volcengine/verl
Feb 2026 – Mar 2026
2 Months active
Languages Used
Python
Technical Skills
CUDADistributed SystemsGPU ProgrammingPyTorchdeep learningmachine learning
Eric Marcus
PROFILE
Eric Marcus
Worked on the volcengine/verl repository to enhance reinforcement learning and vision-language model training pipelines by integrating the Liger kernel and improving GPU memory management. Addressed stability and scalability by fixing out-of-bounds vocabulary handling in logsumexp computations and preventing out-of-memory errors during multi-GPU startup with PyTorch and CUDA. Delivered kernel-level optimizations that reduced memory usage and accelerated training, while maintaining API compatibility for end users. Expanded test coverage to ensure robust handling of edge cases and compatibility across model types. Demonstrated expertise in distributed systems, deep learning, and performance optimization, resulting in faster, more reliable model training workflows.
Overall Statistics
Feature vs Bugs
33%Features
Repository Contributions
3Total
Bugs
2
Commits
3
Features
1
Lines of code
180
Activity Months2
Your Network
444 peopleSame Organization
@kaiko.ai4
Shared Repositories
440
Work History
March 2026
1 Commits • 1 Features
Mar 1, 2026March 2026 monthly summary for volcengine/verl focusing on delivering the Liger kernel integration across vision-language models and reinforcement learning (RL) training, with stability improvements and broader test coverage. The work emphasizes business value through faster training, lower memory usage, and seamless compatibility without API changes. Key features delivered: - Liger Kernel Integration with Vision-Language Models and RL Training: the integration now passes explicit parameters to the kernel application, enables swiglu option, and maintains performance gains during model training. A new test verifies Liger compatibility with VL models (tests/models/test_liger_vl_compat.py). An API usage path remains unchanged for end users. Major bugs fixed: - Fixed Liger integration issues with VL models and RL training when fused kernels are enabled, by updating the call to _apply_liger_kernel_to_instance with explicit kwargs (fused_linear_cross_entropy=False, swiglu=True). This eliminates forward conflicts and related crashes, and broadens compatibility across model types. Overall impact and accomplishments: - Business value: faster training iterations and reduced memory footprint lead to lower training costs and faster experimentation cycles. Benchmarks show significant improvements in representative scenarios: - Up to 1.49x faster forward (with liger enabled) and up to ~28% less peak GPU memory in certain VL configurations; RL training observed ~21% faster step times (4×4 H200 setup). - VL multimodal path shows ~1.41x faster forward/backward (baseline ~411 ms, ligers ~292 ms). - No API changes; config use_liger=True remains the same, but now works correctly with all model types and with fused kernels. Technologies/skills demonstrated: - Python, PyTorch FSDP and distributed training, kernel-level optimization, performance benchmarking, test-driven development, and CI readiness. Demonstrated ability to debug complex integration points, implement stable defaults (fused_linear_cross_entropy=False, swiglu=True), and expand test coverage to prevent regressions.
1 Commits • 1 Features
Mar 1, 2026March 2026 monthly summary for volcengine/verl focusing on delivering the Liger kernel integration across vision-language models and reinforcement learning (RL) training, with stability improvements and broader test coverage. The work emphasizes business value through faster training, lower memory usage, and seamless compatibility without API changes. Key features delivered: - Liger Kernel Integration with Vision-Language Models and RL Training: the integration now passes explicit parameters to the kernel application, enables swiglu option, and maintains performance gains during model training. A new test verifies Liger compatibility with VL models (tests/models/test_liger_vl_compat.py). An API usage path remains unchanged for end users. Major bugs fixed: - Fixed Liger integration issues with VL models and RL training when fused kernels are enabled, by updating the call to _apply_liger_kernel_to_instance with explicit kwargs (fused_linear_cross_entropy=False, swiglu=True). This eliminates forward conflicts and related crashes, and broadens compatibility across model types. Overall impact and accomplishments: - Business value: faster training iterations and reduced memory footprint lead to lower training costs and faster experimentation cycles. Benchmarks show significant improvements in representative scenarios: - Up to 1.49x faster forward (with liger enabled) and up to ~28% less peak GPU memory in certain VL configurations; RL training observed ~21% faster step times (4×4 H200 setup). - VL multimodal path shows ~1.41x faster forward/backward (baseline ~411 ms, ligers ~292 ms). - No API changes; config use_liger=True remains the same, but now works correctly with all model types and with fused kernels. Technologies/skills demonstrated: - Python, PyTorch FSDP and distributed training, kernel-level optimization, performance benchmarking, test-driven development, and CI readiness. Demonstrated ability to debug complex integration points, implement stable defaults (fused_linear_cross_entropy=False, swiglu=True), and expand test coverage to prevent regressions.
March 2026
February 2026
2 Commits
Feb 1, 2026February 2026 (2026-02) focused on RL training reliability and multi-GPU memory stability for volcengine/verl. Key outcomes include a correctness fix in the logsumexp pathway and a memory-management improvement enabling robust startup when vLLM runs colocated with FSDP. The changes enhance model reliability, training stability, and scalability with minimal operational risk. Key deliverables: - Fixed out-of-bounds vocab handling in the logsumexp computation for reinforcement learning training, masking phantom tokens before the running max/exp while preserving the original logits for entropy accumulation. This prevents inflated log-probabilities and entropy and mitigates severe token-mismatch scenarios in peaked softmax, with validation tests added. - Prevented vLLM OOM during startup on colocated GPU setups by appending aggressive_empty_cache(force_sync=True) at the end of ActorRolloutRefWorker.init_model(), a no-op when no memory to free and validated on a large-scale 8B VLM + FSDP configuration. - Expanded test coverage with test_lce_non_divisible_vocab_padding to cover non-divisible vocabulary edge cases and ensure GPU-specific scenarios are exercised. Overall impact and accomplishments: - Increased RL training stability and accuracy by ensuring correct probability/entropy calculations and reducing token-mismatch risk. - Improved deployment resilience in multi-process GPU environments, enabling safer scaling and startup under FSDP/vLLM coexistence. - Maintained API compatibility with no user-facing changes needed. Technologies and skills demonstrated: - Triton-based masking in logsumexp kernels and tl.where usage for safe OOB handling. - PyTorch CUDA allocator memory management practices, including aggressive_empty_cache patterns. - FSDP/vLLM orchestration and multi-process GPU memory considerations. - Test-driven development with GPU-relevant test coverage and CI alignment.
February 2026
2 Commits
Feb 1, 2026February 2026 (2026-02) focused on RL training reliability and multi-GPU memory stability for volcengine/verl. Key outcomes include a correctness fix in the logsumexp pathway and a memory-management improvement enabling robust startup when vLLM runs colocated with FSDP. The changes enhance model reliability, training stability, and scalability with minimal operational risk. Key deliverables: - Fixed out-of-bounds vocab handling in the logsumexp computation for reinforcement learning training, masking phantom tokens before the running max/exp while preserving the original logits for entropy accumulation. This prevents inflated log-probabilities and entropy and mitigates severe token-mismatch scenarios in peaked softmax, with validation tests added. - Prevented vLLM OOM during startup on colocated GPU setups by appending aggressive_empty_cache(force_sync=True) at the end of ActorRolloutRefWorker.init_model(), a no-op when no memory to free and validated on a large-scale 8B VLM + FSDP configuration. - Expanded test coverage with test_lce_non_divisible_vocab_padding to cover non-divisible vocabulary edge cases and ensure GPU-specific scenarios are exercised. Overall impact and accomplishments: - Increased RL training stability and accuracy by ensuring correct probability/entropy calculations and reducing token-mismatch risk. - Improved deployment resilience in multi-process GPU environments, enabling safer scaling and startup under FSDP/vLLM coexistence. - Maintained API compatibility with no user-facing changes needed. Technologies and skills demonstrated: - Triton-based masking in logsumexp kernels and tl.where usage for safe OOB handling. - PyTorch CUDA allocator memory management practices, including aggressive_empty_cache patterns. - FSDP/vLLM orchestration and multi-process GPU memory considerations. - Test-driven development with GPU-relevant test coverage and CI alignment.
Activity
Loading activity data...
Quality Metrics
Correctness100.0%
Maintainability86.6%
Architecture86.6%
Performance93.4%
AI Usage33.4%
Skills & Technologies
Programming Languages
Python
Technical Skills
CUDADeep LearningDistributed SystemsGPU ProgrammingMachine LearningPerformance OptimizationPyTorchTestingdeep learningmachine learningtesting
Repositories Contributed To
Overview of all repositories you've contributed to across your timeline