AI infrastructure and dev tools updates #61
Today's Letter
NVIDIA, DFlash speculative decoding for Blackwell
- NVIDIA introduced DFlash, an open-source block diffusion model for speculative decoding on NVIDIA Blackwell GPUs.
- The drafter generates a full token block in one forward pass, then the target model verifies candidates in parallel.
- On gpt-oss-120b with TensorRT-LLM on an eight-GPU DGX B300 system, NVIDIA reports up to 15x higher throughput at the same interactivity level.
- At 500-600 tokens/sec per user, DFlash is reported at 1.5x higher throughput than EAGLE-3 on the same Blackwell setup.
- NVIDIA also says DFlash nearly doubles interactivity for Llama 3.1 8B at the same concurrency versus EAGLE-3.
- The company released 20 DFlash checkpoints on Hugging Face and is expanding support across TensorRT-LLM, SGLang, and vLLM.
Source: developer.nvidia.com
GitHub, Copilot merge totals by adoption phase
- GitHub added total merged pull request counts by AI adoption phase to Copilot usage reports on June 26, 2026.
- The totals_by_ai_adoption_phase breakdown now includes total_pull_requests_merged for each phase and day.
- The new field complements the existing avg_pull_requests_merged metric instead of replacing it.
- GitHub says the metric is available in both 1-day and 28-day Copilot usage reports.
- Teams can measure each phase's share of merged pull requests and compare absolute throughput across cohorts.
- Access is limited to enterprise administrators and organization owners with Copilot usage metrics REST API access.
- GitHub says total_pull_requests_merged uses the same attribution model as avg_pull_requests_merged for consistency.
Source: github.blog
Jocoletter curates AI, software, and product trends for developers and builders.
#GitHub #NVIDIA