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 callersFunction_run_cuda_ipc
()
benchmarks/benchmark_weight_sync_bridge.py:262
↓ 1 callersFunction_run_cuda_vmm
(args: argparse.Namespace)
benchmarks/benchmark_weight_sync_bridge.py:432
↓ 1 callersFunction_run_cuda_vmm_rollout_update
(args: argparse.Namespace)
benchmarks/benchmark_weight_sync_bridge.py:572
↓ 1 callersFunction_run_local
(args: argparse.Namespace)
benchmarks/benchmark_weight_sync_bridge.py:89
↓ 1 callersFunction_run_paged_kv_only
(args: argparse.Namespace)
benchmarks/benchmark_stateless_executor.py:286
↓ 1 callersFunction_run_rollout_update
(args: argparse.Namespace)
benchmarks/benchmark_weight_sync_bridge.py:511
↓ 1 callersFunction_run_runtime_blockers
()
benchmarks/benchmark_weight_sync_bridge.py:919
↓ 1 callersFunction_run_shared_memory
(args: argparse.Namespace)
benchmarks/benchmark_weight_sync_bridge.py:164
↓ 1 callersFunction_run_vllm_cuda_ipc_hot_update
(args: argparse.Namespace)
benchmarks/benchmark_weight_sync_bridge.py:757
↓ 1 callersFunction_run_vllm_cuda_vmm_external_storage
(args: argparse.Namespace)
benchmarks/benchmark_weight_sync_bridge.py:659
↓ 1 callersFunction_sampling_fn
( *, batch_size: int, vocab_size: int, dtype: torch.dtype, device: torch.device, seed:
benchmarks/profiler.py:560
↓ 1 callersFunction_select_device
(value: str)
scripts/check_operator.py:38
↓ 1 callersFunction_select_flash_attn_backend
Select the installed FlashAttention ROCm backend.
rl_engine/kernels/ops/rocm/attention/flash_attn.py:13
↓ 1 callersFunction_selected_logprob_row
(config: BenchmarkConfig)
benchmarks/benchmark_rl_kernels.py:129
↓ 1 callersFunction_send_fd
(sock: socket.socket, fd: int)
rl_engine/executors/bridge.py:832
↓ 1 callersFunction_should_fallback_attention_backend
( exc: Exception, config: StatelessForwardConfig, )
rl_engine/executors/stateless_executor.py:496
↓ 1 callersFunction_single_threaded_torch
()
tests/test_matmul.py:20
↓ 1 callersFunction_slice_logits
( logits: torch.Tensor, *, logits_start: Optional[int] = None, logits_end: Optional[int] = Non
rl_engine/alignment/model_wrappers.py:24
↓ 1 callersFunction_sm90_available
SM90 forward needs a Hopper GPU and the kernel compiled into the extension.
tests/test_linear_logp.py:33
↓ 1 callersFunction_sm90_logp_available
()
tests/test_op_accuracy.py:17
↓ 1 callersFunction_sm90_tensor_parallel_linear_logp
( hidden: torch.Tensor, lm_head_weight: torch.Tensor, target_ids: torch.Tensor, bias: Optional
rl_engine/kernels/ops/cuda/loss/linear_logp.py:227
↓ 1 callersMethod_snapshot_to_shared_memory
( self, model: torch.nn.Module, )
rl_engine/executors/bridge.py:1770
↓ 1 callersMethod_start_fd_broker
(self, update_id: str, exported_fd: int)
rl_engine/executors/bridge.py:2243
↓ 1 callersMethod_strict_fp32_math
Disable autocast and TF32 for a true fp32 path, restoring state after.
rl_engine/kernels/ops/pytorch/attention/standard_attn.py:183
↓ 1 callersMethod_strict_fp32_matmul
Disable autocast and TF32 for a true fp32 matmul, restoring state after.
rl_engine/kernels/ops/pytorch/linear/lm_head.py:116
↓ 1 callersFunction_summarize
(report: Any)
scripts/check_operator.py:47
↓ 1 callersFunction_tensor_nbytes
(tensor: torch.Tensor)
rl_engine/executors/stateless_executor.py:635
↓ 1 callersFunction_tensor_parallel_world_size
(tp_group: Any)
rl_engine/kernels/ops/pytorch/loss/linear_logp.py:51
↓ 1 callersFunction_tensor_sha256
(tensor: torch.Tensor)
rl_engine/executors/bridge.py:790
↓ 1 callersFunction_time_block
(device: torch.device, fn)
tests/linear_logp_tp.py:202
↓ 1 callersFunction_tiny_bert_reward_model
()
tests/test_stateless_hf_integration.py:69
↓ 1 callersFunction_unwrap_training_model
(engine: Any, fallback_model: torch.nn.Module)
rl_engine/executors/deepspeed_trainer.py:444
↓ 1 callersFunction_validate_inputs
(inputs: StatelessForwardInputs, config: StatelessForwardConfig)
rl_engine/executors/stateless_executor.py:440
↓ 1 callersFunction_validate_model_input_token_ids
(token_ids: torch.Tensor, *, vocab_size: int)
rl_engine/executors/deepspeed_trainer.py:572
↓ 1 callersFunction_write_csv_row
(row: Mapping[str, Any], output: Path)
benchmarks/benchmark_weight_sync_bridge.py:66
↓ 1 callersFunction_write_json
(path: Path | None, payload: dict[str, Any])
benchmarks/benchmark_stateless_executor.py:159
↓ 1 callersFunction_write_rows
(rows: list[dict[str, Any]], output: Path | None)
benchmarks/benchmark_rl_kernels.py:349
↓ 1 callersFunction_write_rows
(rows: list[dict[str, Any]], output: Path | None)
benchmarks/benchmark_ratio_kl.py:231
↓ 1 callersMethodapply
( self, hidden: torch.Tensor, lm_head_weight: torch.Tensor, target_ids: torch.
rl_engine/kernels/ops/triton/loss/linear_logp.py:187
↓ 1 callersMethodapply
Backward-compatible alias for forward.
rl_engine/kernels/ops/pytorch/loss/logp.py:60
↓ 1 callersMethodapply
(self, logits: torch.Tensor, labels: torch.Tensor)
rl_engine/kernels/ops/cuda/loss/logp.py:41
↓ 1 callersMethodapply
(self, logits: torch.Tensor, token_ids: torch.Tensor)
rl_engine/kernels/ops/cuda/loss/logp.py:126
↓ 1 callersMethodapply
(self, logits_arg, token_ids_arg)
tests/test_op_accuracy.py:210
↓ 1 callersFunctionattach_stateless_scores_to_payload
Return a rollout payload augmented with stateless scoring outputs.
rl_engine/executors/training_contract.py:351
↓ 1 callersFunctionbuild_arg_parser
()
benchmarks/benchmark_rl_kernels.py:365
↓ 1 callersFunctionbuild_arg_parser
()
benchmarks/benchmark_ratio_kl.py:247
↓ 1 callersFunctionbuild_parser
()
tests/linear_logp_tp.py:426
↓ 1 callersFunctionbuild_script_parser
Extend the core profiler CLI with report-management flags.
scripts/run_profile_suite.py:33
↓ 1 callersFunctioncollect_paged_kv_metrics
( inputs: StatelessForwardInputs, reservation: PagedKVCacheReservation, *, config: PagedKVScor
rl_engine/executors/paged_kv_baseline.py:260
↓ 1 callersFunctioncollect_stateless_metrics
Collect common scoring metrics without importing optional runtimes.
rl_engine/executors/stateless_executor.py:295
↓ 1 callersMethodcompact_completion_values
(self, values: torch.Tensor)
rl_engine/testing/rl_batch.py:63
↓ 1 callersMethodcompact_token_ids
(self)
rl_engine/testing/rl_batch.py:67
↓ 1 callersMethodcompute_logp
Pre-allocated block logic for computing log probabilities of selected tokens.
rl_engine/kernels/sampling.py:81
↓ 1 callersFunctionconfigure_stateless_model
Apply best-effort no-cache scoring knobs without importing optional runtimes. Hugging Face-style models decide attention kernels from their
rl_engine/executors/stateless_executor.py:335
↓ 1 callersMethodcopy_tensor_to_address
( self, destination_address: int, tensor: torch.Tensor, *, stream: Opt
rl_engine/executors/bridge.py:1302
↓ 1 callersMethodcreate_allocation
(self, requested_nbytes: int)
rl_engine/executors/bridge.py:1207
↓ 1 callersFunctioncuda_flash_attention_availability
()
tests/test_attention_correctness.py:135
↓ 1 callersFunctiondefault_reward_adapter
Default adapter for common scalar reward model outputs.
rl_engine/executors/stateless_executor.py:278
↓ 1 callersMethoddense_completion_values
(self, values: torch.Tensor)
rl_engine/testing/rl_batch.py:55
↓ 1 callersMethodengine
(self)
rl_engine/executors/vllm_sampler.py:94
↓ 1 callersMethodestimate_sampling_tflops
Estimate FLOPs for a single sampling step. softmax ~= 5 * B * V FLOPs top-k and multinomial are comparatively negligible.
benchmarks/profiler.py:238
↓ 1 callersMethodexecute_rollout
Execute sampling using optimized fused kernels. Solves the O(G * L * V) memory wall for GRPO rollout.
rl_engine/executors/rollout.py:172
↓ 1 callersMethodexpand_advantages
Broadcast per-sequence advantages to per-token, zeroing masked tokens.
rl_engine/kernels/ops/triton/loss/grpo_loss.py:168
↓ 1 callersMethodexpand_advantages
Broadcast per-sequence advantages to per-token, zeroing masked tokens.
rl_engine/kernels/ops/pytorch/loss/grpo_loss.py:87
↓ 1 callersMethodforward
Canonical entry: gather in the weight's native dtype, then cast the gathered rows to weight.dtype (a no-op here, kept for symmetry).
rl_engine/kernels/ops/pytorch/linear/embedding.py:23
↓ 1 callersMethodforward
Canonical entry: compute in fp32, cast the result back to x.dtype. This is the dtype-behavior path used as the Axis-B accuracy candid
rl_engine/kernels/ops/pytorch/activation/swiglu.py:23
↓ 1 callersMethodforward
Canonical entry: compute in fp32, cast the result back to gate.dtype. This is the dtype-behavior path used as the Axis-B accuracy can
rl_engine/kernels/ops/pytorch/activation/swiglu.py:58
↓ 1 callersMethodforward
( self, policy_logits: torch.Tensor, ref_logits: torch.Tensor, action_ids: tor
rl_engine/kernels/ops/pytorch/loss/ratio_kl.py:32
↓ 1 callersMethodforward
(self, input_ids: torch.Tensor, **model_kwargs: Any)
rl_engine/alignment/model_wrappers.py:84
↓ 1 callersMethodfreeze
(self)
rl_engine/alignment/model_wrappers.py:129
↓ 1 callersFunctionget_cmdclass
()
setup.py:145
↓ 1 callersFunctionget_extensions
()
setup.py:34
↓ 1 callersMethodgranularity
(self)
rl_engine/executors/bridge.py:1197
↓ 1 callersFunctiongrpo_loss
( current_logps: torch.Tensor, old_logps: torch.Tensor, ref_logps: torch.Tensor, advantages: t
examples/grpo_single_gpu.py:144
↓ 1 callersMethodimport_allocation
(self, fd: int, mapped_nbytes: int)
rl_engine/executors/bridge.py:1263
↓ 1 callersMethodimport_model_weights
(self, ipc_handles: Mapping[str, Any])
rl_engine/executors/bridge.py:2522
↓ 1 callersMethodimport_update
(self, manifest: WeightUpdateManifest)
rl_engine/executors/bridge.py:1421
↓ 1 callersMethodimport_update
(self, incoming_manifest)
tests/test_weight_sync_bridge.py:887
↓ 1 callersFunctioninit_logger
Initializes logger and patches extended methods. Ensures consistent logging style across the entire engine.
rl_engine/utils/logger.py:52
↓ 1 callersMethodinitialize
(self, **kwargs)
tests/test_deepspeed_training_worker.py:53
↓ 1 callersFunctioninstall
(model: torch.nn.Module)
rl_engine/executors/bridge.py:550
↓ 1 callersMethodinstall
( self, manifest: WeightUpdateManifest, tensors: Mapping[str, torch.Tensor], )
rl_engine/executors/bridge.py:1435
↓ 1 callersFunctionis_cuda_platform
()
tests/test_attention_correctness.py:127
↓ 1 callersFunctionis_fused_logp_backend
(backend_name: str)
examples/grpo_single_gpu.py:101
↓ 1 callersFunctionis_rocm_platform
()
tests/test_attention_correctness.py:131
↓ 1 callersMethodkill
(self, actor, no_restart=True)
tests/test_ray_actor_manager.py:92
↓ 1 callersFunctionmain
()
scripts/run_profile_suite.py:62
↓ 1 callersFunctionmain
()
scripts/run_perf.py:46
↓ 1 callersFunctionmain
()
scripts/check_rocm_env.py:14
↓ 1 callersFunctionmain
()
scripts/check_operator.py:108
↓ 1 callersFunctionmain
()
tests/linear_logp_tp.py:450
↓ 1 callersFunctionmain
()
examples/grpo_single_gpu.py:291
↓ 1 callersFunctionmain
()
benchmarks/benchmark_weight_sync_bridge.py:956
↓ 1 callersFunctionmain
()
benchmarks/benchmark_rl_kernels.py:399
↓ 1 callersFunctionmain
()
benchmarks/profiler.py:789
↓ 1 callersFunctionmain
()
benchmarks/benchmark_ratio_kl.py:267
↓ 1 callersFunctionmain
()
benchmarks/benchmark_stateless_executor.py:472
↓ 1 callersFunctionmake_candidate
(args: argparse.Namespace)
rl_engine/kernels/gtest/operator_specs.py:94
↓ 1 callersFunctionmake_group_advantages
( batch_size: int, completion_len: int, samples_per_prompt: int, completion_mask: torch.Tensor
examples/grpo_single_gpu.py:105
↓ 1 callersFunctionmake_operator_case
( args: argparse.Namespace, dtype: torch.dtype, device: torch.device )
rl_engine/kernels/gtest/operator_specs.py:78
← previousnext →501–600 of 1,290, ranked by callers