@@ -136,57 +136,6 @@ APRIL not only improves training efficiency but also achieves:
136136
137137## 🏗️ Architecture
138138
139- ### System Design
140-
141- ``` mermaid
142- graph TB
143- subgraph Pipeline["🎯 APRIL Training Pipeline"]
144- subgraph Rollout["📊 Rollout Phase"]
145- R1[("🎲 Over-provision<br/>N' > N requests")]
146- R2[("⚡ SGLang<br/>Inference Engine")]
147- R3[("🛑 Active<br/>Interruption")]
148- R1 --> R2
149- R2 --> R3
150- end
151-
152- subgraph Buffer["💾 Buffer Management"]
153- B1[("📦 Partial<br/>Rollouts")]
154- B2[("♻️ Resume<br/>Queue")]
155- B3[("✅ Complete<br/>Samples")]
156- B1 --> B2
157- R3 --> B1
158- R3 --> B3
159- end
160-
161- subgraph Training["🧠 Training Phase"]
162- T1[("🔄 Policy<br/>Update")]
163- T2[("📈 Loss<br/>Computation")]
164- T3[("⚙️ Megatron/<br/>FSDP Backend")]
165- B3 --> T2
166- T2 --> T1
167- T1 --> T3
168- end
169-
170- B2 -.->|Next Iteration| R1
171- T3 -.->|Updated Model| R2
172- end
173-
174- style Pipeline fill:#f9f9ff,stroke:#4a5568,stroke-width:2px
175- style Rollout fill:#e6f7ff,stroke:#1890ff,stroke-width:1px
176- style Buffer fill:#fff7e6,stroke:#fa8c16,stroke-width:1px
177- style Training fill:#f0f5ff,stroke:#597ef7,stroke-width:1px
178-
179- style R1 fill:#e6f7ff,stroke:#40a9ff
180- style R2 fill:#e6f7ff,stroke:#40a9ff
181- style R3 fill:#e6f7ff,stroke:#40a9ff
182- style B1 fill:#fff7e6,stroke:#ffa940
183- style B2 fill:#fff7e6,stroke:#ffa940
184- style B3 fill:#fff7e6,stroke:#ffa940
185- style T1 fill:#f0f5ff,stroke:#85a5ff
186- style T2 fill:#f0f5ff,stroke:#85a5ff
187- style T3 fill:#f0f5ff,stroke:#85a5ff
188- ```
189-
190139### Core Components
191140
192141| Component | Path | Description |
0 commit comments