AITOPAI热报
How-toMulti-sourceSelected

Nvidia's Ujval Kapasi to Discuss Agentic PyTorch Performance on Vera Rubin

官方账号PyTorch@PyTorch
Post· 11 sources
View the post
Why we picked it

Agentic pipelines shift the performance bottleneck from single-model inference to the full chain of tool calls, data movement and distributed execution, so framework-level tuning matters more than raw GPU throughput. Engineers building agents on PyTorch, Triton or vLLM can use the co-design lessons to plan their stack for Vera Rubin-class hardware. The keynote is a marker for where Nvidia's agent infrastructure work is heading.

Nvidia's Ujval Kapasi will present "PyTorch Performance for Agents on Vera Rubin" at PyTorchCon 2026. The keynote examines what agentic workloads, which chain reasoning, tool calls, code execution and distributed inference in a single request, demand from the PyTorch stack. Using Vera Rubin as a case study, he covers hardware and software co-design across CPUs, GPUs, memory and networking. The talk also details optimizations across PyTorch, Triton and vLLM on Vera Rubin.

Original · PyTorch

Ujval Kapasi (@nvidia) examines what agentic workloads need from the PyTorch stack in “PyTorch Performance for Agents on Vera Rubin” at #PyTorchCon 2026.

A single request can trigger long chains of reasoning, tool calls, code execution, data movement, and distributed inference, introducing new bottlenecks and optimization opportunities.

Using Vera Rubin as a case study, he explores hardware and software co-design across CPUs, GPUs, memory, and networking. The keynote covers software optimizations and work across PyTorch, Triton, and vLLM on Vera Rubin.

#Nvidia#PyTorch#Vera Rubin#vLLM#Triton

Keep reading

Recent stories with the same tag or section
IndustryUS vs ChinaMulti-sourceSelected71°16:11
Groq plans DRAM-stacked inference chip to fix SRAM capacity

The bonding method determines the outcome: 3DS adds capacity at DRAM speeds, while hybrid bonding delivers bandwidth much closer to SRAM, deciding how few chips are needed to serve large models fast. Inference infrastructure teams evaluating Groq-class LPUs versus Nvidia GPUs, and anyone tracking Cerebras, which plans DRAM stacking on its CS-6 wafer, should watch which route ships first. Capacity, not bandwidth, is emerging as the binding constraint across SRAM-first architectures.

#Groq#Nvidia#LPU#Cerebras
Event
rohanpaul_ai@rohanpaul_ai11 sourcesSource
IndustryUS vs ChinaMulti-sourceSelected14:19
Morgan Stanley pegs 2027 AI spending at US$1.58T

Treating leases as capex materially inflates measured AI investment and shifts scrutiny onto off-balance-sheet financing structures that most revenue models ignore. The supplier split also shows how much of the spending leaks beyond US-listed names to Korean memory and Taiwanese hardware manufacturers. Investors tracking AI capex sustainability, and neocloud credit risk in particular, get a clearer map of where the financing strain concentrates.

#Morgan Stanley#Nvidia#hyperscalers#neoclouds
Event
rohanpaul_ai@rohanpaul_ai11 sourcesSource
PapersBuildersMulti-sourceSelected23:43
Nvidia paper scores base models by first passing edit

Base-model choice before expensive post-training has lacked a reliable signal, and this proxy gets it at 0.964 correlation with SWE-bench Verified outcomes. Teams running RL or SFT on coding agents can apply it with existing benchmark traces instead of full fine-tuning trials. It compares directly with running standard agent benchmarks, which barely predict downstream agent skill on base models.

#Nvidia#SWE-bench Verified#base models#agents
Event
rohanpaul_ai@rohanpaul_ai11 sourcesSource
PapersBuildersMulti-sourceSelected02:32
Nvidia VERA paper: alternating model training and skill edits lifts agents

Per-step scoring from real artifacts tells you which step broke — a diagnosis final-answer benchmarks cannot give. Teams building agents can apply the alternating update loop without new training infrastructure, using VERA's sandbox design as a template. It also gives a concrete comparison point: on the same 9B model, skipping skill edits drops the score to 43.3, below even skill-edit-only training.

#Nvidia#VERA#agents#open-source models
Event
rohanpaul_ai@rohanpaul_ai11 sourcesSource
ModelsBuildersMulti-sourceSelected00:18
Nvidia Long-WAM framework

The framework demonstrates how extending temporal context dramatically improves robotic task success rates, potentially enabling more complex autonomous behaviors. Developers working on robotics and AI agents should evaluate how this approach could enhance their applications, particularly for tasks requiring longer-term planning and memory.

#Long-WAM#Nvidia#RoboCasa GR-1#world-action models
Event
Aran Komatsuzaki (论文推介)@arankomatsuzaki11 sourcesSource
PapersBuildersMulti-sourceSelected73°19:56
Nvidia paper reveals AI model reliability decline

This reveals a fundamental limitation in current AI systems that affects practical applications more than previously acknowledged. Developers and businesses building AI agents should implement item numbering and batch processing to mitigate reliability issues, especially for long document processing tasks.

#NVIDIA#AI models#reliability#agents
Event
rohanpaul_ai@rohanpaul_ai11 sourcesSource