axolotl-ringmaster 0.1.0__tar.gz

axolotl_ringmaster-0.1.0/.github/workflows/release.yml

@@ -0,0 +1,52 @@
+name: release
+
+on:
+  push:
+    tags:
+      - "v*"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+        with:
+          persist-credentials: false
+          fetch-depth: 0 # setuptools-scm derives the version from the tag
+      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
+        with:
+          python-version: "3.12"
+      - name: Build sdist and wheel
+        run: |
+          python -m pip install --upgrade build twine
+          python -m build
+          twine check dist/*
+      - name: Smoke-test the wheel
+        run: |
+          python -m venv /tmp/smoke
+          /tmp/smoke/bin/pip install torch --index-url https://download.pytorch.org/whl/cpu
+          /tmp/smoke/bin/pip install dist/*.whl
+          /tmp/smoke/bin/python -c "import ringmaster; print(ringmaster.__version__)"
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist
+          path: dist/
+
+  publish:
+    needs: build
+    runs-on: ubuntu-latest
+    environment:
+      name: pypi
+      url: https://pypi.org/p/axolotl-ringmaster
+    permissions:
+      id-token: write # PyPI trusted publishing (OIDC)
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: dist
+          path: dist/
+      - uses: pypa/gh-action-pypi-publish@release/v1

axolotl_ringmaster-0.1.0/.gitignore

@@ -0,0 +1,40 @@
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+build/
+dist/
+*.egg-info/
+*.egg
+.eggs/
+
+# Virtual envs
+.venv/
+venv/
+env/
+
+# Test / lint / type caches
+.pytest_cache/
+.cache/
+.coverage
+.coverage.*
+htmlcov/
+.mypy_cache/
+.ruff_cache/
+.dmypy.json
+
+# Training / example artifacts
+outputs/
+last_run_prepared/
+wandb/
+*.log
+checkpoint-*/
+runs/
+
+# Editor / OS
+.idea/
+.vscode/
+.DS_Store
+*.swp

axolotl_ringmaster-0.1.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

axolotl_ringmaster-0.1.0/PKG-INFO

@@ -0,0 +1,171 @@
+Metadata-Version: 2.4
+Name: axolotl-ringmaster
+Version: 0.1.0
+Summary: Composable long-context attention: Ulysses / Ring / USP sequence parallelism for any framework
+Author: Axolotl AI
+License: Apache-2.0
+Project-URL: Homepage, https://github.com/axolotl-ai-cloud/ringmaster
+Project-URL: Repository, https://github.com/axolotl-ai-cloud/ringmaster
+Classifier: Development Status :: 3 - Alpha
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: torch>=2.11
+Provides-Extra: transformers
+Requires-Dist: transformers>=5.9; extra == "transformers"
+Provides-Extra: dev
+Requires-Dist: pytest; extra == "dev"
+Requires-Dist: transformers>=5.9; extra == "dev"
+Dynamic: license-file
+
+# ringmaster
+
+Composable long-context attention for distributed LLM training. Ulysses / Ring /
+USP sequence parallelism that **wraps existing kernels** instead of shipping its
+own — so it rides HuggingFace `kernels` Flash-Attention (FA2/FA3/FA4), torch SDPA,
+and FlexAttention, with **no `flash_attn` pypi dependency**. Framework-neutral
+core; thin adapters for axolotl, trl/accelerate, and raw PyTorch.
+
+## Why
+
+`ring-flash-attn` (the previous CP path) depends on the `flash_attn` pypi package,
+which is no longer installed now that attention comes from HF `kernels`. ringmaster
+takes two routes that avoid that dependency entirely:
+
+- **Ulysses** wraps the model's *own* attention kernel with all-to-all on the head
+  dim — works with any kernel (FA2/3/4/sdpa/flex), inherently load-balanced, the
+  cheapest comm. Capped by KV-head count.
+- **Ring** rides torch-native context parallel (`torch.distributed.tensor.experimental`)
+  on the SDPA/FlexAttention backends — no head cap, scales linearly. LSE merging
+  is handled inside aten, so we never reimplement a ring kernel.
+- **USP** composes the two on a 2D mesh (Ulysses intra-node, Ring inter-node).
+
+An **auto-selector** picks the `ulysses_size × ring_size` split from KV-head count,
+world size, and intra-node topology.
+
+## Install
+
+```bash
+pip install axolotl-ringmaster
+```
+
+The distribution is `axolotl-ringmaster` (the unrelated `ringmaster` name was taken
+on PyPI); the import name is `ringmaster`.
+
+## Quickstart
+
+Requires **torch ≥ 2.11**. The framework adapters do this for you; the raw shape is:
+
+```python
+import ringmaster as rm
+
+cfg = rm.RingmasterConfig(size=8, backend=rm.Backend.AUTO)
+runtime = rm.setup(cfg, num_kv_heads=8, device_mesh=mesh)   # mesh has a "cp" dim
+model.set_attn_implementation(runtime.attn_implementation)
+rm.wire_recurrent_layers(model)                             # no-op for plain attention
+
+from ringmaster.adapters.raw import context_parallel_region
+for batch in loader:
+    with context_parallel_region(batch) as shard:           # contiguous seq shard
+        loss = model(**shard).loss
+    loss.backward()
+```
+
+Pair CP with **FSDP2** (a `dp_shard × cp` mesh), not plain DDP — FSDP2's gradient
+reduce-scatter sums the CP-partial gradients; DDP would mis-reduce. See
+[docs/concepts.md](docs/concepts.md#device-mesh-and-fsdp2-composition).
+
+## Documentation
+
+- [docs/concepts.md](docs/concepts.md) — backends (Ulysses/Ring/USP), rotate
+  methods, load balancing, device-mesh / FSDP2 composition, and when to use which.
+- [docs/integration.md](docs/integration.md) — raw PyTorch, trl/accelerate
+  (`CPSFTTrainer` + `ParallelismConfig` + `chunked_nll`), and axolotl (plugin + YAML).
+- [docs/api.md](docs/api.md) — concise reference of the public API.
+- [docs/models.md](docs/models.md) — plain attention, gated linear-attention
+  (Qwen3.5/Next), Mamba2 hybrids (Nemotron-H), and the sm_120 shim.
+- [examples/cp_fsdp2/](examples/cp_fsdp2/) — worked TRL + FSDP2 script and axolotl
+  configs for all three model families, plus a CP-vs-non-CP benchmark table.
+
+## Status
+
+Requires **torch ≥ 2.11**.
+
+| Phase | Feature | State |
+|-------|---------|-------|
+| 0 | Package scaffold, config, mesh, runtime, registry, loss | ✅ |
+| 1 | Ulysses backend (wraps HF kernels FA2/3/4/sdpa) | ✅ GPU-validated — sdpa + flash parity, fwd+bwd |
+| 2 | Ring — kernel-agnostic loop; **p2p (memory-optimal, fwd+bwd) + allgather (fast)** rotaters; `hf_kernels`/`torch_native`/`math` providers | ✅ GPU parity (both providers) + gloo fwd+bwd grad parity (p2p **and** allgather) |
+| 3 | USP 2D composition (Ulysses × Ring) + auto-selector | ✅ **2×2 validated** (4 gloo ranks) + auto-selector |
+| 4 | Mamba2 SSM CP (owned: `strategies/mamba.py`) + linear-attention state-passing | ✅ **mamba2 CP validated with the real mamba-ssm kernel (2 GPU)**; **gated-DeltaNet linear attention validated with the real flash-linear-attention kernel (2 GPU)**; cp_linear_scan + mamba2-vs-SSD (gloo); sliding-window via flash `window` |
+| 5 | ALST memory | deferred to host framework (axolotl-native) |
+| 6 | Profiling, trl/accelerate + multimodal adapters, **accelerate ParallelismConfig compat**, benchmark, docs, axolotl plugin (schema-merge tested) | ✅ |
+| 7 | Custom collectives — fused Q/K/V all-to-all (MHA) ✅; **comm/compute overlap (prefetch) in p2p ring** ✅; symmetric-memory / Triton | partial (two items done) |
+| v2 | packing / varlen | deferred |
+
+## Benchmark (2× GPU, Ulysses, Llama hidden=2048/heads=16/layers=6, fwd+bwd)
+
+| seq | non-CP (1 GPU) peak | CP (2 GPU) per-GPU peak | CP step |
+|----:|----:|----:|----:|
+| 16k | 8.4 GB | 4.6 GB | 440 ms |
+| 65k | 31 GB | 16 GB | 2.1 s |
+| 131k | 62 GB (10.1 s) | 31 GB | 6.3 s |
+| 262k | **OOM** | 62 GB | 21 s |
+
+CP halves per-GPU memory at every length (1/world sequence sharding) and reaches
+2× the sequence (262k vs 131k) on the same hardware. `benchmarks/cp_vs_noncp.py`.
+
+**CP is a memory/latency tool, not a throughput win.** It uses `world` GPUs, so
+total GPU-time is higher than non-CP for a sequence that already fits on one GPU:
+at 131k, non-CP = 10.1 GPU-s vs CP = 6.3 s × 2 = 12.6 GPU-s (~1.25×); at 16k the
+overhead is ~2.4× (comm dominates). The overhead shrinks as sequence grows
+(attention compute is superlinear, comm linear) but never breaks even. Use CP for
+context that won't fit on one GPU, or to cut latency when you have spare GPUs; for
+sequences that fit, DP is more GPU-efficient.
+
+**p2p ring is the memory-optimal path** (only O(S/P) KV resident) with a proper
+flash-style reverse-ring backward and DistFlashAttn-style comm/compute overlap
+(prefetch the next KV block while computing the current). allgather is the
+faster/higher-memory option (torchtitan-style). Both train (grad-parity validated).
+
+**Remaining gap:** a live multi-GPU `axolotl train` run for correct *gradient*
+semantics needs CP to be a non-DP mesh dim (via accelerate ParallelismConfig
+`cp_size`, now supported by `ringmaster.parallelism.setup_from_parallelism_config`);
+wiring axolotl's core to route through it (vs the legacy ring path) is the last
+integration step.
+
+Comm/memory are profilable: wrap any step in `rm.profile_comms()` / `rm.profile_memory()`
+to get per-collective bytes+time and peak memory. The Ulysses leg fuses K/V into one
+all-to-all (3 collectives/attn, not 4); the Ring leg offers `allgather` (fast, more
+memory) vs `p2p` (O(S/P) KV resident) rotaters.
+
+## Layout
+
+```
+ringmaster/
+  config.py        RingmasterConfig (framework-neutral)
+  compat.py        torch>=2.11 gate + experimental API isolation
+  runtime.py       process-wide CP groups + active runtime
+  mesh.py          resolve cp group(s), topology probe
+  comm.py          SeqAllToAll4D (Ulysses all-to-all autograd; gloo fallback)
+  profiling.py     profile_comms / profile_memory
+  ring/
+    loop.py        ring loop: allgather (autograd) + p2p (memory-frugal)
+    kernels.py     block kernels: hf_kernels flash / aten flash (pluggable)
+    merge.py       online-softmax LSE merge
+  strategies/
+    ulysses.py     AttentionInterface wrapper around HF kernels
+    ring.py        register kernel-agnostic ring + impl resolution
+    usp.py         2D composition + auto_select
+    state_passing.py  Mamba2/linear exact CP linear scan
+  shard.py         dense contiguous sequence sharding
+  balancers.py     head_tail/contiguous index gen (Ring)
+  loss.py          CP-correct token-weighted loss reduction
+  multimodal.py    shard fused sequence after the encoder
+  memory/          deferred to host framework (axolotl ALST)
+  adapters/        raw PyTorch / trl+accelerate / hatchery-core entry points
+  benchmarks/      profile_cp.py (comm + memory across seq lengths)
+```

axolotl_ringmaster-0.1.0/README.md

@@ -0,0 +1,148 @@
+# ringmaster
+
+Composable long-context attention for distributed LLM training. Ulysses / Ring /
+USP sequence parallelism that **wraps existing kernels** instead of shipping its
+own — so it rides HuggingFace `kernels` Flash-Attention (FA2/FA3/FA4), torch SDPA,
+and FlexAttention, with **no `flash_attn` pypi dependency**. Framework-neutral
+core; thin adapters for axolotl, trl/accelerate, and raw PyTorch.
+
+## Why
+
+`ring-flash-attn` (the previous CP path) depends on the `flash_attn` pypi package,
+which is no longer installed now that attention comes from HF `kernels`. ringmaster
+takes two routes that avoid that dependency entirely:
+
+- **Ulysses** wraps the model's *own* attention kernel with all-to-all on the head
+  dim — works with any kernel (FA2/3/4/sdpa/flex), inherently load-balanced, the
+  cheapest comm. Capped by KV-head count.
+- **Ring** rides torch-native context parallel (`torch.distributed.tensor.experimental`)
+  on the SDPA/FlexAttention backends — no head cap, scales linearly. LSE merging
+  is handled inside aten, so we never reimplement a ring kernel.
+- **USP** composes the two on a 2D mesh (Ulysses intra-node, Ring inter-node).
+
+An **auto-selector** picks the `ulysses_size × ring_size` split from KV-head count,
+world size, and intra-node topology.
+
+## Install
+
+```bash
+pip install axolotl-ringmaster
+```
+
+The distribution is `axolotl-ringmaster` (the unrelated `ringmaster` name was taken
+on PyPI); the import name is `ringmaster`.
+
+## Quickstart
+
+Requires **torch ≥ 2.11**. The framework adapters do this for you; the raw shape is:
+
+```python
+import ringmaster as rm
+
+cfg = rm.RingmasterConfig(size=8, backend=rm.Backend.AUTO)
+runtime = rm.setup(cfg, num_kv_heads=8, device_mesh=mesh)   # mesh has a "cp" dim
+model.set_attn_implementation(runtime.attn_implementation)
+rm.wire_recurrent_layers(model)                             # no-op for plain attention
+
+from ringmaster.adapters.raw import context_parallel_region
+for batch in loader:
+    with context_parallel_region(batch) as shard:           # contiguous seq shard
+        loss = model(**shard).loss
+    loss.backward()
+```
+
+Pair CP with **FSDP2** (a `dp_shard × cp` mesh), not plain DDP — FSDP2's gradient
+reduce-scatter sums the CP-partial gradients; DDP would mis-reduce. See
+[docs/concepts.md](docs/concepts.md#device-mesh-and-fsdp2-composition).
+
+## Documentation
+
+- [docs/concepts.md](docs/concepts.md) — backends (Ulysses/Ring/USP), rotate
+  methods, load balancing, device-mesh / FSDP2 composition, and when to use which.
+- [docs/integration.md](docs/integration.md) — raw PyTorch, trl/accelerate
+  (`CPSFTTrainer` + `ParallelismConfig` + `chunked_nll`), and axolotl (plugin + YAML).
+- [docs/api.md](docs/api.md) — concise reference of the public API.
+- [docs/models.md](docs/models.md) — plain attention, gated linear-attention
+  (Qwen3.5/Next), Mamba2 hybrids (Nemotron-H), and the sm_120 shim.
+- [examples/cp_fsdp2/](examples/cp_fsdp2/) — worked TRL + FSDP2 script and axolotl
+  configs for all three model families, plus a CP-vs-non-CP benchmark table.
+
+## Status
+
+Requires **torch ≥ 2.11**.
+
+| Phase | Feature | State |
+|-------|---------|-------|
+| 0 | Package scaffold, config, mesh, runtime, registry, loss | ✅ |
+| 1 | Ulysses backend (wraps HF kernels FA2/3/4/sdpa) | ✅ GPU-validated — sdpa + flash parity, fwd+bwd |
+| 2 | Ring — kernel-agnostic loop; **p2p (memory-optimal, fwd+bwd) + allgather (fast)** rotaters; `hf_kernels`/`torch_native`/`math` providers | ✅ GPU parity (both providers) + gloo fwd+bwd grad parity (p2p **and** allgather) |
+| 3 | USP 2D composition (Ulysses × Ring) + auto-selector | ✅ **2×2 validated** (4 gloo ranks) + auto-selector |
+| 4 | Mamba2 SSM CP (owned: `strategies/mamba.py`) + linear-attention state-passing | ✅ **mamba2 CP validated with the real mamba-ssm kernel (2 GPU)**; **gated-DeltaNet linear attention validated with the real flash-linear-attention kernel (2 GPU)**; cp_linear_scan + mamba2-vs-SSD (gloo); sliding-window via flash `window` |
+| 5 | ALST memory | deferred to host framework (axolotl-native) |
+| 6 | Profiling, trl/accelerate + multimodal adapters, **accelerate ParallelismConfig compat**, benchmark, docs, axolotl plugin (schema-merge tested) | ✅ |
+| 7 | Custom collectives — fused Q/K/V all-to-all (MHA) ✅; **comm/compute overlap (prefetch) in p2p ring** ✅; symmetric-memory / Triton | partial (two items done) |
+| v2 | packing / varlen | deferred |
+
+## Benchmark (2× GPU, Ulysses, Llama hidden=2048/heads=16/layers=6, fwd+bwd)
+
+| seq | non-CP (1 GPU) peak | CP (2 GPU) per-GPU peak | CP step |
+|----:|----:|----:|----:|
+| 16k | 8.4 GB | 4.6 GB | 440 ms |
+| 65k | 31 GB | 16 GB | 2.1 s |
+| 131k | 62 GB (10.1 s) | 31 GB | 6.3 s |
+| 262k | **OOM** | 62 GB | 21 s |
+
+CP halves per-GPU memory at every length (1/world sequence sharding) and reaches
+2× the sequence (262k vs 131k) on the same hardware. `benchmarks/cp_vs_noncp.py`.
+
+**CP is a memory/latency tool, not a throughput win.** It uses `world` GPUs, so
+total GPU-time is higher than non-CP for a sequence that already fits on one GPU:
+at 131k, non-CP = 10.1 GPU-s vs CP = 6.3 s × 2 = 12.6 GPU-s (~1.25×); at 16k the
+overhead is ~2.4× (comm dominates). The overhead shrinks as sequence grows
+(attention compute is superlinear, comm linear) but never breaks even. Use CP for
+context that won't fit on one GPU, or to cut latency when you have spare GPUs; for
+sequences that fit, DP is more GPU-efficient.
+
+**p2p ring is the memory-optimal path** (only O(S/P) KV resident) with a proper
+flash-style reverse-ring backward and DistFlashAttn-style comm/compute overlap
+(prefetch the next KV block while computing the current). allgather is the
+faster/higher-memory option (torchtitan-style). Both train (grad-parity validated).
+
+**Remaining gap:** a live multi-GPU `axolotl train` run for correct *gradient*
+semantics needs CP to be a non-DP mesh dim (via accelerate ParallelismConfig
+`cp_size`, now supported by `ringmaster.parallelism.setup_from_parallelism_config`);
+wiring axolotl's core to route through it (vs the legacy ring path) is the last
+integration step.
+
+Comm/memory are profilable: wrap any step in `rm.profile_comms()` / `rm.profile_memory()`
+to get per-collective bytes+time and peak memory. The Ulysses leg fuses K/V into one
+all-to-all (3 collectives/attn, not 4); the Ring leg offers `allgather` (fast, more
+memory) vs `p2p` (O(S/P) KV resident) rotaters.
+
+## Layout
+
+```
+ringmaster/
+  config.py        RingmasterConfig (framework-neutral)
+  compat.py        torch>=2.11 gate + experimental API isolation
+  runtime.py       process-wide CP groups + active runtime
+  mesh.py          resolve cp group(s), topology probe
+  comm.py          SeqAllToAll4D (Ulysses all-to-all autograd; gloo fallback)
+  profiling.py     profile_comms / profile_memory
+  ring/
+    loop.py        ring loop: allgather (autograd) + p2p (memory-frugal)
+    kernels.py     block kernels: hf_kernels flash / aten flash (pluggable)
+    merge.py       online-softmax LSE merge
+  strategies/
+    ulysses.py     AttentionInterface wrapper around HF kernels
+    ring.py        register kernel-agnostic ring + impl resolution
+    usp.py         2D composition + auto_select
+    state_passing.py  Mamba2/linear exact CP linear scan
+  shard.py         dense contiguous sequence sharding
+  balancers.py     head_tail/contiguous index gen (Ring)
+  loss.py          CP-correct token-weighted loss reduction
+  multimodal.py    shard fused sequence after the encoder
+  memory/          deferred to host framework (axolotl ALST)
+  adapters/        raw PyTorch / trl+accelerate / hatchery-core entry points
+  benchmarks/      profile_cp.py (comm + memory across seq lengths)
+```