Agentic Compiler Engineer

2 нед. назад

ФранцияSeniorHybrid

compilergpulow-level programmingoptimizationtransformerllmpython

Develop and optimize an agentic compiler for LLM execution on GPUs.

builds a co-designed inference stack for real-time AI agents on standard datacenter GPUs, spanning model architecture, inference engine, compilers, and low-level GPU kernels.
On the model side, we developed Laneformer 2B and Delayed Tensor Parallelism (DTP), a Transformer architecture that overlaps communication with useful computation and weight streaming.
On the systems side, the Inference Engine runs this stack on standard AMD and NVIDIA datacenter GPUs.
generates 3,500 tokens/s per request on 8 AMD MI300X GPUs and 2,100 tokens/s per request on 8 NVIDIA H200 GPUs, in FP16 at batch size 1, with quantization and speculative decoding disabled.
Our next major project is AGCO, our agentic compiler. AGCO is designed to optimize LLMs across different GPUs and optimization targets, including very fast inference.
The team has 10 people, including 9 engineers and researchers and 4 PhDs.
Test it at playground..ai. Read the technical details on the Labs blog.

We look for engineers with deep technical expertise and original work in at least one area relevant to AGCO.
Relevant experience includes:
Compiler engineering, including optimization passes, IRs, lowering, code generation, LLVM, or MLIR.
GPU programming with CUDA, HIP, Metal, Vulkan, or similar technologies.
GPU performance work involving kernels, memory, synchronization, profiling, or hardware behavior.
LLM inference engines and performance optimization.
Attention, MoE, parallelism, communication, or other systems-level parts of LLM execution.
Formal verification, equivalence checking, SAT/SMT, or related methods.
Systems that generate, search, test, benchmark, or optimize code automatically.
We care about what you personally built and the technical decisions behind it. Strong candidates can explain the problem, their approach, the alternatives they explored, and how they measured the result.
We review technical work during the process. This can be public code, an upstream contribution, a paper, a thesis, a technical project, or a detailed write-up based on work you can share.

You will join a small team building AGCO as a core part of 's technology.
Work at the intersection of compilers, GPU systems, and LLM inference.
Direct access to engineers working across the full inference stack.
A fast loop from an optimization idea to compilation, execution, verification, and measurement on real GPUs.
The opportunity to go deep in your strongest technical area while expanding into the other parts of the stack.
High ownership over technical decisions and systems that will shape how optimizes LLM inference.
This role is based in Paris, and we are looking for candidates who can relocate to Paris and work closely with the team.

You will work directly on AGCO.
The goal is to build a system that can explore ways to optimize LLM execution, generate changes, compile them, check correctness, run them on real hardware, measure the results, and use this feedback to guide the next optimization.
You will contribute to areas such as:
Compiler and IR design for representing and transforming LLM computations.
Optimization passes, lowering, and code generation.
Search methods for exploring different implementations and execution strategies.
Verification and correctness checks for generated changes.
GPU execution, profiling, and performance optimization.
LLM inference across operators, memory, parallelism, and communication.
Optimization loops that connect generated changes to measurements on real GPUs.
One direction we are exploring combines an IR, a verifier, a compiler, and a search optimizer. We plan to start with focused problems, build working prototypes, and extend the system from what we learn.
Your main area will depend on your experience, skills, and interests. You may focus more on compilers, GPU systems, or LLM inference while working closely with people across the full stack.