[Flink-ddd]

Summary

This PR reorganizes RL-Kernel so model code, backend implementations, reusable operators, framework adapters, contracts, configuration, and validation utilities have clear ownership.

Public imports, backend IDs, configuration semantics, and validated CUDA/ROCm behavior remain compatible.

Repository structure

The directory-only map below shows the main ownership boundaries. Braces group sibling directories; the complete tracked hierarchy follows in expandable sections.

RL-Kernel/
├── rl_engine/
│   ├── models/                          # Model specifications and assembly
│   │   ├── qwen3/                       # Existing Qwen3 Dense implementation
│   │   └── {deepseek_v4,gemma,minimax_h3}/ # Reserved model packages
│   ├── layers/                          # Reusable compositions of operators
│   ├── ops/                             # Semantic interfaces, shared across models
│   │   ├── {attention,gemm,ffn,norm,activation,rope}/
│   │   └── {embedding,moe,logprob,loss,sampling,packing,autograd}/
│   ├── contracts/                       # Numerical requirements and diagnostics
│   │   ├── {operators,diagnostics}/
│   │   └── profiles/{precision,invariance}/
│   ├── reference/                       # Mathematical references and retained aliases
│   ├── platforms/                       # Device discovery and capabilities
│   ├── backends/                        # Hardware-specific implementations
│   │   ├── cuda/{attention,gemm,ffn,logprob}/
│   │   ├── rocm/{attention,gemm,ffn,logprob}/
│   │   ├── musa/{attention,gemm,logprob}/
│   │   ├── ascend/{attention,gemm,logprob}/
│   │   ├── cpu/{attention,gemm,logprob}/
│   │   └── shared/triton/               # Implementations shared across eligible devices
│   ├── runtime/                         # Dispatch, policies, execution, and provenance
│   │   ├── provenance/
│   │   └── {selector,performance}/      # Reserved selection-policy ownership
│   ├── distributed/
│   │   ├── algorithms/                  # Partitioning and canonical reduction order
│   │   └── transports/                  # Communication bindings
│   ├── integrations/
│   │   ├── common/                      # Shared integration logic and state
│   │   ├── engines/
│   │   │   ├── rollout/vllm/            # Rollout hooks and runtime binding
│   │   │   └── train/{megatron,deepspeed}/ # Training adapters
│   │   └── orchestrators/
│   │       ├── vime/{providers,experiments,patches}/
│   │       └── {miles,areal}/            # Reserved framework integrations
│   ├── validation/                      # Reusable train-rollout consistency harnesses
│   │   ├── operators/                   # WS1 operator accuracy and invariance
│   │   ├── distributed/                 # WS2 distributed validation
│   │   ├── {models,cross_config,ablation}/
│   │   └── {common,fixtures,reports}/
│   ├── config/workloads/                # Packaged workload definitions
│   ├── {entrypoints,utils}/
│   └── {alignment,executors,kernels,testing}/ # Legacy import compatibility
├── csrc/                                # Native sources and extension bindings
│   ├── {bindings,common}/
│   └── {cuda,rocm,musa,ascend}/
├── tests/                               # Mirrors implementation and validation ownership
│   ├── {models,layers,ops,contracts,reference}/
│   ├── {backends,platforms,runtime}/
│   ├── distributed/{collectives,transports,tp,cp,ep,dp}/
│   ├── integrations/
│   │   ├── common/
│   │   ├── engines/rollout/vllm/
│   │   ├── engines/train/{megatron,deepspeed}/
│   │   └── orchestrators/{vime,miles,areal}/
│   ├── validation/{operators,distributed,models,cross_config,ablation,common}/
│   └── {benchmarks,e2e,entrypoints,build,helpers,data}/
├── benchmarks/                          # Performance measurements and tuning
│   ├── {operators,layers,backends,distributed,models,e2e}/
│   └── {common,tuning,profiling,results}/
├── configs/
│   ├── {models,hardware,policies,workloads,ablations,benchmarks,tuning,local}/
│   ├── suites/{ws1,ws2,e2e}/
│   ├── experiments/{cross_config,vime}/
│   └── integrations/
│       ├── engines/{rollout,train}/
│       └── orchestrators/vime/
├── tools/{validation,weights,env,checks,benchmarking,migration}/
├── bin/                                 # Checkout command launchers
├── build_tools/                         # Build selection and compiler configuration
├── ci/{scripts,jobs,runners,providers}/  # Reusable CI execution and configuration
├── .github/workflows/                   # GitHub CI orchestration
├── docker/                              # Container definitions
├── requirements/{integrations,constraints}/ # Dependencies and reserved categories
├── docs/
│   ├── {architecture,contributing,contracts,integrations,validation}/
│   └── {getting_started,usage,operators,benchmarking,api,cli,archive}/
├── examples/{operators,post_training}/
├── reports/{experiments,releases,archive}/
├── scripts/                             # Legacy command compatibility
└── .claude/                             # Retained development skills

Model packages declare topology and operator needs. Shared operator semantics belong in ops and contracts; each platform owns its attention, GEMM, and FFN implementations under backends. Native kernels live in csrc. Cross-platform Triton implementations use backends/shared/triton, while ROCm-only Triton FFN stays in backends/rocm/ffn.

Engine-specific hooks belong in integrations/engines/rollout or integrations/engines/train. Post-training coordination belongs in orchestrators. WS1, WS2, model consistency, and ablation checks reuse these implementations through validation, with corresponding coverage under tests.

Reserved packages define where future work belongs; their presence does not establish model or hardware support. The selector and performance packages reserve space for future selection policies. Compatibility paths remain available for existing imports; new code belongs in the owning layers above.

Full directory inventory

The following groups include all 443 tracked directories at ca74261. Files and untracked build/cache outputs are omitted.

Library packages and compatibility paths (211 directories)

RL-Kernel/
└── rl_engine/
    ├── alignment/  # Legacy compatibility imports
    │   ├── cross_config/
    │   │   └── adapters/
    │   └── testing/
    │       └── smoke_ops/
    ├── backends/
    │   ├── ascend/
    │   │   ├── activation/
    │   │   ├── attention/
    │   │   ├── embedding/
    │   │   ├── gemm/
    │   │   ├── logprob/
    │   │   ├── loss/
    │   │   ├── moe/
    │   │   ├── norm/
    │   │   ├── packing/
    │   │   ├── rope/
    │   │   └── sampling/
    │   ├── cpu/
    │   │   ├── activation/
    │   │   ├── attention/
    │   │   ├── embedding/
    │   │   ├── gemm/
    │   │   ├── logprob/
    │   │   ├── loss/
    │   │   ├── moe/
    │   │   ├── norm/
    │   │   ├── packing/
    │   │   ├── rope/
    │   │   └── sampling/
    │   ├── cuda/
    │   │   ├── activation/
    │   │   ├── attention/
    │   │   ├── embedding/
    │   │   ├── ffn/
    │   │   ├── gemm/
    │   │   ├── logprob/
    │   │   ├── loss/
    │   │   ├── moe/
    │   │   ├── norm/
    │   │   ├── packing/
    │   │   ├── rope/
    │   │   └── sampling/
    │   ├── musa/
    │   │   ├── activation/
    │   │   ├── attention/
    │   │   ├── embedding/
    │   │   ├── gemm/
    │   │   ├── logprob/
    │   │   ├── loss/
    │   │   ├── moe/
    │   │   ├── norm/
    │   │   ├── packing/
    │   │   ├── rope/
    │   │   └── sampling/
    │   ├── rocm/
    │   │   ├── activation/
    │   │   ├── attention/
    │   │   ├── embedding/
    │   │   ├── ffn/
    │   │   ├── gemm/
    │   │   ├── logprob/
    │   │   ├── loss/
    │   │   ├── moe/
    │   │   ├── norm/
    │   │   ├── packing/
    │   │   ├── rope/
    │   │   └── sampling/
    │   └── shared/
    │       └── triton/
    │           ├── activation/
    │           ├── attention/
    │           ├── embedding/
    │           ├── ffn/  # Compatibility import for the ROCm Triton FFN
    │           ├── gemm/
    │           ├── logprob/
    │           ├── loss/
    │           ├── norm/
    │           └── rope/
    ├── config/
    │   └── workloads/
    ├── contracts/
    │   ├── diagnostics/
    │   ├── operators/
    │   └── profiles/
    │       ├── invariance/
    │       └── precision/
    ├── distributed/
    │   ├── algorithms/
    │   └── transports/
    ├── entrypoints/
    ├── executors/  # Legacy compatibility imports
    ├── integrations/
    │   ├── common/
    │   ├── engines/
    │   │   ├── rollout/
    │   │   │   └── vllm/
    │   │   └── train/
    │   │       ├── deepspeed/
    │   │       └── megatron/
    │   ├── orchestrators/
    │   │   ├── areal/  # Reserved orchestrator package
    │   │   ├── miles/  # Reserved orchestrator package
    │   │   └── vime/
    │   │       ├── experiments/
    │   │       │   ├── dense/
    │   │       │   │   ├── companion_patches/
    │   │       │   │   └── profiles/
    │   │       │   ├── rocm_ablation/
    │   │       │   └── rocm_attention/
    │   │       │       ├── companion_patches/
    │   │       │       └── profiles/
    │   │       ├── patches/
    │   │       │   ├── dense/
    │   │       │   └── rocm_attention/
    │   │       └── providers/
    │   └── vime/  # Legacy compatibility imports
    ├── kernels/  # Legacy compatibility imports
    │   ├── gtest/
    │   └── ops/
    │       ├── ascend/
    │       │   ├── activation/
    │       │   ├── attention/
    │       │   ├── linear/
    │       │   ├── loss/
    │       │   ├── matmul/
    │       │   ├── norm/
    │       │   └── rotary_embedding/
    │       ├── cuda/
    │       │   ├── activation/
    │       │   ├── attention/
    │       │   ├── linear/
    │       │   ├── loss/
    │       │   ├── matmul/
    │       │   ├── norm/
    │       │   └── rotary_embedding/
    │       ├── matmul/
    │       ├── pytorch/
    │       │   ├── activation/
    │       │   ├── attention/
    │       │   ├── ffn/
    │       │   ├── linear/
    │       │   ├── loss/
    │       │   ├── matmul/
    │       │   ├── norm/
    │       │   ├── packing/
    │       │   └── rotary_embedding/
    │       ├── rocm/
    │       │   ├── attention/
    │       │   ├── loss/
    │       │   ├── matmul/
    │       │   └── rotary_embedding/
    │       └── triton/
    │           ├── activation/
    │           ├── attention/
    │           ├── ffn/
    │           ├── linear/
    │           ├── loss/
    │           ├── matmul/
    │           └── rotary_embedding/
    ├── layers/
    ├── models/
    │   ├── deepseek_v4/  # Reserved model package
    │   ├── gemma/  # Reserved model package
    │   ├── minimax_h3/  # Reserved model package
    │   └── qwen3/
    ├── ops/
    │   ├── activation/
    │   ├── attention/
    │   ├── autograd/
    │   │   └── musa/
    │   ├── embedding/
    │   ├── ffn/
    │   ├── gemm/
    │   ├── logprob/
    │   ├── loss/
    │   ├── moe/
    │   ├── norm/
    │   ├── packing/
    │   ├── rope/
    │   └── sampling/
    ├── platforms/
    ├── reference/
    │   ├── activation/
    │   ├── attention/
    │   ├── embedding/
    │   ├── ffn/  # Compatibility import for the platform-owned FFN
    │   ├── gemm/
    │   ├── logprob/
    │   ├── loss/
    │   ├── norm/
    │   ├── packing/
    │   └── rope/
    ├── runtime/
    │   ├── performance/  # Reserved selection-policy ownership
    │   ├── provenance/
    │   └── selector/  # Reserved selection-policy ownership
    ├── testing/  # Legacy compatibility imports
    ├── utils/
    └── validation/
        ├── ablation/
        ├── common/
        ├── cross_config/
        │   └── adapters/
        ├── distributed/
        ├── fixtures/
        │   └── smoke_ops/
        ├── models/
        ├── operators/
        └── reports/

Tests (62 directories)

RL-Kernel/
└── tests/
    ├── backends/
    │   ├── ascend/
    │   ├── cuda/
    │   ├── musa/
    │   └── rocm/
    ├── benchmarks/
    ├── build/
    ├── contracts/
    │   └── operators/
    ├── data/
    ├── distributed/
    │   ├── collectives/
    │   ├── cp/
    │   ├── dp/
    │   ├── ep/
    │   ├── tp/
    │   └── transports/
    ├── e2e/
    ├── entrypoints/
    ├── helpers/
    ├── integrations/
    │   ├── common/
    │   ├── engines/
    │   │   ├── rollout/
    │   │   │   └── vllm/
    │   │   │       └── distributed/
    │   │   └── train/
    │   │       ├── deepspeed/
    │   │       └── megatron/
    │   └── orchestrators/
    │       ├── areal/
    │       ├── miles/
    │       └── vime/
    ├── layers/
    ├── models/
    │   ├── deepseek_v4/
    │   ├── gemma/
    │   ├── minimax_h3/
    │   └── qwen3/
    ├── ops/
    │   ├── activation/
    │   ├── attention/
    │   ├── autograd/
    │   ├── embedding/
    │   ├── gemm/
    │   ├── logprob/
    │   ├── loss/
    │   ├── norm/
    │   ├── packing/
    │   ├── rope/
    │   └── sampling/
    ├── platforms/
    ├── reference/
    ├── runtime/
    └── validation/
        ├── ablation/
        ├── common/
        ├── cross_config/
        ├── distributed/
        ├── models/
        └── operators/

Native sources, benchmarks, configuration, build, CI, and tools (94 directories)

RL-Kernel/
├── .github/
│   └── workflows/
├── benchmarks/
│   ├── backends/
│   │   └── rocm/
│   ├── common/
│   ├── distributed/
│   ├── e2e/
│   ├── layers/
│   ├── models/
│   ├── operators/
│   │   ├── attention/
│   │   ├── gemm/
│   │   ├── logprob/
│   │   ├── loss/
│   │   ├── norm/
│   │   ├── packing/
│   │   ├── rope/
│   │   ├── sampling/
│   │   └── suite/
│   ├── profiling/
│   ├── results/
│   │   └── qwen_ffn_h100_trace/
│   └── tuning/
├── bin/
├── build_tools/
├── ci/
│   ├── jobs/
│   ├── providers/
│   ├── runners/
│   └── scripts/
├── configs/
│   ├── ablations/
│   ├── benchmarks/
│   ├── experiments/
│   │   ├── cross_config/
│   │   └── vime/
│   │       └── dense/
│   ├── hardware/
│   ├── integrations/
│   │   ├── engines/
│   │   │   ├── rollout/
│   │   │   └── train/
│   │   └── orchestrators/
│   │       └── vime/
│   │           ├── dense/
│   │           └── rocm_attention/
│   ├── local/
│   ├── models/
│   ├── policies/
│   ├── suites/
│   │   ├── e2e/
│   │   ├── ws1/
│   │   └── ws2/
│   ├── tuning/
│   └── workloads/
├── csrc/
│   ├── ascend/
│   │   ├── attention/
│   │   ├── collectives/
│   │   └── gemm/
│   ├── bindings/
│   ├── common/
│   ├── cuda/
│   │   ├── activation/
│   │   ├── attention/
│   │   ├── collectives/
│   │   ├── embedding/
│   │   ├── gemm/
│   │   ├── logprob/
│   │   ├── norm/
│   │   ├── rope/
│   │   └── utils/
│   ├── musa/
│   │   ├── attention/
│   │   └── gemm/
│   └── rocm/
│       ├── attention/
│       ├── collectives/
│       └── logprob/
├── docker/
├── requirements/
│   ├── constraints/
│   └── integrations/
└── tools/
    ├── benchmarking/
    ├── checks/
    ├── env/
    ├── migration/
    ├── validation/
    │   ├── distributed/
    │   ├── models/
    │   └── operators/
    └── weights/

Documentation, examples, reports, and retained helpers (76 directories)

RL-Kernel/
├── .claude/
│   └── skills/
│       └── ws1-single-card-kernel/
│           └── ascendc_skills/
│               ├── ascendc-api-best-practices/
│               ├── ascendc-kernel-dev/
│               └── ascendc-perf-optimize/
├── docs/
│   ├── api/
│   ├── architecture/
│   ├── archive/
│   │   └── design/
│   ├── assets/
│   │   ├── blog/
│   │   │   └── rl-kernel-v0.1.0/
│   │   ├── readme/
│   │   └── rfc/
│   │       ├── gemma4-31b/
│   │       └── minimax-h3/
│   ├── benchmarking/
│   ├── blog/
│   ├── cli/
│   ├── community/
│   ├── contracts/
│   ├── contributing/
│   ├── getting_started/
│   ├── integrations/
│   │   └── orchestrators/
│   │       └── vime/
│   │           ├── dense/
│   │           ├── rocm_ablation/
│   │           └── rocm_attention/
│   ├── mkdocs/
│   │   ├── javascripts/
│   │   └── stylesheets/
│   ├── operators/
│   ├── usage/
│   │   └── evidence/
│   │       └── pr432-20260920/
│   └── validation/
├── examples/
│   ├── operators/
│   ├── post_training/
│   ├── vime_qwen3_8b_rocm_ablation/
│   ├── vime_qwen3_8b_tp4_cp2_200/
│   │   ├── companion_patches/
│   │   ├── profiles/
│   │   └── results/
│   │       ├── convergence_s1234_g10_g11/
│   │       ├── pr396_rocm_s1234_g10_g11_200/
│   │       └── scale_reference_s1234_g10_g11_optimized/
│   └── vime_rocm_attention_ablation/
│       ├── companion_patches/
│       └── profiles/
├── reports/
│   ├── archive/
│   │   ├── benchmarks/
│   │   │   ├── pr319_rocm_mi300x/
│   │   │   │   └── distributed/
│   │   │   ├── pr325_rocm_mi300x/
│   │   │   ├── pr328_rocm_mi300x/
│   │   │   ├── pr396_rocm_mi300x/
│   │   │   ├── qwen_ffn_h100_trace/
│   │   │   ├── qwen_ffn_layout_h100/
│   │   │   ├── ws2_cpu/
│   │   │   └── ws2_rocm_mi300x/
│   │   └── experiments/
│   │       └── dense/
│   │           ├── convergence_s1234_g10_g11/
│   │           │   └── figures/
│   │           ├── pr396_rocm_s1234_g10_g11_200/
│   │           └── scale_reference_s1234_g10_g11_optimized/
│   ├── experiments/
│   │   └── refactor/
│   └── releases/
└── scripts/

See the Contributor Guide for directory placement rules and the migration map for old-to-new paths.

Main changes

• Separate model implementations from reusable operators and framework adapters.
• Move deterministic FFN implementations under the CUDA and ROCm backends, with compatibility imports for previous paths.
• Keep the ROCm-only Triton FFN under the ROCm backend.
• Use a dedicated ROCm strict FFN autograd function instead of globally replacing shared backward behavior.
• Keep ROCm strict backward on deterministic GEMM paths, including TP4/CP2 reference coverage.
• Package the Vime JSON configuration as a regular file for symlink-disabled checkouts.
• Integrate the latest main branch, including MUSA fused logprob support.

Validation

No refactor-caused functional, numerical, or end-to-end performance regression was observed on CUDA or ROCm. GPU means below use rounds 2 and 3; round 1 is warmup.

CUDA H100

Configuration	Mismatch count	Max absolute logprob difference	Exact equality
P/P native	14,262 / 73,707	0.659206	false
R/R strict	0 / 71,369	0	true

Metric	P/P native	R/R strict	R/R relative to P/P
Rollout time	50.909174 s	43.871006 s	13.82% faster
Effective tokens/GPU/s	86.123703	93.558374	8.63% higher
Update weights	27.278190 s	27.247268 s	0.11% faster
Reference log probs	1.199057 s	1.319472 s	10.04% slower
Log probs	1.758154 s	1.785446 s	1.55% slower
Actor train	5.098225 s	5.986691 s	17.43% slower
Train time	8.221268 s	9.261354 s	12.65% slower
Actor train tok/s	7,335.350589	5,698.934777	22.31% lower
End-to-end step	92.304281 s	86.631667 s	6.15% faster

CUDA strict R/R had zero mismatches in the selected rounds. Main-versus-refactor strict trajectories were raw-byte identical across 2,136 tensors, 6,963,648 tensor elements, 2,968 scalar values, and four publications of 399 parameters.

ROCm MI300X

Configuration	Mismatch count	Max absolute logprob difference	Exact equality
P/P native	25,857 / 96,262	3.142752	false
R/R strict	0 / 93,851	0	true

Metric	P/P native	R/R strict	R/R relative to P/P
Rollout time	42.775298 s	41.860492 s	2.14% faster
Effective tokens/GPU/s	139.920280	139.194297	0.52% lower
Update weights	0.534724 s	0.557623 s	4.28% slower
Reference log probs	2.484850 s	3.729799 s	50.10% slower
Log probs	2.461369 s	2.366084 s	3.87% faster
Actor train	5.770843 s	9.017135 s	56.25% slower
Train time	11.162281 s	15.471597 s	38.61% slower
Actor train tok/s	8,560.598426	5,327.845068	37.76% lower
End-to-end step	55.449389 s	59.192035 s	6.75% slower

ROCm strict R/R had zero bitwise mismatches in all 129,729 checked elements across all three rounds. In rounds 2 and 3, all 93,851 checked elements and all 16 selected samples were bitwise exact. The strict mean end-to-end time improved from the previous 59.866968 s result to 59.192035 s.

Path and test coverage

• P/P used native Megatron and vLLM attention, FFN, and logprob implementations with no fallback.
• R/R used RL-Kernel strict attention, FFN, logprob, deterministic MFMA GEMM, and grouped RCCL implementations with no fallback.
• The final ROCm deterministic GEMM, strict backward, packed FFN, and collective suite passed 71 tests.
• Conflict-resolution validation passed 96 tests with 5 skips; formatting, compilation, and diff checks passed.

Performance numbers apply to the tested configurations and are regression evidence, not universal hardware claims.

[coderabbitai]

Walkthrough

Warning

Review details and warnings were omitted to fit the comment limit.