fix: DML/GPU build crash — skip compile config for EPs with enable_ep_context=False

[DingmaomaoBJTU]

Summary

Fixes #396

winml build crashes with FileNotFoundError when targeting GPU/DML because the compile stage produces no EPContext file, but the pipeline unconditionally advances current_path to the non-existent compiled_path.

Root Cause

Why DML hits this but CPU doesn't:

CPU is protected at config generation time: _DEVICE_TO_PROVIDER["cpu"] = None → WinMLCompileConfig.for_provider(None) returns None → config.compile = None → compile stage is skipped entirely by the first guard in _run_compile_stage.

DML goes through a different path: _DEVICE_TO_PROVIDER["gpu"] = "dml" → WinMLCompileConfig.for_dml() returns a config with enable_ep_context=False. This produced a non-None compile config, so the stage was entered. But enable_ep_context=False means no EPContext file is ever written. The old code then unconditionally set current_path = compiled_path on a file that doesn't exist.

Fix

Two changes:

1. Fix at config level — WinMLCompileConfig.for_provider() returns None for non-EPContext EPs

Rather than checking enable_ep_context inside build.py, the fix teaches for_provider() that EPs with enable_ep_context=False have no offline compile step and should return None:

factory = factories.get(provider)
if factory:
    config = factory()
    # EPs that don't produce EPContext have no offline compile step
    if not config.ep_config.enable_ep_context:
        return None
    return config

This naturally covers DML, CPU, CUDA, NvTensorRTRTX, VitisAI, and MIGraphX without hardcoding EP names. The logic is derived from the enable_ep_context flag already set in each factory. Unknown/custom EPs (generic fallback) are not subject to this rule.

2. Raise explicitly when EPContext-producing EP reports success but output is missing

Replaces the silent compiled_path.exists() fallback with an explicit RuntimeError, so real failures (e.g. QNN) surface immediately:

if not compiled_path.exists():
    raise RuntimeError(
        f"Compile reported success but output not found: {compiled_path}"
    )

3. --no-compile defaults to True

Changed from is_flag=True, default=False to --no-compile/--compile pair with default=True, so compile is disabled by default in the CLI.

[xieofxie]

but I do think DML also supports compile, bug?

[DingmaomaoBJTU]

Good catch - DML does support compile, but in our current flow it is JIT-style (usually no persisted EPContext file). See: https://github.com/microsoft/WinML-ModelKit/blob/77d553b75846178e94e0e1dca0d5d0ec623cfc7e/src/winml/modelkit/compiler/configs.py#L177 So the bug fixed here is the path assumption: _run_compile_stage switched to compiled_path even when no file was produced. Now it only switches when the file actually exists.