Own inference performance and correctness for pre-training evaluations in a research lab focused on open AI models.
Posted by employer 3 days ago
First seen on Joblaze 2 days ago
Last verified on the company career page 10 hours ago
Skills & Technologies
What you'll build
Must have
Requirements
Not disclosed in this posting: compensation, years of experience.
Benefits
Joblaze summary
In this role, the individual is responsible for ensuring the performance and accuracy of inference in pre-training evaluations, working closely with both researchers and engineers. Key skills include a deep understanding of distributed systems, GPU architecture, and inference processes, along with a methodical approach to optimization and validation. This position is suited for someone with a strong background in machine learning infrastructure and a proven track record in performance improvement. Reflection AI offers a collaborative environment focused on foundational models, emphasizing rigorous technical problem-solving.
Joblaze insights
Quick facts
From the original posting
You will own inference performance and correctness for pre-training evaluations. You will be bridging between our training framework to our inference framework. Your work will span model implementations, GPU execution, distributed serving, and the lifecycle of inference workers, from making a checkpoint ready to serve through completing an evaluation.
Working closely with researchers and engineers in pre-training and post-training, you will track ongoing research work, implement them efficiently, and establish the validation needed to deploy them confidently. You will help turn workload-specific tuning and one-off debugging into reliable systems that keep pace with new models, hardware, and research needs.
Establish parity between ongoing ML research modeling in training framework and inference framework via defining numerical acceptance criteria based on activations, logits, and downstream evaluation results.
Optimize inference for pre-training by profiling CPU and GPU execution, memory use, communication, and scheduling to understand where the time and resource go. Improve the components that limit real evaluation workloads, and validate that changes improve throughput or cost while preserving correctness.
Improve autoscaling and fleet efficiency by determining which signals best capture demand and available serving capacity, such as queued work, token processing, cache pressure, and server health.
You have a track record of improving performance or reliability in complex distributed systems and inference infrastructure .
You understand how CPU execution, GPU architecture, memory hierarchy, and interconnects shape performance. You can reason quantitatively about compute limits, memory bandwidth, communication overhead, and the tradeoffs between different forms of parallelism.
You understand inference, including prefill and decode, batching, attention, KV-cache management, and distributed model execution, and can connect those concepts to observed system behavior.
You approach optimization through careful measurement: form a hypothesis, design a useful experiment, identify the bottleneck, and demonstrate the effect of the change.
You bring rigor to correctness. You can distinguish implementation errors, expected numerical differences, and experimental noise, and design validation that catches meaningful regressions.
You take ownership of ambiguous technical problems by discovering and locating issues independently, communicate findings clearly, and work effectively with researchers and engineers to carry improvements into regular use.
Top-tier compensation: Salary and equity structured to recognize and retain our talent globally.
Standard company text repeated across Reflection AI's postings is omitted here.