resolve multi-CUDA_ARCHITECTURES compilation conflicts

[GACLove]

fix(build): resolve multi-CUDA_ARCHITECTURES compilation conflicts

• Isolate build artifacts per CUDA architecture in build/sm_{arch}/ subdirectories
• Add validation for CUDA_ARCHITECTURES format (comma-separated decimals)

[Copilot]

Pull Request Overview

This PR resolves CUDA compilation conflicts when building for multiple GPU architectures by isolating build artifacts and adding validation for CUDA_ARCHITECTURES environment variable.

• Enables building for multiple CUDA architectures by reading from CUDA_ARCHITECTURES environment variable
• Isolates build artifacts per architecture in separate subdirectories to prevent conflicts
• Filters NVCC flags to only include relevant architecture-specific flags for each extension

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[Copilot]

The code doesn't validate that each architecture value is a valid decimal format as mentioned in the PR description. Consider adding validation to ensure each arch value matches the expected pattern (e.g., regex check for decimal format).

Suggested change

	if arch:
	if arch:
	if not re.match(r"^\d+\.\d+$", arch):
	raise ValueError(f"Invalid architecture value '{arch}' in CUDA_ARCHITECTURES. Expected decimal format like '8.0', '8.6', etc.")

[Copilot]

The flag filtering logic is duplicated between sm89 and sm90 extensions. Consider extracting this into a helper function to reduce code duplication and improve maintainability.

Suggested change

	sm89_arch_list = ["sm_89", "compute_89", "sm_90a", "compute_90a", "sm_120", "compute_120"]
	filtered_flags = filter_nvcc_flags(NVCC_FLAGS, sm89_arch_list)

[Copilot]

This is duplicate code from the sm89 extension filtering logic. The same flag filtering pattern should be extracted into a reusable function.

Suggested change

	filtered_flags.append(flag)
	def filter_nvcc_flags_for_arch(nvcc_flags, arch_substrings):
	filtered_flags = []
	skip_next = False
	for i, flag in enumerate(nvcc_flags):
	if skip_next:
	skip_next = False
	continue
	if flag == "-gencode":
	if i + 1 < len(nvcc_flags):
	arch_flag = nvcc_flags[i + 1]
	if any(sub in arch_flag for sub in arch_substrings):
	filtered_flags.append(flag)
	filtered_flags.append(arch_flag)
	skip_next = True
	elif flag not in ["-gencode"]:
	filtered_flags.append(flag)
	return filtered_flags
	filtered_flags = filter_nvcc_flags_for_arch(NVCC_FLAGS, ["sm_90a", "compute_90a"])

[Copilot]

Using 'default' as a fallback directory name could lead to conflicts if multiple threads don't have mapped extension names. Consider using a more unique identifier like thread ID or timestamp.

Suggested change

	self.thread_ext_name_map.get(threading.current_thread().ident, "default")),
	self.thread_ext_name_map.get(
	threading.current_thread().ident,
	f"thread_{threading.current_thread().ident}"
	)),

[GACLove]

testing

pip install -v --no-cache-dir .
CUDA_ARCHITECTURES="9.0,12.0" pip install -v --no-cache-dir .
CUDA_ARCHITECTURES="8.9,9.0" pip install -v --no-cache-dir .