_indices
(self, row_indices: torch.Tensor, logits: torch.Tensor)
rl_engine/kernels/ops/pytorch/loss/logp.py:38
↓ 1 callersFunction_fused_logp_fn
( *, batch_size: int, seq_len: int, vocab_size: int, dtype: torch.dtype, device: torch
benchmarks/profiler.py:528
↓ 1 callersFunction_gloo_available
()
tests/test_linear_logp.py:54
↓ 1 callersFunction_group_outputs
( outputs: Sequence[Any], batch_size: int, num_generations: int, )
rl_engine/executors/vllm_sampler.py:187
↓ 1 callersFunction_has_hidden_state_metadata
(candidate: Any)
rl_engine/executors/deepspeed_trainer.py:525
↓ 1 callersFunction_has_logits_rank
(value: torch.Tensor)
rl_engine/alignment/model_wrappers.py:20
↓ 1 callersFunction_hf_rotate_half_reference
Independent HF rotate-half reference (from ISSUE_108_OPS_DEV §5).
tests/test_rope.py:35
↓ 1 callersMethod_infer_amd_arch
(name: str)
benchmarks/profiler.py:141
↓ 1 callersMethod_infer_nvidia_arch
(name: str, major: int, minor: int)
benchmarks/profiler.py:127
↓ 1 callersMethod_init_backend_assets
Preload backend-specific dependencies to avoid runtime import overhead.
rl_engine/kernels/sampling.py:25
↓ 1 callersFunction_init_distributed
()
tests/linear_logp_tp.py:67
↓ 1 callersFunction_install_vllm_cuda_vmm_aliases_on_worker
( manifest: WeightUpdateManifest, *, device_index: int, source_worker: str, source_rank: i
rl_engine/executors/bridge.py:543
↓ 1 callersFunction_linear_logp_op_for_device
(device: torch.device | str)
rl_engine/executors/deepspeed_trainer.py:437
↓ 1 callersFunction_linear_logp_parameter_context
( deepspeed_runtime: Any, model: torch.nn.Module, *, zero_stage: int, world_size: int, )
rl_engine/executors/deepspeed_trainer.py:584
↓ 1 callersMethod_load_backend
Dynamic loading technique: Import modules only when needed and check environment dependencies.
rl_engine/kernels/registry.py:257
↓ 1 callersFunction_load_deepspeed
()
rl_engine/executors/deepspeed_trainer.py:314
↓ 1 callersFunction_load_op
(source: str)
tests/linear_logp_tp.py:143
↓ 1 callersFunction_load_ray
()
rl_engine/executors/ray_actor_manager.py:225
↓ 1 callersFunction_logits_for
(inputs: StatelessForwardInputs, vocab_size: int = 8)
tests/test_paged_kv_baseline.py:66
↓ 1 callersFunction_looks_like_cuda_home
(path: Path)
rl_engine/executors/deepspeed_trainer.py:371
↓ 1 callersFunction_looks_like_kv_cache
(value: Any)
rl_engine/executors/stateless_executor.py:630
↓ 1 callersFunction_make_grad_outputs
( outputs: list[torch.Tensor], *, grad_mode: str, seed: int, )
rl_engine/kernels/gtest/op_checks.py:348
↓ 1 callersFunction_make_inputs
(num_prompts, spp, completion_len, vocab, device, dtype)
benchmarks/benchmark_grpo_loss.py:35
↓ 1 callersFunction_make_inputs
(num_tokens, hidden_dim, vocab, device, dtype)
benchmarks/benchmark_linear_logp.py:55
↓ 1 callersMethod_manifest_gpu_uuid
(self, manifest: WeightUpdateManifest)
rl_engine/executors/bridge.py:300
↓ 1 callersMethod_manifest_ipc_handles
(self, manifest: WeightUpdateManifest)
rl_engine/executors/bridge.py:280
↓ 1 callersFunction_manual_reference
The semantic definition: materialize logits, log_softmax, gather.
tests/test_linear_logp.py:162
↓ 1 callersFunction_manual_rms_norm
Independent hand-written fp32 reference (NOT the op under test).
tests/test_rms_norm.py:22
↓ 1 callersFunction_match_grad_outputs
( grad_outputs: list[torch.Tensor], outputs: list[torch.Tensor], )
rl_engine/kernels/gtest/op_checks.py:381
↓ 1 callersFunction_materialized_logp
( hidden: torch.Tensor, weight: torch.Tensor, target: torch.Tensor, bias: Optional[torch.Tenso
tests/linear_logp_tp.py:122
↓ 1 callersFunction_max_abs
(actual: torch.Tensor, expected: torch.Tensor)
tests/linear_logp_tp.py:163
↓ 1 callersFunction_max_rel
(actual: torch.Tensor, expected: torch.Tensor)
tests/linear_logp_tp.py:167
↓ 1 callersFunction_maybe_sm90_op
The Hopper TMA+MMA op, or None when unavailable (non-Hopper / not built).
benchmarks/benchmark_linear_logp.py:29
↓ 1 callersFunction_memory_gb
(device: torch.device)
benchmarks/benchmark_rl_kernels.py:123
↓ 1 callersFunction_model_config_snapshot
( model: torch.nn.Module, config: StatelessForwardConfig, )
rl_engine/executors/stateless_executor.py:583
↓ 1 callersFunction_model_with_empty_buffer
()
tests/test_weight_sync_bridge.py:45
↓ 1 callersFunction_next_pow2
(x: int)
rl_engine/kernels/ops/triton/loss/grpo_loss.py:14
↓ 1 callersMethod_next_published_weight_version
(self, consumed_weight_version: int)
rl_engine/executors/deepspeed_trainer.py:240
↓ 1 callersFunction_normalize_mapping_output
( raw_output: Mapping[str, Any], *, prompt_index: int, candidate_index: int, )
rl_engine/executors/vllm_sampler.py:250
↓ 1 callersFunction_normalize_prompts
( prompts: str | Mapping[str, Any] | Sequence[str | Mapping[str, Any]], )
rl_engine/executors/vllm_sampler.py:160
↓ 1 callersFunction_pack_tensors_for_cuda_vmm
( state_dict: Mapping[str, torch.Tensor], )
rl_engine/executors/bridge.py:1477
↓ 1 callersFunction_paged_kv_config
(args: argparse.Namespace)
benchmarks/benchmark_stateless_executor.py:307
↓ 1 callersFunction_paged_kv_correctness_notes
(stateless_result, paged_result)
benchmarks/benchmark_stateless_executor.py:364
↓ 1 callersFunction_paged_kv_row
( args: argparse.Namespace, inputs: StatelessForwardInputs, result, )
benchmarks/benchmark_stateless_executor.py:321
↓ 1 callersFunction_parse_dtype
(value: str)
scripts/check_operator.py:27
↓ 1 callersFunction_parse_dtype
(value: str)
benchmarks/benchmark_rl_kernels.py:80
↓ 1 callersFunction_parse_dtype
(value: str)
benchmarks/profiler.py:488
↓ 1 callersFunction_parse_dtype
(value: str)
benchmarks/benchmark_ratio_kl.py:75
↓ 1 callersFunction_parse_float_list
(value: str)
benchmarks/benchmark_rl_kernels.py:76
↓ 1 callersFunction_parse_float_list
(value: str)
benchmarks/benchmark_ratio_kl.py:71
↓ 1 callersFunction_parse_workloads
(value: str)
benchmarks/profiler.py:732
↓ 1 callersFunction_peak_memory_gb
(device: torch.device)
tests/linear_logp_tp.py:196
↓ 1 callersMethod_prepare_kernels
Hardware-aware operator initialization. Dynamically retrieves optimal operator objects for CUDA or ROCm environments.
rl_engine/executors/rollout.py:121
↓ 1 callersMethod_prepare_sampler
Lazily construct the vLLM-backed sampler. vLLM import and engine construction are deferred so CPU-only tests and kernel-only
rl_engine/executors/rollout.py:136
↓ 1 callersFunction_python_cuda_home_candidates
()
rl_engine/executors/deepspeed_trainer.py:349
↓ 1 callersFunction_rank_env
()
tests/linear_logp_tp.py:60
↓ 1 callersFunction_ratio_kl_row
(config: BenchmarkConfig)
benchmarks/benchmark_ratio_kl.py:130
↓ 1 callersFunction_recv_fd
(sock: socket.socket)
rl_engine/executors/bridge.py:841
↓ 1 callersFunction_reference_logp_groups_match
( candidate_groups: Sequence[Sequence[Sequence[int]]], reference_logp_groups: Sequence[Sequence[Sequen
rl_engine/executors/training_contract.py:899
↓ 1 callersFunction_reference_loss
Independent reference: logits -> selected logp -> clipped surrogate + KL.
tests/test_grpo_loss.py:67
↓ 1 callersFunction_reference_ratio_kl
Independent reference using the testing log-prob helper + mask-before-exp.
tests/test_ratio_kl.py:62
↓ 1 callersFunction_reset_peak_memory
(device: torch.device)
tests/linear_logp_tp.py:190
↓ 1 callersMethod_resolve_group_ids
( num_sequences: int, *, device: torch.device, samples_per_prompt: Optional[in
rl_engine/kernels/ops/pytorch/loss/grpo_loss.py:171
↓ 1 callersMethod_resolve_manifest_install_callable
(self)
rl_engine/executors/bridge.py:175
↓ 1 callersMethod_resolve_reduce_tensor
(self)
rl_engine/executors/bridge.py:307
↓ 1 callersMethod_resolve_reduce_tensor
(self)
rl_engine/executors/bridge.py:2615
↓ 1 callersMethod_resolve_release_model_fn
(self, update_id: str)
rl_engine/executors/bridge.py:713
↓ 1 callersMethod_resolve_reload_weights
(self)
rl_engine/executors/bridge.py:409
↓ 1 callersMethod_resolve_reload_weights
(self)
rl_engine/executors/bridge.py:516
↓ 1 callersMethod_resolve_weights_path
( self, manifest: WeightUpdateManifest, tensors: Mapping[str, torch.Tensor], )
rl_engine/executors/bridge.py:497
↓ 1 callersMethod_resolved_deepspeed_config
(self)
rl_engine/executors/deepspeed_trainer.py:302
↓ 1 callersFunction_resolved_zero_stage
(config: Mapping[str, Any], *, fallback: int)
rl_engine/executors/deepspeed_trainer.py:402
↓ 1 callersFunction_restore_model_config_snapshot
(saved: list[tuple[Any, str, Any]])
rl_engine/executors/stateless_executor.py:609
↓ 1 callersFunction_reward_groups_match
( candidate_groups: Sequence[Sequence[Sequence[int]]], reward_groups: Sequence[Sequence[float]], )
rl_engine/executors/training_contract.py:887
↓ 1 callersFunction_run
(args: argparse.Namespace)
benchmarks/benchmark_stateless_executor.py:216
↓ 1 callersFunction_run_autograd_linear_logp
(hidden, weight, target, bias, grad_out)
tests/test_linear_logp.py:207
↓ 1 callersFunction_run_cache_forward
( model: torch.nn.Module, inputs: StatelessForwardInputs, *, use_cache: bool, )
rl_engine/executors/paged_kv_baseline.py:342
↓ 1 callersFunction_run_candidate
( candidate: CandidateSpec, cases: Sequence[OperatorCase], contract: Mapping[str, Any], *,
rl_engine/kernels/gtest/op_checks.py:134
↓ 1 callersFunction_run_case
( candidate: CandidateSpec, case: OperatorCase, contract: Mapping[str, Any], )
rl_engine/kernels/gtest/op_checks.py:172
↓ 1 callersFunction_run_case_backward
( candidate: CandidateSpec, case: OperatorCase, contract: Mapping[str, Any], *, grad_mode:
rl_engine/kernels/gtest/op_checks.py:182
↓ 1 callersFunction_run_chunked_backward
(hidden, weight, target, bias, grad_out, *, chunk_elems)
tests/test_linear_logp.py:191
← previousnext →401–500 of 1,290, ranked by callers