中文版

CodeChecker

1. Purpose

Static checker executed on generated kernel code before KernelVerifier. Uses ast.parse, py_compile, importlib, and regex; does not invoke an LLM or run code on a device. Catches errors that would fail verification (syntax, unresolvable imports, non-identifier text in code) and enforces project conventions that the runtime verifier does not detect (kernel defined and launched; forward() does not call torch compute APIs; @triton.autotune declares restore_value).

2. Callers

Invoked by every workflow that generates kernel code:

Caller	File
default_workflow / default_workflow_v2	op/workflows/default_workflow*.py
coder_only_workflow	op/workflows/coder_only_workflow.py
kernelgen_only_workflow	op/workflows/kernelgen_only_workflow.py
autoresearch_workflow	op/workflows/autoresearch_workflow.py
AutoResearch agent tools	op/autoresearch/agent/tools.py

Each caller constructs CodeChecker(backend, dsl) and awaits checker.check(code). On failure the caller forwards error_message to the coder / KernelGen for the next retry.

3. Pipeline

Six stages, run in order. Stages 3–6 are skipped when stage 1 or 2 reports an error.

#	Stage	Implementation	Rejected cases
1	Syntax	ast.parse	unmatched brackets, fullwidth punctuation, indentation errors
2	Compile	py_compile.compile(doraise=True)	errors ast.parse does not raise (duplicate kwargs, SyntaxWarning promoted to error)
3	Imports	importlib.util.find_spec on each top-level import target	module name not resolvable in the current environment
4	Non-identifier text	tokenize → regex over code tokens (comments and strings excluded)	runs of ≥ min_run consecutive characters from a configured Unicode range
5	DSL compliance	AST walk over the module and the kernel class body	no @<triton>.<decorator> kernel defined; kernel defined but never launched via kernel[grid](...); forward() calls a method in torch_compute_ops_hard; forward() calls a method in torch_compute_ops_soft while no kernel is launched
6	Autotune	regex + paren matching	@<triton_module_name>.<decorator_attr>(...) missing <required_kwarg>=

Stages 5 and 6 run only when dsl.startswith(dsl_compliance_prefix) (default prefix triton).

4. Policy (op/config/code_checker.yaml)

The checker's keyword sets and identifiers are defined in op/config/code_checker.yaml. The Python module has no fallback defaults. A missing or ill-typed key raises KeyError or re.error at module import.

Key	Type	Meaning
triton_decorators	list[str]	attribute names under <triton_module_name> recognized as kernel decorators
torch_call_prefixes	list[str]	local aliases or dotted torch namespaces whose .method(...) calls are scanned in forward()
torch_compute_ops_hard	list[str]	methods rejected in forward() regardless of whether a kernel is launched
torch_compute_ops_soft	list[str]	methods rejected in forward() only when no kernel is launched
kernel_class_name	str	class name the AST walker searches for in stage 5
kernel_forward_method	str	method within that class whose body is scanned
triton_module_name	str	top-level namespace for the Triton decorator match
dsl_compliance_prefix	str	DSL-string prefix that activates stages 5–6
stray_text.min_run	int	minimum consecutive characters required to flag
stray_text.unicode_ranges	list of [lo, hi]	code-point ranges scanned in stage 4
autotune.decorator_attr	str	attribute name off <triton_module_name> that triggers stage 6
autotune.required_kwarg	str	keyword argument that must appear in the decorator call

The YAML is read once at module import via importlib.resources.files("akg_agents.op.config") and is packaged as package_data (see setup.py **/*.yaml).

5. Customizing

All rule changes are YAML edits; no Python changes required.

• Move an op between torch_compute_ops_hard and torch_compute_ops_soft to change when it is rejected. The shipped policy places layer_norm, batch_norm, pooling, interpolate, cumsum, cumprod in the soft list so Ascend seeds using them as kernel pre-processing are accepted.
• Remove an op from both lists to accept it unconditionally.
• Append method names to either list.
• Append local aliases or dotted namespaces to torch_call_prefixes; examples include F, torch.nn.functional, nn.functional, torch.linalg, torch.fft, torch.special, and torch.sparse.
• Update kernel_class_name / kernel_forward_method when the kernel module convention changes.
• Extend stray_text.unicode_ranges with additional script ranges (Hiragana [0x3040, 0x309f], Katakana [0x30a0, 0x30ff], Hangul [0xac00, 0xd7af]).

6. Interface

CodeChecker(backend: str, dsl: str, config: Optional[dict] = None)

The config parameter is retained for call-site signature compatibility; CodeChecker does not read it.

await checker.check(code: str) -> Tuple[bool, str, List[Dict]]

• passed — overall result.
• error_message — markdown-formatted report with context lines, suitable for insertion into the next coder / KernelGen prompt.
• errors — list of dicts. Each entry has line, error_type, detail, suggestion, code_snippet. The error_type values (syntax_error, compile_error, import_error, stray_chinese_text, no_triton_kernel, triton_kernel_not_called, torch_api_instead_of_kernel, torch_api_without_kernel, autotune_missing_restore_value, empty_code) are referenced by op/autoresearch/agent/tools.py and the failure extractor; rename only with the consumers updated.