Repository navigation
Conversation
tbennun
force-pushed
the
worktree-ai-library-nodes
branch
from
September 7, 2026 07:49
7e32b3a to
9943722
Compare
…bject, and default schedule/storage inference
Updated pytest commands to exclude tests marked with 'ai'.
| ml = ["onnx", "torch", "onnxsim", "onnxscript", "onnxruntime", "protobuf", "ninja"] | ||
| mpi = ["mpi4py", "pytest-mpi"] | ||
| gpu = ["cupy-cuda13x", "cutensor-cu13"] | ||
| ai-anthropic = ["anthropic>=1.0"] |
Collaborator
There was a problem hiding this comment.
For simplicity we can enforce installing both anthropic and open ai when ai is requested?
Collaborator
Author
There was a problem hiding this comment.
sure, I wanted to not call the dependency ai though to not confuse users with DaCeML. Thoughts?
Collaborator
There was a problem hiding this comment.
llm, llm-expansions, agent, gen-ai are some that come to my mind, agent might make sense the most?
| :param sdfg: The SDFG containing the state. | ||
| :return: The complete expansion context. | ||
| """ | ||
| # Computed once and shared: it deep-copies the SDFG, and every part of the context that |
Collaborator
There was a problem hiding this comment.
Same, do we really need the full SDFG? Isn't the concept of connector+tasklet is to have information needed for a tasklet locally and derive it just from the input output data?
Collaborator
Author
There was a problem hiding this comment.
The deep copy can be removed in principle. However, we will need to refactor some code to do a read-only/immutable connector type inference. It is the same code that sets the default schedule and storage types (infer_types.py) when a library node is expanded.
…odes uv.lock: took main's lockfile and re-ran uv lock to add the ai-anthropic and ai-openai extras. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Rename the "responses" provider to "openai" (OpenAIProvider). ai.api_key_envvar now defaults to empty, meaning each provider's conventional variable. - Manual provider: replace fence-stripping regex with strip(). - Add a prompt/response exchange API: LibraryNode.generate_prompt(state) and read_prompt_response(state, response), plus SDFG-level dace.libraries.ai.generate_prompts / read_prompt_responses. They share the manual provider's files; a supplied response that does not compile is reported instead of sent to a provider for repair. - Context: use dtypes.can_access instead of ad-hoc storage tuples, define the missing logger, and only deep-copy the SDFG for default inference when the node's slot still says Default. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Warnings are reported through warnings.warn with the exception text in the message; informational messages are printed only when debugprint is set. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Collaborator
Author
|
cscs-ci run |
Remove the yapf pre-commit hook, .style.yapf and the yapf pin in the linting group.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nstead of editing sys.path Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Environment names must be valid identifiers and are used verbatim, so that "my-env", "my env" and "my_env" no longer map to one module. A different definition under an existing name is no longer overwritten. Both cases are sent back to the model as a repair round, like a compile failure. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The answer cache, generated environments, sessions and manual-provider prompts now live in ai-cache, ai-environments, ai-sessions and ai-prompts under build_folder_root() (.dacecache by default) instead of ~/.dace or the system temporary directory. A hyphen cannot occur in an SDFG name, so these cannot clash with a program's build folder. The four *_dir configuration entries are removed. A session id is now the node's path from the root SDFG (states, nested SDFG nodes, node), percent-encoded so that different names never share a session. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Storage inference gives a nested SDFG's non-transient the storage of the outer container, whatever it declares. A node expanded before compilation read the declared storage, so a single element of GPU memory became a by-value connector, and the host loaded it straight out of device memory (a segfault). Connector types now use the resolved storage, and the inference runs whenever a nested container's storage differs from the outer one. infer_types.get_storage_from_parent is made public for this. The documentation now says when to expand relative to transformations that move data. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ts own The check no longer assembles a translation unit from strings. The tasklet is placed in a single-tasklet SDFG with the containers, memlets, symbols and environments of the node it replaces (inside a one-iteration GPU_Device map for device code), and built with sdfg.compile() in <build folder root>/ai-check/<session id>, which later attempts on the same slot reuse. The check therefore uses the real frame code, state struct, CMake environment resolution and CUDA or HIP toolchain, and has no hardcoded timeout. The model gets the compiler errors and the generated sources of the check program. The configuration entry ai.verify is renamed to ai.compile_check. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Parallel builds interleave the compiler output of different files, which breaks attributing diagnostics to tasklets. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ubset helpers - Connector types come from infer_types (infer_in_connector_type is extracted from infer_connector_types for this), run on the copy whose defaults are resolved when needed, instead of a hand-copied rule. - Layout is read from the memlet's own container, which is what codegen offsets the connector pointer from, also for a chain of views. - The index formula uses Range.absolute_strides and contiguity uses Range.is_contiguous_subset; both were wrong for stepped subsets. - Symbols in scope come from SDFGState.symbols_defined_at, which also covers loop and interstate symbols. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- compiler.host_cpu() describes the CPU through py-cpuinfo when it is installed (covering macOS and Windows) and through /proc/cpuinfo otherwise; host_isa_id() reads the same fields as before. - common.local_gpus() lists the GPUs of the selected backend, through pynvml (nvidia-ml-py) or amdsmi (ROCm) when they can be imported and through nvidia-smi or rocm_agent_enumerator otherwise. Only the selected vendor is queried, and none of these packages is a dependency. dace.libraries.ai.sysinfo is removed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…stalled Environments without an is_installed check were listed as available on the machine. They are now listed separately, as registered but possibly missing, and the prompt asks to reuse an environment whenever one fits. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…evel getattr/hasattr give way to the attributes the node and descriptor classes define, isinstance checks, and the environment registry. Nested imports remain only for the optional SDKs and the provider modules, which import the backend. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The offline isolation moves from the conftest into ai_test_utils.offline(), which the __main__ blocks use as well, so a direct run cannot be served cached stub answers. Live tests run with the defaults, so their answers stay cached. The Anthropic and manual provider tests keep pytest.main, since every test there needs a fixture. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Docstrings and comments of the AI library and of the core code it touches are reduced to what the code does. The manual provider gets the node name from the expansion instead of parsing it out of the prompt, and an unknown environment in use_environments is sent back to the model like other environment errors. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ons and sample Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n created An environment written by a failed attempt stayed registered, so the repair of that attempt could not change it and lost a round to a name conflict. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…age for the FFTW test The FFTW test now skips unless a program using FFTW compiles and links, and uses the default build folder, so that saved manual-provider answers and the answer cache are found again. The GPU tests that need cupy are skipped before they make a model request. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The system prompt claimed that every thread of the block runs the body and that threadIdx and blockIdx are free to use. DaCe maps each iteration of the innermost map onto one thread and skips the body on threads past the iteration space, so code that split its work across threadIdx computed only thread 0's share. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A fresh expansion of a node whose session already had rounds salted the cache key with the round number, so every rerun of a program made a new request. Refinements never needed the salt, since their feedback already changes the conversation. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ed GPU scalar test A copy from pageable host memory can return before the data reaches the device, and DaCe copies the result out on a stream of its own, so the test read zero about half of the time. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… schedule in the prompt - The state struct is emitted before code_global, so the system prompt now says that a state field's type must be built in or come from an environment's headers, and that a type defined in code_global is held through a `struct name *` pointer, allocated in code_init and deleted in code_exit. A test compiles and runs that pattern. - The environment listing shows the headers each environment includes in the file of the generated code, and its state fields, which the system prompt told the model to consult. - The capabilities no longer show the node's own schedule and the enclosing block size. A node in a kernel resolves to GPU_ThreadBlock, while its tasklet runs on one thread; the outline already shows the schedules of the enclosing scopes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The tile GEMM description of ai_cpu_test named a reduction bound N that the SDFG does not define; it now states the size. - The iteration feedback asked for a helper with static linkage, which the system prompt advises against, because every tasklet's code_global shares one translation unit. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ports The instrumentation environments (LIKWID, PAPI) register when code is first generated, and the FFT, MPI, PBLAS and sparse environments when their library is imported. The environment listing in the prompt, and so the answer cache key, depended on whether a program had been compiled before in the same process: a test run alone missed the answers cached by a run of its file. The context module now imports them. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ydantic The AI library needs pydantic, which only the agent extras provide, so the offline AI tests failed collection in CI. General and GPU CI now install the agent extra, and the test directory is skipped when pydantic is missing. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Loading the check program and unloading it when the handle was collected broke the later kernels of the process: after the check of a host cuBLAS call, the kernel launch of the next GPU program hung in cudaLaunchKernel (5 of 5 runs of the two live GPU tests from the answer cache) or failed with an invalid device kernel image. The check only needs the build. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The ResolvedDefaults.nodes field shadowed the nodes module in its own annotation, which Python 3.10-3.13 evaluate eagerly, so importing dace.libraries.ai failed there. Python 3.14 defers annotations and was unaffected. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR provides two elements:
'ai'implementation, available on every library node. Expanding with it asks a language model to write the replacement tasklet, given the node's context in the SDFG and on the local machine.dace.libraries.ai.nodes.ai_node.AINode, a library node that carries only a natural-language description, for microkernels that have no library node of their own.ai.refineallows AI sessions to refine the implementation (e.g., if it crashes or not fast enough)The model is reached through a provider selected by the
ai.providerconfiguration entry. It supports the Anthropic Messages API, the OpenAI Responses API, and arbitrary AI models through amanualmode that lets you paste prompts and responses.