Framework Comparison

How Flux compares to VERL, AReaL, and Slime.

Quick Comparison

Aspect	VERL	AReaL	Slime	Flux
Sync Strategy	Fixed sync	Fixed async	Manual	Adaptive
Orchestration	Ray	Custom	HTTP	asyncio
Training	Megatron/FSDP	Custom	Megatron	Megatron
Inference	vLLM/SGLang	Custom	SGLang	SGLang
Weight Sync	Ray Store	Custom	CUDA IPC	CUDA IPC
Staleness	N/A	Aware	APRIL	Unified
Code Size	~15k LOC	~25k LOC	~8k LOC	<5k LOC

Why Not Ray?

Ray adds overhead for LLM training:

Ray Provides	LLM Needs	Mismatch
Task scheduling	NCCL collectives	Ray doesn't understand NCCL
Actor lifecycle	Fine-grained GPU control	Ray treats GPU as "count"
Flexible placement	Fixed TP/PP/DP config	Changing requires restart
Object Store	CUDA IPC	Serialization is slow

Both VERL and Slime bypass Ray for critical paths.

Sync vs Async

VERL:  ████████████░░░░░░░░░░░░░░░░  Stable but slow
AReaL: ░░░░░░░░░░░░░░░░████████████  Fast but risky
Flux:  ◄═══════ adapts here ═══════►  Best of both

Performance Targets

Metric	VERL	AReaL	Flux
GPU Util	~45%	~95%	>80%
Stability	High	Medium	High
Throughput	1x	2x	1.8x

When to Use Each

Use Case	Recommendation
Maximum stability	VERL or Flux
Maximum throughput	AReaL or Flux
SGLang native	Slime or Flux
Simple setup	Flux
Production	Flux

See Also

• Architecture
• Technical Specification