HAMi Meetup Shanghai Deep Dive (1): Where Does HAMi Fit in the CNCF × PyTorch Cloud Native AI Foundation?
The model has been able to run and the operation is slow to go online. Keith Chan, Director General of the China District, gave an opening speech at the HAMMI Meetup Shanghai Station, giving a structured collaborative idea of how to create a base for the development of claud-native AI: Hami is located in Scheduling with resource management, assuming the role of heterogenouscompute resources sharing and managing.
Read articleHAMi Meetup Shanghai Deep Dive (7): From User to Co-Builder — Why Enterprises Need a Sustainable HAMi Community
Once an enterprise uses open source software in its production environment, the concerns will gradually extend beyond code: can it be responded to when problems are encountered? How do new needs enter public routes? The Hami Meetup Shanghai Station round-table discussion revolves around real use, contribution thresholds, AI ' s liability boundary for assistive development and community knowledge deposition, providing a different set of judgement than technology selection.
Read articleHAMi Meetup Shanghai Deep Dive (2): After GPU Sharing — Isolation, Scheduling, and Observability in HAMi 2.10
The distribution of a GPU to multiple tasks is only the beginning of resource sharing. Dynamia co-founder and CTO, HAMi Mainlainer Mengxuan Li, at the HAMi Meetup Shanghai Station, presented the progress of V2.10 capabilities, DRA ecological appliance, and v2.11 planning (Remote GPU, CPU NUMA alignment). The main thread for these changes is to enable shared compute resources to be distributed on demand and to be subject to monitoring, observation and diagnosis during their operation.
Read articleHAMi Meetup Shanghai Deep Dive (4): Turning GPU Performance into Inference Service Capability — Lessons from Iluvatar CoreX
GPU computing performance is good, and users are still waiting for the first token. Iluvatar CoreX Infra R&D Director Yip Sung Lin has deployed a high-performance GPU deployment practice at HAMi Meetup Shanghai Station: to implement optimization, to expand the division of labour between Scheduling and inference configurations, to end-to-end SLO standards for acceptance, and to use the HAMi shared GPU path for small model scenarios.
Read articleHAMi Meetup Shanghai Deep Dive (3): Inference Control vs. GPU Resource Management — How llm-d and HAMi Can Work Together
The same model deploys multiple copies, and requests are divided on average, but do not necessarily respond faster. CTO Zhang Jiayong in Red Hat Grand Central China shares the llm-d distributed inference structure at the Hami Meetup Shanghai station: prefixing sensory paths, how PD separation and cache offloading can be coordinated, and the HAMi resource layer collaboration direction discussed during the question and answer session - requesting scheduling and GPU cut is a two-tiered task that needs to be linked and cannot be replaced by one another.
Read articleHAMi Meetup Shanghai Deep Dive (6): How UCloud Turns HAMi Sharing into Ready-to-Use Development Environments
The GPU is in place, and the algorithm team is still waiting for the environment. UCloud Senior Research and Development Engineer Peng shares the Platform’s Engineering Challenges and Practice at Hami Meetup Shanghai Station: Harmonization of Examples, Multi-cluster Access, Tool Injection and Mirror Acceleration to Shorten the Environmental Delivery Chain; privatization scenarios use Hami vGPU with common clouds and QEMU transmission paths.
Read article