Skip to content
CAI
Software that uses CAICheck a score

verl-project/verl

40.8

Weak · 16 November 2026

96.2k

lines of production code

Python

primary language

5

measurements over time

CAI band scale
CAI trend line
CAI lens gauges

What this system is

Features

Add GSPO training script for Qwen3-32B on Ascend NPU

A new example script (run\_qwen3\_32b\_fsdp.sh) is provided to train the Qwen3-32B model using the GSPO algorithm on Ascend NPUs. The script configures FSDP2 for the actor and reference models, utilizes vLLM for rollouts, and sets up specific hyperparameters for the DAPO-Math-17k dataset, enabling users to reproduce this specific optimization workflow.

_examples/ascend\_extras/gspo\trainer · high confidence

Add ReMax trainer examples with uv integration

Added new example scripts in \examples/remax\_trainer\ to demonstrate the ReMax policy-gradient method, including a synchronous FSDP training setup (\run\_qwen2.5\_math\_7b\_sync\_fsdp.sh\) and a standard FSDP setup (\run\_qwen3\_8b\_fsdp.sh\). These examples configure the trainer with \algorithm.adv\_estimator=remax\ and use vLLM for rollouts. The launch scripts now support \uv\ for dependency management on GPU environments, allowing users to run the examples via \uv run\ with the \vllm\ and \fsdp\ extras.

_examples/remax\trainer · high confidence

Add TPU Supervised Fine-Tuning (SFT) example using TorchTitan

Adds a new example for running Supervised Fine-Tuning on Google Cloud TPU v6e clusters using the TorchTitan engine. The entry includes a shell script (\run\_qwen3\_0\_6b\_torchtitan.sh\) and documentation (\README.md\) that support both \torchrun\ (SPMD multi-controller) and Ray (single-controller) launch modes, with defaults configured for pure FSDP parallelism and sequence bucketing to optimize XLA compilation.

examples/tpu/sft · high confidence

Add router replay example for MoE models

Added the \examples/router\_replay\ directory containing documentation and shell scripts for running Mixture of Experts (MoE) training with deterministic router replay. The new \README.md\ explains the R2 (record during log-prob, replay during update) and R3 (record during rollout, replay during update) modes, including configuration for Megatron and VeOmni engines and vLLM version requirements. The \run\_qwen3\_30b\_a3b\_megatron.sh\ script demonstrates R2/R3 replay with the Megatron engine on Qwen3-30B-A3B, while \run\_qwen3\_5\_35b\_a3b\_veomni.sh\ demonstrates the same for the VeOmni engine on Qwen3.5-35B-A3B, showing how to enable \enable\_rollout\_routing\_replay\ for R3 mode.

_examples/router\replay · high confidence

New Ascend Docker images for SGLang, vLLM, and Qwen3.5 across CANN 8.2–9.0

Added a comprehensive set of new Dockerfiles in docker/ascend to build Ascend NPU images for SGLang, vLLM, and Qwen3.5 on CANN 8.2, 8.3, 8.5, and 9.0, covering both A2 and A3 hardware variants. These images bundle the required PyTorch/torch\_npu stacks, vLLM-ascend, MindSpeed, and Megatron-LM, and include specific build steps such as compiling sgl-kernel-npu for A2/A3, installing mbridge, and patching mbridge for DeepSeek models where applicable.

docker/ascend · high confidence

New DPPO and GSPO RL training examples for Qwen3 models

Added new example scripts and documentation for Divergence Proximal Policy Optimization (DPPO) and Group Sequence Policy Optimization (GSPO) in the \examples/dppo\_trainer\ and \examples/gspo\_trainer\ directories. These examples demonstrate how to train Qwen3 models (including the 30B-A3B MoE variant) using Megatron and FSDP backends with vLLM rollouts, introducing specific loss modes (\dppo\_tv\, \dppo\_kl\, \gspo\) and configuration flags for users experimenting with advanced trust-region and sequence-level policy optimization techniques.

_examples/dppo\_trainer, examples/gspo\trainer · high confidence

New Docker image hierarchy and uv-based multi-backend image

The project introduces a new Docker image release hierarchy. Stable images for vLLM and SGLang are now built on top of the official vLLM and SGLang base images, respectively, with additional dependencies like Megatron-LM, Apex, and TransformerEngine baked in. A new universal \uv\-managed image (\Dockerfile.uv.cu130\) is added, which bakes a full package cache for all backend combinations (vLLM, SGLang, FSDP, Megatron, VeOmni, CPU) into a single image, allowing users to select their desired backend at runtime using \uv run --extra\. Additionally, new AWS-specific Dockerfiles are provided for SageMaker (vLLM) and EFA support.

docker · high confidence

New Docker images for SGLang and vLLM integration

Added new Dockerfiles to support specific inference engines: \Dockerfile.app.sglang\ installs SGLang 0.5.2 and torch-memory-saver, while \Dockerfile.vllm011.mcore\_gpt-oss\ sets up vLLM 0.11.0 with GPT-OSS dependencies. These images are based on the updated base environment defined in \Dockerfile.base\, which upgrades PyTorch to 2.8.0, Flash Attention to 2.7.4, and includes DeepEP and Megatron-LM components.

docker/verl0.6-cu128-torch2.8.0-fa2.7.4 · high confidence

New Docker images for verl v0.4 and v0.5 with updated base stacks and DeepEP support

This change introduces new Docker image definitions for two major release tracks. The v0.4 track (CUDA 12.4, PyTorch 2.6, FlashAttention 2.7.4) adds application images for SGLang and vLLM paired with Megatron-LM core v0.12 and v0.13 (preview), including variants with DeepEP integration. The v0.5 track (CUDA 12.6, PyTorch 2.7.1, FlashAttention 2.8.0) provides a base image with DeepEP pre-built and application images for SGLang 0.4.8 with Megatron-LM core v0.12 and v0.13 (preview). These images standardize the underlying framework versions and provide explicit support for DeepEP in the v0.4 app images and the v0.5 base image.

docker/verl0.4-cu124-torch2.6-fa2.7.4, docker/verl0.5-cu126-torch2.7.1-fa2.8.0 · high confidence

New Docker images for verl v0.5 with CUDA 12.8 and PyTorch 2.7.1

This location introduces the base and application Dockerfiles for the verl v0.5 preview release, upgrading the underlying stack to CUDA 12.8, cuDNN 9.8, and PyTorch 2.7.1. The base image includes Flash Attention 2.8.0, FlashInfer 0.2.6, and TransformerEngine 2.5, while the application image adds support for sglang 0.4.8 and Megatron-LM core\_r0.13.0.

docker/verl0.5-preview-cu128-torch2.7.1-fa2.8.0 · high confidence

New Docker images for verl v0.5 with PyTorch 2.7 and updated inference engines

This change introduces a new Docker image family (verl0.5) based on PyTorch 2.7.1 and CUDA 12.6. It provides a base image with DeepEP support and several application images: SGLang (0.4.9.post6 and 0.4.10.post2) and vLLM (0.10.0) variants, both integrated with Megatron-LM (0.13.0) and TransformerEngine (2.2.1). A separate vLLM variant is also provided with Megatron-LM 0.15.0 and TransformerEngine 2.7. These images update the underlying framework versions and add specific fixes for Qwen-VL and other dependencies.

docker/verl0.5-cu126-torch2.7-fa2.7.4 · high confidence

New GMPO trainer example and AIME reward scoring fix

The examples/gmpo\_trainer directory now includes a README and a run script for Geometric-Mean Policy Optimization (GMPO), a new training mode that uses the geometric mean of token-level rewards to improve stability over standard GRPO, particularly for MoE models. Additionally, a custom reward function for AIME validation data has been added to ensure correct scoring by normalizing data source tags.

_examples/gmpo\trainer · high confidence

New GRPO trainer examples for DeepSeek V4, GPT-OSS, GLM-4.1V, and Hy3

The \examples/grpo\_trainer\ directory now includes new launch scripts for Group Relative Policy Optimization (GRPO) training on several additional model families. Users can now run GRPO on DeepSeek-V3 (671B) and DeepSeek-V4-Flash using the Megatron backend, as well as DeepSeek-V4-Flash with the VeOmni engine. Support has also been added for GPT-OSS-20B (using FSDP and VeOmni backends), GLM-4.1V-9B (FSDP), and Hy3 (MoE) via Megatron. These scripts provide pre-configured parameters for data, model parallelism, and rollout settings, enabling users to fine-tune these specific models using the GRPO algorithm out of the box.

_examples/grpo\trainer · high confidence

New GRPO training launch scripts for GLM-5.2, Qwen3, and Qwen3.5 on Ascend NPUs

Added new shell scripts in the GRPO trainer examples directory to enable reinforcement learning with Group Relative Policy Optimization (GRPO) on Ascend NPUs for several large language models. The new files include \run\_glm5\_2\_megatron.sh\ for GLM-5.2 using Megatron parallelism, \run\_qwen3\_235b\_256k\_megatron.sh\ and \run\_qwen3\_30b\_a3b\_megatron.sh\ for Qwen3 variants (235B and 30B-A3B) with Megatron configurations, \run\_qwen3\_32b\_fsdp.sh\ for Qwen3-32B using FSDP, \run\_qwen3\_5\_122b\_a10b\_32k\_megatron.sh\ for Qwen3.5-122B-A10B with long-context support, and \run\_qwen3\_next\_80b\_fsdp.sh\ for Qwen3-Next-80B using FSDP. These scripts provide pre-configured parameters for data, model, algorithm, and hardware-specific settings (such as \use\_mbridge\, \use\_remove\_padding\, and specific Megatron/FSDP overrides) to facilitate training on Ascend hardware.

_examples/ascend\_extras/grpo\trainer · high confidence

New MTP training examples for MiMo-7B-RL with async and hybrid rollout support

Added a new \examples/mtp\_trainer\ directory providing ready-to-run scripts for Multi-Token-Prediction (MTP) training using the MiMo-7B-RL model with the Megatron backend. The entry includes a README and three shell scripts: a standard sync hybrid-engine setup (\run\_mimo\_7b\_mtp\_megatron.sh\), a hybrid setup supporting both SGLang and vLLM rollouts with speculative decoding options like EAGLE (\run\_mimo\_7b\_mtp\_rl\_vllm\_sgl\_megatron.sh\), and a fully-async one-step-off trainer variant for multi-node split-placement (\run\_mimo\_7b\_mtp\_fully\_async\_megatron\_multinode.sh\). These examples demonstrate configuring MTP auxiliary heads, loss scaling, and integration with the \verl\ training framework, including support for the \uv\ package manager for dependency resolution.

_examples/mtp\trainer · high confidence

New PrefixGrouper example for GRPO training acceleration

Added a new example in the \examples/prefix\_grouper\ directory that demonstrates how to use PrefixGrouper to optimize GRPO training by grouping samples with shared prompts, reducing redundant computations. The entry includes a README explaining the installation, configuration (\use\_prefix\_grouper=True\), and limitations (e.g., FSDP only, incompatible with certain padding/parallelism options), along with a shell script (\run\_qwen3\_8b\_fsdp.sh\) showing a complete setup for Qwen3-8B on GSM8K using vLLM rollout and FSDP training. The example also integrates with the new uv-based dependency management system.

_examples/prefix\grouper · high confidence

New RL training examples and uv integration

Added canonical training scripts and documentation for CISPO, GDPO, GPG, OTB, REINFORCE++, RLOO, and SAPO algorithms, providing ready-to-run configurations for vLLM rollout with FSDP or Megatron training on NVIDIA and Ascend hardware. The examples now support uv-based dependency management on GPU, allowing users to run training via \uv run\ with frozen environments while falling back to system Python on NPU.

(repo-wide) · high confidence

New ROCm Dockerfiles with DeepSeek-V4-Flash GRPO support and bug fixes

The docker/rocm directory now includes a new Dockerfile.rocm and Apptainerfile.rocm that enable DeepSeek-V4-Flash GRPO on AMD GPUs, based on the sglang base image. The primary Dockerfile.rocm upgrades the stack to ROCm 7.14, PyTorch 2.12, vLLM 0.22.1rc1, and SGLang 0.5.15, while applying specific patches to vLLM (VRAM zeroing and MoE weight storage) and SGLang (fused\_add\_rms\_norm crash fix). It also installs mbridge, megatron-core, and verl, and patches megatron-core to warn instead of asserting on attention backend conflicts. A separate Dockerfile.rocm.deepseek-v4 provides a vLLM 0.26.0-based image with Megatron Bridge v0.5.0 and Transformer Engine 2.6.0 for DeepSeek-V4. Historical Dockerfiles for ROCm 6.x and older verl versions are retained as references.

docker/rocm · high confidence

New dataset preprocessing scripts for multiturn, tool-use, and multimodal training

The examples/data\_preprocess directory now includes a comprehensive set of new scripts to prepare datasets for advanced training workflows. These include multiturn and tool-augmented preprocessors for math datasets (AIME 2024, DAPO-Math-17k, GSM8k, MATH, and Geometry3k) that inject system prompts, tool definitions, and ground-truth reward information into the data format. Additionally, new scripts support multimodal and video data, with specific preprocessors for the Open-R1 multimodal dataset and the TinyLLaVA-Video-R1 training data, handling image and video inputs alongside text. The collection also features a generic multiturn conversation generator for testing, a tool-agent loop preprocessor for GSM8k, a SearchR1 dataset preprocessor that handles HuggingFace downloads and search tool integration, and a script for the Pokemon-GPT4O captions dataset. All scripts standardize output to parquet format and support local or HDFS storage.

_examples/data\preprocess · high confidence

New diagnostic and validation scripts for chat templates, model conversion, and environment setup

The \scripts/\ directory now includes several new utilities: \chat\_template\_checker.py\ and \chat\_template\_mock\_trajectories.py\ validate that chat templates support continuous token merging by checking token-level append-only behavior across single-turn and tool-use trajectories; \converter\_hf\_to\_mcore.py\ provides a distributed HuggingFace-to-Megatron Core checkpoint converter with a built-in test mode; \diagnose.py\ collects OS, hardware, Python, pip, network, and package version information to aid in troubleshooting; \init\_random\_model.py\ creates small random-weight models from existing HuggingFace configs for debugging; \generate\_trainer\_config.sh\ flattens Hydra trainer configs into reference YAML files; \legacy\_model\_merger.py\ merges FSDP and Megatron checkpoints into HuggingFace format; \megatron\_merge\_lora.py\ merges LoRA adapters into base weights using the unified engine; \migrate\_megatron\_checkpoint\_layout.py\ updates checkpoint directory structures; and \install\_vllm\_sglang\_mcore.sh\, \install\_vllm\_mcore\_npu.sh\, and \install\_sglang\_mcore\_npu.sh\ automate installation of inference frameworks and dependencies for GPU and NPU backends.

scripts · high confidence

New experimental Docker images for SGLang and vLLM backends

Added two new Dockerfiles (\Dockerfile.sglang056exp\ and \Dockerfile.vllm012exp\) in the \docker/verl0.6.1-experimental\ directory, providing pre-built environments for running verl v0.6.1 with SGLang v0.5.6 and vLLM v0.12.0 respectively. These images include necessary dependencies such as TransformerEngine, Flash Attention, HybridEP, and specific libraries for Qwen3-Next support, enabling users to experiment with these backend configurations.

docker/verl0.6.1-experimental · high confidence

New learning-oriented tutorials for agent loops, Ray, and cloud deployment

Added a new \examples/tutorial\ directory containing documentation and runnable examples for users new to the platform. This includes a Jupyter notebook demonstrating how to train a ReAct agent with a code sandbox using the agent-loop API, a Ray API crash-course notebook, a SLURM job template for running Ray-on-SLURM with configurable network interfaces, and SkyPilot task specifications (YAML files) for launching PPO and GRPO training jobs on Kubernetes or cloud platforms.

examples/tutorial · high confidence

New on-policy distillation trainer examples for Qwen3 and Qwen3.5

The \examples/on\_policy\_distillation\_trainer\ directory now provides a comprehensive set of launch scripts and documentation for on-policy distillation (OPD). Users can now distill knowledge from larger teacher models (e.g., Qwen3-32B, Qwen3.5-35B-A3B) into smaller student models (e.g., Qwen3-0.6B, Qwen3-2B, Qwen3-4B, Qwen3-VL-8B) using vLLM for rollouts and either FSDP2 or Megatron for training. The examples cover single-teacher and multi-teacher (MOPD) setups, support both text and vision-language modalities, and include a VeOmni engine variant that enables fused top-K distillation kernels to reduce GPU memory usage by avoiding full logit materialization. Scripts are also updated to support Huawei Ascend NPUs and integrate with \uv\ for dependency management.

_examples/on\_policy\_distillation\trainer · high confidence

New profiling examples for GRPO runs on NVIDIA and NPU hardware

The examples/profile directory now includes ready-to-run scripts for capturing performance traces during GRPO training. Users can profile PyTorch execution and memory usage on NVIDIA GPUs using vLLM or SGLang rollouts, or capture NPU traces on Ascend hardware. These examples demonstrate how to enable the torch, torch\_memory, and NPU profilers via the Hydra API, allowing users to generate Chrome/Perfetto traces or memory snapshots without writing custom launchers.

examples/profile · high confidence

New rollout correction example scripts for score centering, RLOO, and multi-criteria rejection sampling

Added three new bash scripts in the examples/rollout\_correction directory demonstrating advanced rollout correction configurations: one for REINFORCE with score centering (removing training-inference drift via token-level importance sampling), one for RLOO (REINFORCE Leave-One-Out) with sequence-level self-normalized importance sampling, and one for PPO-clip with multi-criteria rejection sampling (chaining token-level and sequence-level criteria). These scripts provide ready-to-run templates for configuring bypass mode, loss types, and specific rollout correction parameters using the vLLM rollout engine.

_examples/rollout\correction · high confidence

Removals

Removal of Megatron-LM v4 patches

The \megatron\_v4.patch\ file has been deleted, removing the custom modifications previously applied to the Megatron-LM v4 codebase. These patches included updates to argument validation logic in \megatron/arguments.py\, adjustments to pipeline parallel communication in \megatron/core/pipeline\_parallel/p2p\_communication.py\, and changes to scheduling and tensor shape handling in \megatron/core/pipeline\_parallel/schedules.py\. With this patch removed, the project no longer applies these specific behavioral changes to the underlying Megatron-LM library.

patches · high confidence

Removal of single\_controller base and Ray implementation modules

The \single\_controller\ package has removed its core implementation files, including the base worker and decorator logic (\single\_controller/base/worker.py\, \single\_controller/base/decorator.py\), the Ray-specific execution engine (\single\_controller/ray/base.py\, \single\_controller/ray/decorator.py\, \single\_controller/ray/dp.py\), and the Megatron integration classes (\single\_controller/base/megatron/worker\_group.py\, \single\_controller/ray/megatron.py\). This change eliminates the distributed worker group management, data parallelism dispatching, and Megatron model parallelism support previously provided by this module, effectively removing the \single\_controller\ capability from the codebase.

_single\controller · high confidence

Removed Ray tutorial example

The \examples/ray/tutorial.ipynb\ notebook has been deleted. This removes the introductory guide that demonstrated how to initialize a Ray cluster, configure environment variables for Megatron and NCCL, and run distributed training workers with tensor and pipeline parallelism.

examples/ray · high confidence

Behavioural changes

1 commit (0 fixes) modifying docker

A change to existing behaviour in docker — 1 commit, 1 file: “\[hardware, docs\] feat: add multi-stage Intel GPU (XPU) Docker image for vLLM 0.32.0 (\#8401)” (refs \#8401).

docker/xpu, examples/mxfp8 · medium confidence · unverified

1 commit (0 fixes) modifying examples/mxfp4

A change to existing behaviour in examples/mxfp4 — 1 commit, 1 file: “\[megatron, doc\] feat: MXFP4 MoE recipe and Qwen3-30B-A3B example (\#8426)” (refs \#8426).

examples/mxfp4 · medium confidence · unverified

10 commits (3 fixes) modifying docker/verl0.5-cu126-torch2.7-fa2.7.4

A change to existing behaviour in docker/verl0.5-cu126-torch2.7-fa2.7.4 — 10 commits (3 fixs), 13 files. Also in this area: 1 commit (0 fixes) modifying verl; 4 commits (2 fixes) modifying verl; 1 commit (1 fix) fixing verl; Fix backward pass for chunked vocab-parallel entropy; Add VeOmni backend to model merger; and 19 more.

verl · low confidence · unverified

13 commits (4 fixes) modifying examples/tpu/grpo

A change to existing behaviour in examples/tpu/grpo — 13 commits (4 fixs), 6 files: “\[hardware, recipe\] feat: Qwen3-30B-A3B (MoE) GRPO recipe for TPU (\#8451)”; “\[hardware, doc\] fix: make examples/tpu/grpo run on main and add a v6e-4 RayCluster and Dockerfile (\#8225)”; “\[BREAKING\]\[trainer, recipe, ci, doc\] refactor: remove V0 trainer and migrate legacy async trainers to recipes (\#8625)”; and 10 more (refs \#8451, \#8225, \#8625, \#8456, \#8382, …).

examples/tpu/grpo · medium confidence · unverified

2 commits (1 fix) modifying examples

A change to existing behaviour in examples — 2 commits (1 fix), 4 files: “\[fsdp, ci\] fix: optimize Qwen3.5 fsdp-turbo training with fused kernels (\#8508)”; “\[BREAKING\]\[megatron, vllm\] feat: per-token activations for NVFP4 QAT (\#8494)” (refs \#8508, \#8494). Also in this area: 2 commits (0 fixes) modifying examples.

_examples/real\nvfp4 · medium confidence · unverified

Enable FULL\_DECODE\_ONLY cudagraph mode in Ascend PPO example

The run script for the Qwen3-8B FSDP PPO example on Ascend NPU now enables the FULL\_DECODE\_ONLY cudagraph mode via the vLLM engine configuration. This change optimizes the rollout phase for decode-only generation, potentially improving inference throughput during the PPO training loop.

_examples/ascend\_extras/ppo\trainer · high confidence

Migrate to uv for dependency management and pre-commit for code quality

The project has replaced the legacy setup.py/yapf/pylint workflow with a modern, unified dependency and linting system. Users now manage environments via \uv\ (using a single \uv.lock\ file) instead of \requirements.txt\ or \pip\, and code quality is enforced through \pre-commit\ hooks running \ruff\ and \mypy\. This change simplifies environment setup, ensures consistent formatting across the codebase, and improves build reproducibility.

(repo-wide) · high confidence

PPO trainer examples updated to canonical scripts with uv integration and deprecated legacy runs

The PPO trainer examples have been restructured to use a new set of canonical, user-adjustable shell scripts (e.g., \run\_qwen3\_8b\_fsdp.sh\, \run\_qwen3\_8b\_megatron.sh\) that replace the previous legacy scripts (such as \run\_deepseek7b\_llm.sh\ and \run\_qwen2-7b.sh\). These new scripts standardize the configuration interface, support both FSDP and Megatron training backends, and integrate \uv\ for dependency management. The update also includes a comprehensive README documenting the new configuration options, including advanced features like KL divergence control and dual-clip PPO.

_examples/ppo\trainer · high confidence

SFT examples migrated to verl.trainer.sft\_trainer with modern engine support

The GSM8K, multiturn, and VLM SFT examples have been updated to use the unified \verl.trainer.sft\_trainer\ entry point, replacing the legacy \verl.trainer.fsdp\_sft\_trainer\. This change introduces support for multiple training engines including FSDP, Megatron, Megatron Lite, and the new Nemo-Automodel engine. The examples now utilize the \messages\ key for data formatting, support distributed strategies like Ulysses sequence parallelism and LoRA/PEFT via environment variables, and use local directories for checkpoints instead of HDFS. Legacy scripts relying on the old trainer and HDFS paths have been removed.

examples/sft · high confidence

Updated DeepSeek generation example with vLLM 0.8.2 rollout configuration

The DeepSeek LLM generation example has been updated to use the vLLM rollout engine with explicit configuration for multi-node support and correct load formatting. The new script (run\_deepseek\_llm\_7b.sh) sets the rollout load format to 'auto' to ensure compatibility in generation-only runs without a trainer, whereas the previous example (run\_deepseek\_v2\_lite\_math.sh) used a generic rollout configuration. The updated example also introduces user-adjustable parameters for node count, tensor parallel size, and GPU memory utilization, and defaults to the DeepSeek-LLM-7B-Chat model.

examples/generation · high confidence

Fixes

1 commit (1 fix) fixing examples/ascend\_extras/low\_precision

A fix in examples/ascend\_extras/low\_precision — 1 commit (1 fix), 1 file: “\[doc\] fix: fix ascend tutorial links and navigation (\#8573)” (refs \#8573).

_examples/ascend\_extras/low\precision · low confidence · unverified

The .claude and .codex directories now use symlinks to point to the shared skills located in .agent/skills. This change ensures that both .claude and .codex reference the same underlying skill definitions, maintaining consistency across different agent configurations.

.claude, .codex · high confidence

Test coverage

1 commit adding/updating tests in tests/utils/qat; 3 commits adding/updating tests in tests; 7 commits adding/updating tests in tests/trainer; Added Ascend NPU quick-start test scripts for Qwen3-0.6B; Added CPU tests for reward manager configurations and registry; Added CPU-based tests for dataset utilities; Added CPU-based unit tests for PPO core algorithms and metrics; Added CPU-only unit tests for V1 PPO trainer components; Added NPU test scripts for Qwen3 and Qwen3.5 models; Added SFT engine accuracy alignment tests; Added comprehensive test suite for the checkpoint engine; Added detached worker test suite for Ray-based training; Added end-to-end test for PPO trainer with custom function rewards; Added integration tests for sandbox fusion correctness checks; Added model regression and integration tests; Added multi-GPU unit tests for distributed training stability and correctness; Added nightly CI test scripts for multiple models and parallelism strategies on Ascend NPU; Added performance benchmarking script for vLLM async rollout backends; Added rollout worker test coverage; Added tests for FSDP model merger and output validation; Added tests for Megatron router replay utilities; Added tests for Muon optimizer support and precision-aware optimizer behavior; Added tests for PRIME reward scoring and sandbox fusion; Added tests for SGLang rollout HTTP server and LoRA handling; Added tests for experimental VLA simulation environments; Added tests for experimental reward loop and async policy configurations; Added tests for function-based tool registration and mixed tool loading; Added tests for single controller dispatch mode registration and updates; Added tests for the experimental agent loop; Added tests for the platform abstraction layer and plugin system; Added tests for utility modules in tests/utils; Added tests for vLLM rollout stability and determinism on Ascend NPU and CPU; Added tests for worker engine correctness and stability; Added unit and integration tests for TRT-LLM rollout workers; Added unit tests for VeOmni MoE parameter export and router replay utilities; Added unit tests for checkpoint manager logic and configurations; Added unit tests for the single controller module; Added unit tests for worker configuration dataclasses; Expanded test coverage for protocol, configuration, and distributed worker health; New CI sanity checks for code quality and conventions; New end-to-end test suite for multi-GPU and specialized trainer configurations.

Dependencies

Migrate to pyproject.toml and uv for dependency management

The project replaces the legacy requirements.txt-based setup with a structured pyproject.toml configuration, introducing a unified dependency resolution system via uv. This change organizes dependencies into modular extras (verl-core, vllm, sglang, fsdp, megatron, ci) to handle mutually exclusive backends and specific hardware requirements (such as CUDA 13 and NPU support) within a single lockfile. It also updates core library versions, including upgrading numpy to \>=2.0.0, pinning transformers to 5.12.1 for vllm/sglang backends, and adding new dependencies like tensorboard, fastapi, and uvicorn.

(dependencies) · high confidence

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

How this codebase got here

Score

  • CAI 48.5 → 40.8 (-7.7)
  • Rubric changed (rubric-2026.11.2 → rubric-2026.11.4) — scores are not directly comparable.

Lenses

  • Code Health 59.9 → 59.1 (-0.8)
  • Architecture 84.8 → 96.0 (+11.2)
  • Maturity 71.5 → 72.2 (+0.7)
  • Readiness 50.3 → 49.5 (-0.8)
  • Security 34.3 → 17.1 (-17.2)
  • Performance 100.0 → 100.0 (0.0)

Resolved (165)

  • Duplicated block (10 lines × 2) (verl/utils/qat/fp4.py)
  • Duplicated block (10 lines × 2) (verl/utils/qat/fp4.py)
  • Duplicated block (10 lines × 3) (verl/experimental/one_step_off_policy/ray_trainer.py)
  • Duplicated block (10–15 lines × 2) (verl/experimental/fully_async_policy/fully_async_main.py)
  • Duplicated block (11 lines × 2) (verl/utils/skip/rollout_skip.py)
  • Duplicated block (11 lines × 3) (verl/experimental/separation/ray_trainer.py)
  • Duplicated block (12 lines × 2) (verl/experimental/fully_async_policy/fully_async_trainer.py)
  • Duplicated block (12 lines × 2) (verl/trainer/ppo/ray_trainer.py)
  • Duplicated block (12 lines × 2) (verl/utils/kernel/fp8_kernel.py)
  • Duplicated block (12 lines × 3) (verl/models/transformers/dense_common.py)
  • Duplicated block (12–13 lines × 2) (verl/experimental/one_step_off_policy/ray_trainer.py)
  • Duplicated block (12–14 lines × 2) (verl/experimental/fully_async_policy/fully_async_rollouter.py)
  • Duplicated block (13 lines × 2) (verl/experimental/one_step_off_policy/ray_trainer.py)
  • Duplicated block (13 lines × 2) (verl/experimental/separation/ray_trainer.py)
  • Duplicated block (13 lines × 2) (verl/experimental/separation/ray_trainer.py)
  • Duplicated block (13 lines × 2) (verl/trainer/ppo/ray_trainer.py)
  • Duplicated block (13–14 lines × 2) (verl/experimental/fully_async_policy/fully_async_trainer.py)
  • Duplicated block (14 lines × 3) (verl/experimental/one_step_off_policy/ray_trainer.py)
  • Duplicated block (14–15 lines × 2) (verl/experimental/one_step_off_policy/ray_trainer.py)
  • Duplicated block (14–15 lines × 2) (verl/trainer/main_ppo_v0.py)
  • …and 145 more

New (181)

  • Duplicated block (10 lines × 3) (verl/models/mcore/model_forward_fused.py)
  • Duplicated block (11 lines × 3) (verl/models/transformers/qwen3_5.py)
  • Duplicated block (11 lines × 7) (verl/models/transformers/dense_common.py)
  • Duplicated block (13 lines × 2) (verl/models/transformers/glm4v.py)
  • Duplicated block (15 lines × 2) (verl/workers/rollout/sglang_rollout/async_sglang_server.py)
  • Duplicated block (16 lines × 2) (verl/utils/kernel/tilelang/linear_cross_entropy.py)
  • Duplicated block (22 lines × 3) (verl/models/transformers/dense_common.py)
  • Duplicated block (5 lines × 2) (verl/utils/kernel/tilelang/linear_cross_entropy.py)
  • Duplicated block (5 lines × 2) (verl/workers/engine/fsdp/transformer_impl.py)
  • Duplicated block (6 lines × 2) (verl/models/mcore/model_forward_fused.py)
  • Duplicated block (7 lines × 7) (verl/models/mcore/model_forward_fused.py)
  • Duplicated block (8 lines × 2) (verl/workers/engine/fsdp/transformer_impl.py)
  • Duplicated block (8 lines × 5) (verl/models/transformers/dense_common.py)
  • Edited copies of one member (3 members, 11+ corresponding lines) (verl/utils/tokenizer/continuous_token.py)
  • Edited copies of one member (3 members, 12+ corresponding lines) (verl/checkpoint_engine/hccl_checkpoint_engine.py)
  • Edited copies of one member (3 members, 16+ corresponding lines) (verl/workers/engine/fsdp/transformer_impl.py)
  • Edited copies of one member (3 members, 18+ corresponding lines) (verl/experimental/agent_loop/tool_parser.py)
  • Edited copies of one member (3 members, 24+ corresponding lines) (verl/utils/megatron/tensor_parallel.py)
  • Edited copies of one member (3 members, 26+ corresponding lines) (verl/workers/engine/automodel/transformer_impl.py)
  • Edited copies of one member (3 members, 34+ corresponding lines) (verl/utils/flops_counter.py)
  • …and 161 more

Changes since last survey

  • 47 commits — 25 feature/other, 22 fixes

By area

  • .github/workflows — 8 commits
  • docs/ascend_tutorial — 8 commits
  • verl/workers — 5 commits
  • examples/tpu — 4 commits
  • tests/workers — 3 commits
  • verl/trainer — 3 commits
  • docker/ascend — 2 commits
  • tests/utils — 2 commits
  • (root) — 1 commit
  • docker/rocm — 1 commit
  • docs/advance — 1 commit
  • docs/index.rst — 1 commit
  • examples/grpo_trainer — 1 commit
  • tests/checkpoint_engine — 1 commit
  • tests/experimental — 1 commit
  • tests/trainer — 1 commit
  • verl/checkpoint_engine — 1 commit
  • verl/experimental — 1 commit
  • verl/models — 1 commit
  • verl/single_controller — 1 commit

Notable commits

  • fix: [ci, hardware] feat: Expand NPU regression coverage and fix HCCL cleanup (#8619)
  • fix: [ci] fix: apply deterministic rollout settings to multimodal agent tests (#8521)
  • fix: [ci] fix: create log directory before multinode Ascend test (#8540)
  • fix: [ci] fix: run vLLM Ascend reward-loop CI on 16 NPUs (#8671)
  • fix: [ci] fix: shrink engine test models (#8539)
  • fix: [ci] fix: use all 16 NPUs in ci tests (#8561)
  • fix: [doc] fix: add Ascend recipe repository links (#8665)
  • fix: [doc] fix: clarify Ascend image and Qwen3-ASR patch guidance (#8630)
  • fix: [doc] fix: fix ascend tutorial links and navigation (#8573)
  • fix: [doc] fix: repair navigation in Ascend installation guide (#8566)
  • fix: [doc] fix: require Python >=3.12 for SGLang on Ascend (#8569)
  • fix: [doc] fix: standardize Atlas training and Ascend 950 product names (#8606)
  • fix: [env, doc] fix: align Ascend source installation with Dockerfiles (#8679)
  • fix: [env] fix: avoid duplicate Qwen3-ASR registration on Ascend (#8563)
  • fix: [reward] fix: parse bold Answer labels in math_dapo Minerva scoring (#8653)
  • fix: [rollout, vllm, sglang] fix: launch rollout servers when workers hold no accelerator; per-replica env-var hook (#8529)
  • fix: [trainer, single_controller, perf] fix: stop applying the OMP_NUM_THREADS default to every Ray worker (#8592)
  • fix: [trainer] fix: exclude all-padding rows from per-sequence off-policy metrics (#7785)
  • fix: [vllm, hardware] fix: enable vLLM-Ascend RL mode by default (#8585)
  • fix: [vllm, rollout] fix: preserve the length stop reason in token outputs (#8680)
  • …and 27 more

Architecture

  • Unchanged — 0 containers · 1 contexts · 0 edges

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

Survey your own repository

verl-project/verl was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.

About this page

  • The score is its most recent published measurement, taken on 16 November 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
  • Measured at commit 909bb91387b96fd20dddc86cb9c026b536f0f05e — the exact code this score is about.
  • Scored under rubric-2026.11.4 — the same rubric and the same method as every other entry in this index.
  • Measured by Canine Code Analyzer (cca.canine.dev) using cca measure-2026.11.83+preprod-a489942cf94b.