lightmoe 0.1.0__tar.gz

lightmoe-0.1.0/BENCHMARKS.md

@@ -0,0 +1,174 @@
+# Benchmarks
+
+Operator measurements use an NVIDIA B200 (`sm100`, 148 SMs),
+`nvcr.io/nvidia/pytorch:26.07-py3`, PyTorch
+`2.13.0a0+9186a08b2c.nv26.07`, CUDA 13.3, driver 580.95.05, CUTLASS DSL
+4.7.0, FlashInfer 0.6.17, and Transformer Engine 2.17. The full-model result
+uses eight B200s, vLLM 0.27.1, PyTorch 2.13.0+cu130, and FlashInfer
+0.6.16.post3.
+
+## Measurement
+
+| Scope | Included | Excluded |
+|---|---|---|
+| Grouped GEMM | one prepacked NVFP4 grouped GEMM launch | quantization, packing, allocation, compilation, autotuning |
+| Complete MoE inference | BF16 input quantization, dispatch, two expert projections, activation, routing weights, combine | router logits, top-k selection, communication, weight packing, compilation |
+| Routed-expert training | dispatch, expert forward and backward, routing weights, combine | router logits, top-k selection, communication, optimizer |
+| Full-model serving | one warmed `LLM.generate` call | model load, JIT compilation, dataset loading, tokenization |
+
+Operator backends run on identical shapes and route tensors in randomized,
+alternating order. Each point uses 21 CUDA-event samples and reports median
+`[IQR]`. Training arms run for two seconds before measurement. The complete
+arm sequence is then repeated; a point is rejected after three attempts if the
+two LightMoE medians differ by more than 5%. Runs are also rejected when host
+wall time divided by enclosed CUDA-event time exceeds 1.5. An unavailable
+backend is marked unavailable, not replaced by another implementation.
+
+For expert `e`, `M_e` is its local row count. **Ragged** means that the `M_e`
+values are unequal; a zero-row expert is valid. The stress matrix uses unequal
+counts, a route distribution concentrated on one expert, and counts around
+tile boundaries with one zero-row expert.
+
+Grouped-GEMM useful work is `2 × sum(M_e) × N × K`, with one multiply-add
+counted as two FLOPs. Peak percentages use the B200 dense FP4 specification of
+9,000 TFLOP/s; the 18,000 TFLOP/s structured-sparse figure is not used. This is
+logical throughput, so work added by tile padding is not counted as useful
+FLOPs.
+
+## Grouped NVFP4 GEMM
+
+These are model-derived expert projection shapes with 8,192 input tokens per
+EP rank and balanced routing. The local grouped GEMM contains
+`sum(M_e) = tokens × top-k` rows across `experts / EP size` local experts.
+LightMoE and FlashInfer write to preallocated outputs. PyTorch
+`scaled_grouped_mm` allocates its output because its API has no `out=` argument.
+The baseline column selects the faster runnable baseline for each row.
+
+| Model / projection | Experts / top-k / EP size | `N × K` | LightMoE µs `[IQR]` | Fastest baseline µs | Baseline / LightMoE | TFLOP/s | Dense FP4 peak |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| Qwen3.5-35B-A3B gate/up | 256 / 8 / 16 | 1,024 × 2,048 | **77.0 [0.2]** | FlashInfer 95.8 | 1.243× | 3,569 | 39.7% |
+| Qwen3.5-35B-A3B down | 256 / 8 / 16 | 2,048 × 512 | **67.8 [2.0]** | FlashInfer 68.0 | 1.003× | 2,027 | 22.5% |
+| Qwen3.5-397B-A17B gate/up | 512 / 10 / 32 | 2,048 × 4,096 | **296.2 [9.3]** | PyTorch 342.2 | 1.155× | 4,640 | 51.6% |
+| Qwen3.5-397B-A17B down | 512 / 10 / 32 | 4,096 × 1,024 | **214.3 [1.0]** | FlashInfer 248.4 | 1.159× | 3,207 | 35.6% |
+| DeepSeek-V4-Flash gate/up | 256 / 6 / 32 | 4,096 × 4,096 | **353.4 [12.1]** | PyTorch 392.4 | 1.110× | 4,667 | 51.9% |
+| DeepSeek-V4-Flash down | 256 / 6 / 32 | 4,096 × 2,048 | **195.9 [2.0]** | PyTorch 226.5 | 1.156× | 4,210 | 46.8% |
+| Kimi-K3 gate/up | 896 / 16 / 32 | 6,144 × 3,584 | **1,191.0 [27.6]** | FlashInfer 1,777.7 | 1.493× | 4,847 | 53.9% |
+| Kimi-K3 down | 896 / 16 / 32 | 3,584 × 3,072 | **644.4 [3.1]** | FlashInfer 893.9 | 1.387× | 4,479 | 49.8% |
+| GLM-5.2 gate/up | 256 / 8 / 16 | 4,096 × 6,144 | **627.7 [2.3]** | PyTorch 691.4 | 1.101× | 5,255 | **58.4%** |
+| GLM-5.2 down | 256 / 8 / 16 | 6,144 × 2,048 | **406.7 [3.9]** | PyTorch 437.4 | 1.075× | 4,055 | 45.1% |
+| MiniMax-M3 gate/up | 128 / 4 / 16 | 6,144 × 6,144 | **486.6 [2.2]** | PyTorch 530.5 | 1.090× | 5,084 | 56.5% |
+| MiniMax-M3 down | 128 / 4 / 16 | 6,144 × 3,072 | **275.6 [2.0]** | PyTorch 304.2 | 1.104× | 4,489 | 49.9% |
+| Nemotron-3.5-Lightning-30B-A3B up | 128 / 6 / 16 | 1,920 × 2,688 | **156.1 [3.4]** | PyTorch 158.3 | 1.014× | 3,250 | 36.1% |
+| Nemotron-3.5-Lightning-30B-A3B down | 128 / 6 / 16 | 2,688 × 1,920 | **167.2 [0.3]** | PyTorch 168.5 | 1.008× | 3,035 | 33.7% |
+
+All 14 balanced rows pass the two validity gates. The maximum LightMoE repeat
+deviation is 4.10%, the maximum host/CUDA ratio is 1.035, and sampled cosine
+against FP32 GEMM from the same BF16 operands is at least 0.99086. Nemotron's
+1,856-wide expert intermediate is padded to 1,920 for the NVFP4 kernel.
+Kimi-K3 uses its 3,584-wide routed latent expert input; its 7,168-wide outer
+projections are outside this grouped-GEMM boundary.
+
+The ragged matrix covers imbalanced, single-expert-skew, and alignment-stress
+routing, including a zero-row expert. PyTorch is unavailable for these rows
+because its NVFP4 grouped GEMM requires every group to have a 128-row-aligned
+size.
+
+## Complete MoE inference
+
+[![Complete MoE inference latency](https://raw.githubusercontent.com/roycho96/lightmoe/main/assets/moe-inference-latency.svg)](https://github.com/roycho96/lightmoe/blob/main/assets/moe-inference-latency.svg)
+
+[![Complete MoE inference speedup](https://raw.githubusercontent.com/roycho96/lightmoe/main/assets/moe-inference-speedup.svg)](https://github.com/roycho96/lightmoe/blob/main/assets/moe-inference-speedup.svg)
+
+`M` is the number of input tokens before top-k routing; the layer processes
+`M × top-k` token–expert assignments. Each point uses the model-derived local
+EP shard, fixed synthetic BF16 inputs, deterministic balanced routes, prepacked
+weights, preallocated outputs, and CUDA Graph replay. The FlashInfer 0.6.17 arm
+captures `fp4_quantize` and `trtllm_fp4_block_scale_routed_moe` together,
+matching the complete `InferenceMoE` call boundary. Package initialization and
+tactic loading happen before capture.
+
+| Model | `M=1` µs | `M=128` µs | `M=8,192` µs | FlashInfer / LightMoE at `M=8,192` | Lowest ratio in sweep |
+|---|---:|---:|---:|---:|---:|
+| Qwen3.5-35B-A3B | 14.4 | 28.8 | 252.5 | **1.290×** | 0.857× at `M=128` |
+| Qwen3.5-397B-A17B | 20.5 | 53.6 | 770.7 | **1.211×** | 0.907× at `M=32` |
+| DeepSeek-V4-Flash | 20.5 | 48.9 | 679.5 | **1.365×** | 0.828× at `M=32` |
+| GLM-5.2 | 40.5 | 91.4 | 1,364.0 | **1.369×** | 0.920× at `M=128` |
+| MiniMax-M3 | 34.7 | 71.8 | 900.5 | **1.524×** | 0.932× at `M=128` |
+| Nemotron-3.5-Lightning-30B-A3B | 16.4 | 32.7 | 387.3 | **1.175×** | 0.821× at `M=32` |
+
+All 42 points pass the measurement gates. The maximum LightMoE repeat
+deviation is 3.64% and the maximum host/CUDA ratio is 1.054. Ratios below 1.0
+identify short-token points where FlashInfer is faster; they are retained in
+the speedup chart.
+
+## Routed-expert training
+
+[![Complete MoE training latency](https://raw.githubusercontent.com/roycho96/lightmoe/main/assets/moe-training-latency.svg)](https://github.com/roycho96/lightmoe/blob/main/assets/moe-training-latency.svg)
+
+[![Complete MoE training speedup](https://raw.githubusercontent.com/roycho96/lightmoe/main/assets/moe-training-speedup.svg)](https://github.com/roycho96/lightmoe/blob/main/assets/moe-training-speedup.svg)
+
+The complete forward-and-backward curve uses the same model-derived local EP
+shards and deterministic balanced routes as the inference curve. The speedup
+chart includes Transformer Engine only for standard SwiGLU models with an
+exact activation contract. Bounded SwiGLU, SwiGLU-OAI, and ReLU² remain in the
+LightMoE latency chart and are not compared with a different TE activation.
+
+| Model | `M=32` ms `[IQR]` | `M=512` ms `[IQR]` | `M=8,192` ms `[IQR]` | TE / LightMoE at `M=8,192` |
+|---|---:|---:|---:|---:|
+| Qwen3.5-35B-A3B | 2.140 [0.037] | 2.163 [0.034] | 2.174 [0.036] | **1.914×** |
+| Qwen3.5-397B-A17B | 2.188 [0.032] | 2.207 [0.042] | 4.066 [0.023] | **1.505×** |
+| DeepSeek-V4-Flash | 2.085 [0.140] | 2.090 [0.118] | 3.791 [0.115] | — |
+| GLM-5.2 | 2.207 [0.046] | 2.215 [0.047] | 6.465 [0.117] | **1.369×** |
+| MiniMax-M3 | 2.069 [0.116] | 2.120 [0.150] | 4.646 [0.046] | — |
+| Nemotron-3.5-Lightning-30B-A3B | 2.058 [0.111] | 2.066 [0.093] | 2.480 [0.028] | — |
+
+All 30 LightMoE points and all 15 exact-contract TE points pass the measurement
+gates. The maximum LightMoE repeat deviation is 1.98% and the maximum
+host/CUDA ratio is 1.014. Input, router, expert-weight, and output gradients
+are finite. Minimum sampled cosine against the BF16 reference is 0.9726 for
+outputs and 0.9617 for input gradients. These are precision checks, not
+training-convergence results.
+
+## Full-model vLLM serving
+
+`nvidia/DeepSeek-V4-Flash-NVFP4` revision
+`e3cd60e7de98e9867116860d522499a728de1cf9` runs on eight B200s with tensor
+parallel size 8 and an FP8 KV cache. Attention, scheduling, collectives, and all
+non-MoE operators are fixed; only the NVFP4 MoE backend changes. Prompts are
+fixed SWE-bench Verified samples from revision
+`78f471bf655a3137b2e8a75af1501690ec009ec3`.
+
+| Workload | Fixed batch | LightMoE initial median `[IQR]` | FlashInfer TensorRT-LLM median `[IQR]` | LightMoE repeat median `[IQR]` | Baseline / LightMoE |
+|---|---:|---:|---:|---:|---:|
+| Prefill | 8 × 1,024 input → 1 output | 221.670 [2.869] ms | 250.464 [4.300] ms | 226.927 [1.803] ms | 1.10–1.13× |
+| Decode | 32 × 256 input → 256 output | 2.4488 [0.0125] s | 2.5395 [0.0667] s | 2.4587 [0.0653] s | 1.033–1.037× |
+
+Each backend has 21 warmed runs. Initial-to-repeat drift is 2.37% for prefill
+and 0.40% for decode. Inputs, outputs, and generated token sequences match
+across backends. This is the only full-model backend swap claimed here; an
+operator result is never presented as end-to-end performance.
+
+## Run
+
+Pinned model sources and revisions live in `benchmarks/model_shapes.py`.
+
+```bash
+python -m benchmarks.nvfp4_gemm \
+  --models all --tokens 8192 --routing balanced \
+  --projections gate_up,down \
+  --backends lightmoe,flashinfer_cutedsl,torch_scaled_grouped_mm \
+  --mode prepacked --warmup 5 --iterations 21
+
+python -m benchmarks.nvfp4_gemm \
+  --models all --tokens 8192 \
+  --routing imbalanced,single_expert_skew,alignment_stress \
+  --projections gate_up,down \
+  --backends lightmoe,flashinfer_cutedsl,torch_scaled_grouped_mm \
+  --mode prepacked --warmup 5 --iterations 21
+
+python -m benchmarks.moe \
+  --models all --tokens 8192 --routing balanced,imbalanced \
+  --backends lightmoe,transformer_engine_nvfp4_fused,pytorch_bf16 \
+  --scope full-layer --pass fwd_bwd --warmup 15 --iterations 21 \
+  --interleave-training
+```

lightmoe-0.1.0/LICENSE

@@ -0,0 +1,201 @@
+                                 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.

lightmoe-0.1.0/MANIFEST.in

@@ -0,0 +1,6 @@
+include LICENSE
+include README.md
+include BENCHMARKS.md
+include NOTICE
+recursive-include benchmarks *.py
+prune tests

lightmoe-0.1.0/NOTICE

@@ -0,0 +1,24 @@
+LightMoE third-party notices
+============================
+
+The following files retain upstream copyright and license notices.
+
+BSD-3-Clause
+------------
+
+- lightmoe/kernels/dense/kernel.py
+- lightmoe/kernels/grouped/kernel.py
+- lightmoe/kernels/grouped/pipeline.py
+
+These files are based on NVIDIA CUTLASS CuTe DSL examples.
+Upstream: https://github.com/NVIDIA/cutlass
+
+Apache-2.0
+----------
+
+- lightmoe/kernels/grouped/scheduler.py
+- lightmoe/kernels/grouped/wgrad_kernel.py
+- lightmoe/kernels/grouped/wgrad_tma.py
+
+These files are based on NVIDIA cuDNN Frontend examples.
+Upstream: https://github.com/NVIDIA/cudnn-frontend