[monolithic change] Rework the way CUDA-Q embeds in Python #3693

schweitzpgi · 2025-12-18T00:45:36Z

This is a huge change that crosscuts almost every aspect of the system (except for the C++ frontend, maybe?).

The main drivers are:

Replace all the random, ad hoc compiler pipelines being used in the Python compiler front-end with pipelines that the C++ compiler uses
To achieve the previous bullet required
- Changing pipelines that were run backwards or multiple times
- Getting rid of ad hoc caches of information kept at all levels of the code
- Deploying proper lambda lifting and beta reduction steps at definition and call, resp.
- Radical brain surgery on Python kernel launchers (algorithms) so that they could properly resolve symbols in context
- Fixing the misuse of MLIR symbols
- Removing ad hoc, blind, and error prone pruning of MLIR code
- Fixing bugs due to the conflation of kernel definition and kernel call semantics
- Fixing tests with errors, bugs, etc.
This includes folding (back) in changes to the Python ast_bridge (Python bridge revision #3537)

Signed-off-by: Eric Schweitz <[email protected]>

The current implementation of the Python handling of CUDA-Q has baked in various attempts to deal with the language coupling between Python and CUDA-Q kernels. These solutions have been accumulating and making it more and more difficult to work on the Python implementation. These changes are a total rewrite to bring the Python implementation more closely aligned with the C++ implementation. Changes: - The kernel builder and kernel decorator are fundamentally different and will no longer share a duck-typed interface. It doesn't work well. The builder assembles a CUDA-Q kernel dynamically. As such all symbolic references are known immediately. The decorator converts a static AST of code into a CUDA-Q kernel. Symbolic references are either local or not. Non-local symbols are unknown at the point the decorator is processed. All non-local symbols in a decorator are recorded with the decorator itself and lambda lifted as actual arguments. - MLIR requires that symbols be uniqued. The previous implementation ignored this requirement. - Lazy state maintenance in Python and the C++ runtime layers is buggy and not needed. It is removed. This includes dangling MLIR bindings from the AST bridge's python MLIR bindings. - Kernels are no longer built with assumptions, then rebuilt when those guesses prove wrong. Kernels are no longer built and rebuilt for different steps in the process. A kernel decorator builds a target agnostic, context independent kernel, and saves that MLIR ModuleOp under a unique name. - Launch scenarios have been reworked and consolidated to use the ModuleOp directly instead of shuffling between string representations (possibly under maps that were not thread-safe) and ModuleOp instances. - Every step of the process creating a brand new MLIRContext and loading all the dialects into that context, etc. is removed. This is done once and the Python interpreter uses the same context to build all modules. Other changes include: Fix GIL issue in get_state_async. Restructure lambda lifting so it handles recursive ops. Clone the ModuleOps as they go into the make_copyable_function closure to prevent them from being erased along the way. Remove VQE tests. Use VQE from CUDA-QX! Simplifying cudaq::state support. Handle kernel decorator from import module Simplify the symbol table. Python is not a scoped language other than LEGB. Convert kernel builder to generate code compatible with C++ for state initialization of veqs. Refactor the AST bridge to generate state objects from the runtime. Fixes for various tests. Signed-off-by: Eric Schweitz <[email protected]> Co-authored-by: Thien Nguyen <[email protected]>

Signed-off-by: Eric Schweitz <[email protected]>

* Fixes for alternate module resolution. Teaches kernel launching to understand how to deal with qualified symbols and the symbol lookup rules that are entailed from that. Signed-off-by: Eric Schweitz <[email protected]> * Fix module seeking issues. Signed-off-by: Eric Schweitz <[email protected]> * Remove debug print. Signed-off-by: Eric Schweitz <[email protected]> --------- Signed-off-by: Eric Schweitz <[email protected]>

sync with main

* Don't implicit handle density matrix as a flattened vector Signed-off-by: Thien Nguyen <[email protected]> * Format Signed-off-by: Thien Nguyen <[email protected]> * Fix Clang build Signed-off-by: Thien Nguyen <[email protected]> * Code review: fix style and add test Signed-off-by: Thien Nguyen <[email protected]> * Fix clang build Signed-off-by: Thien Nguyen <[email protected]> * More clang build fix Signed-off-by: Thien Nguyen <[email protected]> --------- Signed-off-by: Thien Nguyen <[email protected]>

Much of this no longer makes much sense, but port the parts of the tests that do make sense to the new Python refactoring. Specifically verboten: It does not make sense to try and synthesize Python kernel decorators or look them up by non-unique short names in C++ code. The python interpreter knows what kernel decorators it is referencing and must do the resolution of symbols to bindings. Signed-off-by: Eric Schweitz <[email protected]>

Signed-off-by: Eric Schweitz <[email protected]>

Signed-off-by: Thien Nguyen <[email protected]>

Signed-off-by: Eric Schweitz <[email protected]>

Fixes for Python/C++ interop.

Signed-off-by: Eric Schweitz <[email protected]>

Signed-off-by: Thien Nguyen <[email protected]>

Signed-off-by: Eric Schweitz <[email protected]>

Signed-off-by: Thien Nguyen <[email protected]>

Signed-off-by: Eric Schweitz <[email protected]>

Correcting the last weight (#3654)

… with a nullptr guard Signed-off-by: Thien Nguyen <[email protected]>

Signed-off-by: Eric Schweitz <[email protected]>

…ests Signed-off-by: Eric Schweitz <[email protected]>

[features/py-redesign] Fix for C++ mock server tests

Signed-off-by: Eric Schweitz <[email protected]>

sacpis · 2025-12-30T19:28:21Z

/ok to test 47806d2