[spectrometerHBH]

CUDA compiler settings need to be consistent across build defaults, target presets, and individual device entries without maintaining a Python field for every native compiler flag. Introduce a lightweight per-backend mapping accepted directly by tvm.compile(..., backend_config=...) and T.device_entry(..., backend_config=...).

from tvm.backend.cuda import BackendConfig

cuda_config = BackendConfig(
    arch="sm_100a",
    compiler="nvrtc",
    nvrtc=["--use_fast_math", "--ftz=false"],
)
executable = tvm.compile(func, backend_config={"cuda": cuda_config})

• CUDA exposes six optional keys: arch, compiler, target_format, nvcc, nvrtc, and ptxas. Native argv lists are forwarded to the selected toolchain. BackendConfig is a TypedDict whose constructor returns a regular dictionary.
• Resolve backend defaults, Target/tag defaults, compile overrides, and entry overrides before architecture-sensitive lowering. Lists replace inherited lists; an empty list clears them. Snapshot inputs and use canonical configuration cache keys.
• Compile entries with different architectures or options in separate existing CUDA modules. Specialize shared private device helpers before tile dispatch and dtype legalization so each helper follows its caller's configuration.
• Keep CUDA runtime metadata to one opaque backend configuration string appended after the existing serialized module prefix. Remove the versioned metadata helper and persisted compiler logs. Accept legacy payloads, avoid invoking a compiler for source-only fallback modules, and replay saved source independently of an unrelated current Target.
• Export cuda_host as ordinary C++ with embedded independently compiled device binaries. Exported libraries require TVM FFI and CUDA, with no device compiler required at runtime.
• Replace CompileConfig, compiler-policy environment variables, and decorator-level architecture settings. Obsolete compiler-policy environment variables are ignored. Update build, Relax, external-kernel, and IKET consumers and document native compiler options separately from the runtime schema.

The caller migration is in mlc-ai/TIRx-kernels#8. The earlier CUDA launch configuration PR #20639 is merged.

Validation on the current implementation:

• TVM C++ build and changed-file pre-commit checks passed.
• Full TIRx suite: 2,691 passed / 290 skipped on seven B200 GPUs (160 seconds). Workload import gate (274 workloads / 93 kernels) and strict registry import gate passed.
• CUDA codegen and Relax VM build regressions: 415 passed / 4 skipped.
• Backend configuration tests: 45 passed, covering native flags and FTZ behavior, Target/tag defaults, mixed NVCC/NVRTC entries, helper specialization, concurrent compilation, explicit offline architecture, source fallback, legacy payloads, fresh-process artifact replay, and standalone C++ host export.
• TIRx-kernels unit tests: 225 passed, including configuration forwarding, CPU preparation without CUDA initialization, mutable-input snapshots, and cache keys across Target defaults.

[tqchen]

source remember the code and backend_config as string

[tqchen]

ffi::String backend_config_; // serialized json string of backend config;