Functions — github.com/RL-Align/RL-Kernel

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 callersFunctionnormalize_output_candidate

( raw_output: Any, *, prompt_index: int, candidate_index: int, )

rl_engine/executors/vllm_sampler.py:219

↓ 1 callersFunctionobjective_reference_logps

Return payload reference logps when available, otherwise a synthetic offset.

rl_engine/executors/training_contract.py:950

↓ 1 callersMethodonline_indexed_fp32

( self, logits: torch.Tensor, token_ids: torch.Tensor, row_indices: torch.Tensor )

rl_engine/kernels/ops/pytorch/loss/logp.py:104

↓ 1 callersFunctionoperator_names

()

rl_engine/kernels/gtest/operator_specs.py:74

↓ 1 callersMethodoptions

(self, **kwargs)

tests/test_ray_actor_manager.py:56

↓ 1 callersFunctionparse_args

()

scripts/check_operator.py:68

↓ 1 callersFunctionparse_args

()

examples/grpo_single_gpu.py:59

↓ 1 callersFunctionparse_args

()

benchmarks/benchmark_grpo_loss.py:148

↓ 1 callersFunctionparse_args

()

benchmarks/benchmark_linear_logp.py:170

↓ 1 callersFunctionparse_args

()

benchmarks/benchmark_stateless_executor.py:402

↓ 1 callersFunctionprint_attention_diff

(actual, expected, *, dtype, atol, rtol, causal, softmax_scale)

tests/test_attention_correctness.py:112

↓ 1 callersMethodprint_panel

(metrics: Dict[str, Any])

scripts/run_perf.py:19

↓ 1 callersMethodprofile_sampling

Profile a sampling operator and compute end-to-end metrics.

benchmarks/profiler.py:325

↓ 1 callersMethodpublish_weights

(self, *, weight_version, metadata=None)

tests/test_ray_actor_manager.py:141

↓ 1 callersMethodreject

(self, update_id: str, reason: str)

rl_engine/executors/bridge.py:1425

↓ 1 callersMethodreject

(self, update_id: str, reason: str)

rl_engine/executors/bridge.py:1609

↓ 1 callersMethodrelease

(self, update_id: str)

rl_engine/executors/bridge.py:166

↓ 1 callersMethodrelease

(self, update_id: str)

rl_engine/executors/bridge.py:405

↓ 1 callersMethodrelease

(self, update_id: str)

rl_engine/executors/bridge.py:493

↓ 1 callersMethodrelease_weight_update

Release a specific manifest update, active or already superseded.

rl_engine/executors/rollout.py:96

↓ 1 callersFunctionresolve_logp_op

(device: torch.device)

examples/grpo_single_gpu.py:90

↓ 1 callersMethodresolved_init_kwargs

(self)

rl_engine/executors/ray_actor_manager.py:25

↓ 1 callersFunctionrun_benchmark

(args)

benchmarks/benchmark_grpo_loss.py:73

↓ 1 callersFunctionrun_benchmark

(args, return_data: bool = False)

benchmarks/benchmark_sampling.py:43

↓ 1 callersFunctionrun_benchmark

(args)

benchmarks/benchmark_linear_logp.py:89

↓ 1 callersFunctionrun_benchmark

()

benchmarks/benchmark_attention.py:10

↓ 1 callersFunctionrun_benchmark

(args, return_data: bool = False)

benchmarks/benchmark_grpo_op.py:53

↓ 1 callersFunctionrun_correctness

(args, rank: int, world_size: int, device: torch.device, op: Any)

tests/linear_logp_tp.py:228

↓ 1 callersFunctionrun_stress

(args, rank: int, world_size: int, device: torch.device, op: Any)

tests/linear_logp_tp.py:336

↓ 1 callersFunctionrun_training

(args: argparse.Namespace)

examples/grpo_single_gpu.py:163

↓ 1 callersFunctionsdpa_math_backend

()

tests/test_attention_correctness.py:35

↓ 1 callersFunctionselect_device

(requested: str)

examples/grpo_single_gpu.py:82

↓ 1 callersMethodselected_logprobs

( self, input_ids: torch.Tensor, token_ids: torch.Tensor, *, mask: Opt

rl_engine/alignment/model_wrappers.py:138

↓ 1 callersFunctionselected_logps_with_op

( logp_op: Any, logits: torch.Tensor, token_ids: torch.Tensor, mask: torch.Tensor, )

examples/grpo_single_gpu.py:131

↓ 1 callersFunctionshould_print_attention_diff

()

tests/test_attention_correctness.py:53

↓ 1 callersFunctionstep

()

tests/linear_logp_tp.py:378

↓ 1 callersMethodsupports_posix_fd_vmm

(self)

rl_engine/executors/bridge.py:1185

↓ 1 callersFunctiontest_accuracy

()

tests/test_op_accuracy.py:143

↓ 1 callersFunctiontest_device_and_registry

()

rl_engine/tests/test_dispatch.py:20

↓ 1 callersFunctiontest_executor_flow

()

rl_engine/tests/test_dispatch.py:45

↓ 1 callersFunctiontest_hal_routing_and_fallback

()

tests/test_cpu_hal.py:10

↓ 1 callersFunctiontest_logger_enhancements

()

rl_engine/tests/test_dispatch.py:12

↓ 1 callersMethodto_metadata

(self)

rl_engine/executors/bridge.py:72

↓ 1 callersMethodto_ray_options

(self)

rl_engine/executors/ray_actor_manager.py:45

↓ 1 callersMethodto_tensor

(self)

rl_engine/executors/bridge.py:972

↓ 1 callersFunctiontop_k_renorm_probs

(probs, top_k)

tests/test_sampler_temperature.py:21

↓ 1 callersFunctiontop_p_sampling_from_probs

(probs, top_p, deterministic=True)

tests/test_sampler_temperature.py:24

↓ 1 callersMethodupdate_weights

(self, manifest)

tests/test_ray_actor_manager.py:160

↓ 1 callersMethodupdate_weights_via_ipc

Backward-compatible IPC entry point. Raw CUDA IPC handle imports are intentionally unavailable until issue #13 validates CUD

rl_engine/executors/rollout.py:105

↓ 1 callersFunctionvisit

(node: Any)

rl_engine/executors/stateless_executor.py:416

MethodCUmemAllocFlags

(self)

rl_engine/executors/bridge.py:1094

FunctionPYBIND11_MODULE

PyBind11 Module Registration

csrc/ops.cpp:75

Method__call__

( self, manifest: WeightUpdateManifest, tensors: Mapping[str, torch.Tensor], )

rl_engine/executors/bridge.py:215

Method__call__

( self, policy_logits: torch.Tensor, ref_logits: torch.Tensor, action_ids: tor

rl_engine/kernels/ops/triton/loss/ratio_kl.py:186

Method__call__

( self, hidden: torch.Tensor, lm_head_weight: torch.Tensor, target_ids: torch.

rl_engine/kernels/ops/triton/loss/linear_logp.py:166

Method__call__

( self, policy_logits: torch.Tensor, ref_logits: torch.Tensor, action_ids: tor

rl_engine/kernels/ops/triton/loss/grpo_loss.py:72

Method__call__

Standard attention forward pass. Args: q: (batch, seqlen, nheads, headdim) k: (batch, seqlen, nheads_k, headd

rl_engine/kernels/ops/rocm/attention/flash_attn.py:43

Method__call__

( self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, dropout_p:

rl_engine/kernels/ops/pytorch/attention/__init__.py:11

Method__call__

Alias for ``forward`` so the op is callable like a module.

rl_engine/kernels/ops/pytorch/attention/kv_cache.py:54

Method__call__

Alias for ``forward`` so the op is callable like a module.

rl_engine/kernels/ops/pytorch/attention/standard_attn.py:46

Method__call__

Alias for ``forward`` so the op is callable like a module.

rl_engine/kernels/ops/pytorch/linear/lm_head.py:37

Method__call__

( self, x: torch.Tensor, weight: torch.Tensor, *, eps: float = 1e-6,

rl_engine/kernels/ops/pytorch/norm/rms_norm.py:18

Method__call__

( self, policy_logits: torch.Tensor, ref_logits: torch.Tensor, action_ids: tor

rl_engine/kernels/ops/pytorch/loss/ratio_kl.py:19

Method__call__

(self, logits: torch.Tensor, token_ids: torch.Tensor)

rl_engine/kernels/ops/pytorch/loss/logp.py:17

Method__call__

( self, hidden: torch.Tensor, lm_head_weight: torch.Tensor, target_ids: torch.

rl_engine/kernels/ops/pytorch/loss/linear_logp.py:455

Method__call__

( self, policy_logits: torch.Tensor, ref_logits: torch.Tensor, action_ids: tor

rl_engine/kernels/ops/pytorch/loss/grpo_loss.py:24

Method__call__

(self, x: Tensor, positions: Tensor, *, theta: float = 1_000_000.0)

rl_engine/kernels/ops/pytorch/rotary_embedding/rope.py:24

Method__call__

Standard attention forward pass. Args: q: (batch, seqlen, nheads, headdim) k: (batch, seqlen, nheads_k, headd

rl_engine/kernels/ops/cuda/attention/flash_attn.py:36

Method__call__

Prefix-shared attention forward pass. Args: q: Query tensor of shape [bs, G, seq_len_q, head_dim] k: Shared

rl_engine/kernels/ops/cuda/attention/prefix_shared_attn.py:34

Method__call__

(self, logits: torch.Tensor, labels: torch.Tensor)

rl_engine/kernels/ops/cuda/loss/logp.py:25

Method__call__

(self, logits: torch.Tensor, token_ids: torch.Tensor)

rl_engine/kernels/ops/cuda/loss/logp.py:102

Method__call__

( self, hidden: torch.Tensor, lm_head_weight: torch.Tensor, target_ids: torch.

rl_engine/kernels/ops/cuda/loss/linear_logp.py:268

Method__call__

(self, logits: torch.Tensor, token_ids: torch.Tensor)

rl_engine/kernels/gtest/operator_specs.py:67

Method__call__

(self, *args, **kwargs)

tests/test_deepspeed_training_worker.py:31

Method__call__

(self, hidden, lm_head_weight, target_ids, bias=None, **kwargs)

tests/test_deepspeed_training_worker.py:122

Method__enter__

(self)

rl_engine/executors/ray_actor_manager.py:88

Method__enter__

(self)

tests/test_deepspeed_training_worker.py:73

Method__exit__

(self, exc_type, exc, traceback)

rl_engine/executors/ray_actor_manager.py:92

Method__exit__

(self, exc_type, exc, traceback)

tests/test_deepspeed_training_worker.py:79

Method__getattr__

(self, name)

tests/test_ray_actor_manager.py:46

Method__getitem__

(cls, name: str)

rl_engine/kernels/registry.py:16

Method__init__

( self, engine: Any, *, install_callable: Optional[Any] = None, reques

rl_engine/executors/bridge.py:121

Method__init__

( self, *, is_checkpoint_format: bool = True, reduce_tensor_fn: Optional[Any]

rl_engine/executors/bridge.py:203

Method__init__

( self, engine: Any, *, target_dtype: Optional[torch.dtype] = None, ta

rl_engine/executors/bridge.py:339

Method__init__

( self, engine: Any, *, weights_path: Optional[str] = None, weights_pa

rl_engine/executors/bridge.py:447

Method__init__

( self, engine: Any, *, device_index: int = 0, source_worker: str = "v

rl_engine/executors/bridge.py:627

Method__init__

(self, *, device_index: int = 0)

rl_engine/executors/bridge.py:986

Method__init__

(self, *, source_worker: str = "local-training", source_rank: int = 0)

rl_engine/executors/bridge.py:1519

Method__init__

(self, *, source_worker: str = "local-training", source_rank: int = 0)

rl_engine/executors/bridge.py:1708

Method__init__

( self, *, source_worker: str = "cuda-training", source_rank: int = 0,

rl_engine/executors/bridge.py:2007

Method__init__

( self, *, source_worker: str = "cuda-training", source_rank: int = 0,

rl_engine/executors/bridge.py:2385

Method__init__

( self, vocab_size: int, hidden_dim: int, *, bias: bool = True,

rl_engine/executors/deepspeed_trainer.py:55

Method__init__

( self, config: Optional[DeepSpeedTrainingConfig] = None, *, weight_bridge: Op

rl_engine/executors/deepspeed_trainer.py:83

Method__init__

( self, model: torch.nn.Module, config: Optional[StatelessForwardConfig] = None,

rl_engine/executors/stateless_executor.py:102

Method__init__

( self, model: torch.nn.Module, config: Optional[PagedKVScoringConfig] = None,

rl_engine/executors/paged_kv_baseline.py:97

Method__init__

( self, executor: StatelessForwardExecutor, collate_inputs: Callable[[RolloutStageResu

rl_engine/executors/training_contract.py:319

Method__init__

( self, config: VLLMSamplerConfig, *, engine: Optional[Any] = None, ll

rl_engine/executors/vllm_sampler.py:80

Method__init__

( self, runtime_config: Optional[RayRuntimeConfig] = None, *, ray_module: Any

rl_engine/executors/ray_actor_manager.py:78

Method__init__

(self, actor: Any, ray_module: Any)

rl_engine/executors/ray_actor_manager.py:141

Method__init__

(self, actor: Any, ray_module: Any)

rl_engine/executors/ray_actor_manager.py:158

← previous next →601–700 of 1,290, ranked by callers