AI-generated implementations for DaCe library nodes

.github/workflows/general-ci.yml

@@ -65,7 +65,7 @@ jobs:
         else
             export DACE_optimizer_automatic_simplification=${{ matrix.simplify }}
         fi
-        uv run  --group testing pytest -n auto --cov-report=xml --cov=dace --tb=short --timeout_method thread --timeout=300 -m "not gpu and not autodiff and not torch and not onnx and not tensorflow and not mkl and not sve and not papi and not mlir and not lapack and not mpi and not scalapack and not datainstrument and not long and not sequential"
+        uv run  --group testing pytest -n auto --cov-report=xml --cov=dace --tb=short --timeout_method thread --timeout=300 -m "not gpu and not autodiff and not torch and not onnx and not tensorflow and not mkl and not sve and not papi and not mlir and not lapack and not mpi and not scalapack and not datainstrument and not long and not sequential and not ai"
         ./codecov
 
     - name: Test OpenBLAS LAPACK

dace/config_schema.yml

@@ -584,6 +584,140 @@ required:
                         default_Darwin: ''
                         default_Windows: ''
 
+    ai:
+        type: dict
+        title: AI-generated code
+        description: >
+            Preferences for the "ai" library node implementation, which expands a library node
+            into a tasklet written by a language model.
+        required:
+            provider:
+                type: str
+                default: 'anthropic'
+                title: Model provider
+                description: >
+                    Which backend to generate code with. "anthropic" uses the Anthropic Messages
+                    API (pip install 'dace[ai-anthropic]'); "responses" uses the Responses API of
+                    the openai package (pip install 'dace[ai-openai]'); "manual" prints the prompt and
+                    reads the reply back from the terminal, needing neither a key nor an SDK. The
+                    SDK is only imported when an expansion actually runs.
+
+            manual_dir:
+                type: str
+                default: ''
+                title: Manual provider directory
+                description: >
+                    Directory the "manual" provider writes prompts to and reads replies from.
+                    Defaults to a "dace_ai_prompts" directory under the system temporary directory
+                    when empty.
+
+            api_key_envvar:
+                type: str
+                default: 'ANTHROPIC_API_KEY'
+                title: API key environment variable
+                description: >
+                    Name of the environment variable holding the provider's API key. If
+                    the variable is unset, the provider's own credential chain is used instead.
+
+            model:
+                type: str
+                default: 'claude-opus-5'
+                title: Model
+                description: Identifier of the model used to generate tasklets.
+
+            effort:
+                type: str
+                default: 'high'
+                title: Reasoning effort
+                description: >
+                    How much effort the model should spend per expansion, for providers that
+                    support it (low, medium, high, xhigh, max).
+
+            max_tokens:
+                type: int
+                default: 64000
+                title: Maximum response tokens
+                description: Upper bound on the length of a single generated response.
+
+            timeout:
+                type: int
+                default: 600
+                title: Request timeout
+                description: Timeout, in seconds, for one request to the model provider.
+
+            verify:
+                type: bool
+                default: true
+                title: Verify generated code
+                description: >
+                    Compile the generated tasklet in isolation before it is inserted into the
+                    SDFG, and send any compiler diagnostics back to the model. This catches
+                    syntax and type errors; it does not check that the result is correct.
+
+            max_repair_attempts:
+                type: int
+                default: 2
+                title: Maximum repair attempts
+                description: >
+                    How many times the model may be asked to fix generated code that does not
+                    compile. Only used when verification is enabled.
+
+            environment_dir:
+                type: str
+                default: ''
+                title: Generated environment directory
+                description: >
+                    Directory where library environments requested by the model are written as
+                    Python modules, so that they can be inspected, edited, and resolved again in
+                    later sessions. Defaults to ~/.dace/ai_environments when empty.
+
+            cache:
+                type: bool
+                default: true
+                title: Reuse model answers
+                description: >
+                    Store every answer on disk, keyed by the provider, the model, the effort
+                    setting and the whole conversation, and reuse it when the same question is
+                    asked again. Re-expanding an unchanged library node then costs nothing. Any
+                    change to the SDFG, the node, or the model changes the key and misses the
+                    cache, so a stale answer cannot be picked up.
+
+            cache_dir:
+                type: str
+                default: ''
+                title: Answer cache directory
+                description: >
+                    Where reusable answers are stored. Defaults to ~/.dace/ai_cache when empty.
+                    Deleting the directory, or any file in it, only forces the question to be
+                    asked again.
+
+            transcripts:
+                type: bool
+                default: true
+                title: Record expansion transcripts
+                description: >
+                    Write the prompts, the model's answers, and the source and diagnostics of every
+                    probe compilation to disk, one directory per expansion. An expansion is a paid
+                    and non-reproducible call whose result is baked into the SDFG, so the material
+                    that produced it is kept by default. Set DACE_debugprint=verbose to see the
+                    same material on the terminal as it happens.
+
+            transcript_dir:
+                type: str
+                default: ''
+                title: Transcript directory
+                description: >
+                    Where expansion transcripts are written. Defaults to ~/.dace/ai_transcripts
+                    when empty.
+
+            extra_instructions:
+                type: str
+                default: ''
+                title: Extra prompt instructions
+                description: >
+                    Additional text appended to every code generation prompt, e.g. project-wide
+                    coding conventions or libraries that should be preferred.
+
     instrumentation:
         type: dict
         title: Instrumentation

dace/libraries/ai/__init__.py

@@ -0,0 +1,32 @@
+# Copyright 2019-2026 ETH Zurich and the DaCe authors. All rights reserved.
+"""
+AI-generated implementations for DaCe library nodes.
+
+This library provides two things:
+
+1. The reserved ``'ai'`` implementation, available on **every** library node -- including ones
+   defined outside DaCe. 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. See
+   :mod:`dace.libraries.ai.expansion`.
+2. :class:`~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.
+
+The model is reached through a provider selected by the ``ai.provider`` configuration entry; its
+SDK is imported lazily, so nothing here requires one to be installed until an expansion actually
+runs.
+"""
+
+from dace.library import register_library
+from dace.sdfg.nodes import AI_IMPLEMENTATION_NAME
+
+from .expansion import ExpandAI
+from .exceptions import AIExpansionError
+from .nodes import AINode
+
+# Re-register environments generated by earlier sessions, so that an SDFG expanded previously
+# still resolves the libraries its tasklets were linked against.
+from .environments import load_generated_environments
+
+load_generated_environments()
+
+register_library(__name__, "AI")