TorchSpec merges DFlash2 training support with Qwen3-8B recipe
zhyncs42 · x · 2026-08-21
TorchSpec has merged a PR adding DFlash2 training support. The update introduces DFlash2 configurations, dynamic causal convolution, and candidate selection mechanisms. It also exports the public model configuration and checkpoint schemas used by SGLang and vLLM, and adds a Qwen3-8B training recipe with focused tests.
More from Infra
- AI Datacenters Bypass Grid with On-site Power, Rely on Model Routing for Uptime — shensi · 2026-08-21
- AI Data Centers Bypass Grid Delays with On-Site Gas and Solar — shensi · 2026-08-21
- Unsloth Dynamic V3 GGUFs: Q3 Outperforms Larger Q4 Models — danielhanchen · 2026-08-21
- VC Trend: Invest in Own GPU Clusters to Offer Compute at Cost to Portfolio Companies — beffjezos · 2026-08-21
- Heterogeneous Systems Outperform Frontier Models in New Benchmarks — ShahabBakht · 2026-08-21
- AMD MI300x vs NVIDIA H100: Real-world agent coding benchmark — locker73 · 2026-08-21