PhD · AI Infrastructure · ML Systems · Distributed Training
Paris, France
Open to work & relocation
I build AI infrastructure for large-scale LLM training, focusing on hybrid parallelism, performance modelling and automated strategy planning. I received my PhD in Computer Science from Sorbonne Université in July 2026, with doctoral research at EURECOM and Huawei Paris through the CIFRE programme.
This research led to planners now integrated into MindSpore and MindFormers. The systems have been evaluated on dense, mixture-of-experts (MoE) and multimodal models. Across these projects, experiments span Ascend-910 clusters with up to 16,384 NPUs and models with over 1T parameters. This work has also supported production training, including China Telecom’s DeepSeek-MoE deployment and TeleChat3-MoE. See the systems and publications below.
How a model is partitioned and scheduled across accelerators affects training time, memory use and cost. Choosing a parallelisation strategy often requires extensive profiling and manual tuning for each model and cluster.
How can we plan parallelism for large-scale training without per-cluster profiling, while retaining portability across models and hardware?
I develop analytical cost models and optimisation methods to estimate the performance and memory requirements of a training strategy before execution. The systems below address memory prediction, pipeline scheduling, joint optimisation, incremental re-planning and user-specified constraints.
Can Hybrid-Parallel Planning Support Alternating Model–Strategy Design? Dependency-Keyed Per-Layer Primitives for Re-Planning.
Reuses per-layer primitives during hybrid-parallel re-planning, recomputing only those affected by a model edit.
Median re-planning speedup of ~2.7× across audited model edits, with provably identical plan selection to a full search; evaluated on dense and MoE models using up to 4,096 NPUs.
PRISM: Profiling-Free Symbolic Memory-Driven Strategy Planner for Large DNN Model Training.
Osaka, Japan · Jan 26–29, 2026
Plans hybrid-parallel training without profiling, using symbolic memory models, estimates that account for recomputation, and topology-aware communication costs.
92–96% memory prediction accuracy (median error ≈ 7%) across DeepSeek3, Llama2-3, Mistral, TextHawk and Llama-MoE on up to 1,024 Ascend-910 NPUs; up to 1.43× the model FLOPs utilisation (MFU) of Megatron-LM.
BMPipe: Bubble-Memory Co-optimization Strategy Planner for Very-Large DNN Training.
27th IEEE CLUSTER · Edinburgh, UK · Sep 2–5, 2025
Uses integer linear programming (ILP) to jointly optimise pipeline execution and memory use, accounting for recomputation, load imbalance and preparation bubbles.
1.36× speedup over Megatron-Even on a 16,384-NPU Ascend-910 cluster; up to 1.70× with 1F1B-Interleave; the ILP is solved in under 200 ms at production scale.
H²O: Holistic Hyper-Parameter Optimization for Large-Scale Deep Neural Network Training.
Poster · 31st Euro-Par · Dresden, Germany · Aug 25–29, 2025
A two-level optimiser combines symbolic cost modelling with an ILP to jointly plan hybrid-parallel configurations, layer-to-stage assignments and recomputation.
36.7% speedup over the D-Rec baseline on 128-device clusters; 35.7% MFU on DeepSeek-141B — without profiling on accelerators.
ManuMatic: Strategy Injection for Robust Automatic Hybrid Parallelism in Distributed DNN Training.
22nd IFIP NPC · Nha Trang, Vietnam · Nov 14–16, 2025
Lets users specify parallelisation strategies for selected operators, while a D-Rec-based planner determines a consistent strategy for the rest of the graph.
Up to 2.24× speedup over D-Rec on Mixtral-8×7B with expert parallelism (8 NPUs); 2.04× on Llama3-8B at 8K context; 1.45× / 1.30× on Qwen2.5-72B at 32K context (64 NPUs, combined with BMPipe).
Explore training strategies with PRISM, BMPipe and H²O. Configure a model and cluster, compare candidate strategies and view their pipeline timelines.
Open live demo ↗Select Launch search study to start a search.
MLLM Pipeline Bubble Modeling for Large-Scale Training.
SCA / HPCAsia 2026, Osaka, Japan · Jan 26–29, 2026
SCOPE — Symbolic Computation-Memory Optimization for Pipeline Efficiency in Ultra-Scale DNN Training.
DP2E-AI 2025 Workshop, Paris, France · Jun 2–6, 2025
Devices and methods for generating execution plans for a machine learning model.
WO 2025/020165 A1 · PCT/CN2023/109533 · Huawei Technologies · filed Jul 27, 2023 · published Jan 30, 2025
A load balancing method to improve large model performance.
Patent application · Huawei Technologies · publication pending
A multi-dimension parallelism configuration method for large language model training.
Patent application · Huawei Technologies · publication pending
Advisors, committee members, mentors, co-authors and friends who have supported my research.