We accidentally built an LLVM compiler for Jax

2 days ago

The Catalyst quantum compiler for PennyLane accidentally supports pure classical JAX code, compiling it to MLIR and LLVM without quantum instructions.
The team chose JAX because it can trace Python functions, supports NumPy and SciPy APIs, and lowers to MLIR, while adding native Python control flow and dynamic shapes.
The goal was to compile entire hybrid quantum-classical workflows into standalone executables, not just quantum instructions.
Catalyst bypasses XLA, lowering JAX representations directly to standard MLIR dialects and then to LLVM, enabling use of LLVM tools like Enzyme for backpropagation.
Benefits include standalone AOT binaries without ML runtime dependencies, support for dynamically-shaped arrays, native Python control flow, and ease of custom MLIR passes.
This approach may be useful for edge ML, custom hardware, and compiler infrastructure, though it won't beat XLA for standard deep learning workloads.

Differentiable Fortran allows backpropagation through existing Fortran, C, or C++ simulation code using LFortran and Enzyme, enabling integration with modern ML frameworks like JAX and PyTorch.
Enzyme applies autodifferentiation at the LLVM IR level, allowing differentiation of any code that compiles to LLVM, avoiding the need to rewrite validated legacy physics code.
LFortran is used instead of Flang because it emits cleaner LLVM IR with plain pointer arithmetic, making it easier for Enzyme to trace memory accesses without opaque runtime calls.

Nativ allows running open models from Google, Cohere, and Liquid AI locally on Apple Silicon Macs, with model recommendations tailored to hardware.
Features include a clean interface with streaming, markdown, code highlighting, image input, and performance monitoring like tokens/sec and thermal state.
Built on MLX-VLM and optimized for M-series unified memory and Metal, with no wrappers or translation layers.

Package management functionality moved from compiler to build system, improving modularity and allowing patching without compiler rebuild.
SPIR-V backend progressed with new @SpirvType builtin, execution mode via calling convention, and multi-threaded codegen.
New @bitCast semantics defined by logical bit layout, making behavior endian-agnostic and enabling new operations like integer to vector conversion.

AI infrastructure in 2026 is driven by cycles where useful techniques become marketed as features, often as walled-garden products.
Loop engineering is a basic programming construct (a 'while' statement), not a product category or feature to be purchased.
The core of loop engineering is the turn orchestrator, which manages loop turns and exposes lifecycle hooks.

NoiseLang is a probabilistic programming language where every value is a probability distribution, including constants as Dirac distributions.
It enables operations on random variables directly, with features like independent draws and Monte Carlo simulation for queries such as probability, expectation, or variance.
The language includes an efficient runtime with multiple backends: a columnar interpreter, a Cranelift JIT for native kernels, and WASM for browser execution, ensuring deterministic results across platforms.

Differentiable Fortran allows backpropagation through existing Fortran, C, or C++ simulation code using LFortran and Enzyme, enabling integration with modern ML frameworks like JAX and PyTorch.
Enzyme applies autodifferentiation at the LLVM IR level, allowing differentiation of any code that compiles to LLVM, avoiding the need to rewrite validated legacy physics code.
LFortran is used instead of Flang because it emits cleaner LLVM IR with plain pointer arithmetic, making it easier for Enzyme to trace memory accesses without opaque runtime calls.

Nativ allows running open models from Google, Cohere, and Liquid AI locally on Apple Silicon Macs, with model recommendations tailored to hardware.
Features include a clean interface with streaming, markdown, code highlighting, image input, and performance monitoring like tokens/sec and thermal state.
Built on MLX-VLM and optimized for M-series unified memory and Metal, with no wrappers or translation layers.

Package management functionality moved from compiler to build system, improving modularity and allowing patching without compiler rebuild.
SPIR-V backend progressed with new @SpirvType builtin, execution mode via calling convention, and multi-threaded codegen.
New @bitCast semantics defined by logical bit layout, making behavior endian-agnostic and enabling new operations like integer to vector conversion.

AI infrastructure in 2026 is driven by cycles where useful techniques become marketed as features, often as walled-garden products.
Loop engineering is a basic programming construct (a 'while' statement), not a product category or feature to be purchased.
The core of loop engineering is the turn orchestrator, which manages loop turns and exposes lifecycle hooks.

NoiseLang is a probabilistic programming language where every value is a probability distribution, including constants as Dirac distributions.
It enables operations on random variables directly, with features like independent draws and Monte Carlo simulation for queries such as probability, expectation, or variance.
The language includes an efficient runtime with multiple backends: a columnar interpreter, a Cranelift JIT for native kernels, and WASM for browser execution, ensuring deterministic results across platforms.

Hasty Briefsbeta