[Feature]: Windows Support

[astrelsky]

Suggestion Description

After a very quick look through, it doesn't appear to be too bad and there shouldn't be much blocking this. Here are the observed issues.

1. The bash scripts are obviously no good.
2. This is not a standard rocm tool and it is not available on Windows.
   

   FlyDSL/python/flydsl/runtime/device.py

   Lines 10 to 14 in 0c1805f

   	def _arch_from_rocm_agent_enumerator() -> Optional[str]:
   	"""Query rocm_agent_enumerator (standard ROCm tool) for the first GPU arch."""
   	try:
   	out = subprocess.check_output(
   	["rocm_agent_enumerator", "-name"],

3. These need to handle dlls.

   FlyDSL/python/flydsl/compiler/backends/rocm.py

   Lines 72 to 86 in 0c1805f

   	def native_lib_patterns(self) -> List[str]:
   	return [
   	"_fly*.so",
   	"_fly_rocdl*.so",
   	"libFly*.so",
   	"libfly_jit_runtime.so",
   	"libmlir_rocm_runtime.so",
   	"_mlirRegisterEverything*.so",
   	]
   	def jit_runtime_lib_basenames(self) -> List[str]:
   	return [
   	"libfly_jit_runtime.so",
   	"libmlir_c_runner_utils.so",
   	]

It is difficult to spot other issues without attempting to build it and seeing what happens. I don't current have mlir python bindings on hand so I'm going to hold off for now. I might give it a shot this weekend and if I do I'll report any other problems I encounter.

Seeing that the first commit to this project was in November 2025 (native windows rocm support launched in September 2025), it would have been cool if FlyDSL has Windows support to begin with, it would have saved some extra work. Especially if aiter is going to use it and then flash-attention is going to use aiter.

Operating System

Win11

GPU

No response

ROCm Component

FlyDSL

[0xDELUXA]

Would be awesome to have Windows support, and RDNA support too: https://github.com/ROCm/FlyDSL/blob/main/docs/architecture_guide.md#6-target-hardware

[astrelsky]

Would be awesome to have Windows support, and RDNA support too: https://github.com/ROCm/FlyDSL/blob/main/docs/architecture_guide.md#6-target-hardware

FlyDSL/python/flydsl/runtime/device.py

Lines 46 to 53 in f63022e

	def is_rdna_arch(arch: Optional[str] = None) -> bool:
	"""Check if architecture is RDNA-based (gfx10/11/12, wave32).
	This is the single source of truth for CDNA vs RDNA classification.
	RDNA architectures use wave32 and have different buffer descriptor flags.
	If arch is None, the current GPU arch is auto-detected.
	"""

[0xDELUXA]

Ah, okay. Then shouldn’t the documentation at https://github.com/ROCm/FlyDSL/blob/main/docs/architecture_guide.md#6-target-hardware be updated to reflect the RDNA implementation?

[coderfeli]

Ah, okay. Then shouldn’t the documentation at https://github.com/ROCm/FlyDSL/blob/main/docs/architecture_guide.md#6-target-hardware be updated to reflect the RDNA implementation?

Thanks. will update the doc.

[coderfeli]

I haven't accessed the feasibility. Not sure how much changes needed.

[0xDELUXA]

Especially if aiter is going to use it and then flash-attention is going to use aiter.

Flash Attention uses aiter on Windows if we build it now after our commits to both Dao-AILab/flash-attention and ROCm/aiter, but without FlyDSL.