[arif-ahmed-nv]

Summary

Stock cosmos-framework cannot be imported or started on Windows: several modules assume POSIX primitives at module scope or at trainer construction. This PR guards each assumption so that Linux behaviour is unchanged while a single-GPU native Windows run (x64 and Windows on ARM64 / RTX Spark) can import, train, evaluate and export.

All platform shims live in one new module, cosmos_framework/utils/win_compat.py, so the port is reviewable in one place. On Linux every helper there delegates to the stdlib primitive it replaces.

Fixes, one per POSIX assumption

Assumption	Where	Fix
os.sched_getaffinity / sched_setaffinity	utils/distributed.py	hasattr guard in _set_gpu_cpu_affinity; AttributeError added to the handler. On Windows the AttributeError is not an OSError, so it escaped the handler and aborted every entry point.
signal.SIGALRM / signal.alarm	trainer/__init__.py	Guarded; the step-timeout watchdog is disabled where alarm(2) does not exist instead of failing trainer init.
NCCL backend	utils/distributed.py	Fall back to gloo on win32 (is_nccl_available() is False). A process group is still mandatory: Trainer.__init__ runs an unconditional dist.barrier(). env:// defaults are set for the single-rank case, which has no torchrun launcher on native Windows.
gloo collectives on CUDA tensors	utils/distributed.py, win_compat.py	The Windows gloo build faults on CUDA tensors; collectives are staged through host memory.
DDP reducer on CUDA grads	utils/distributed.py, trainer/__init__.py	DDP wrapper skipped at world_size == 1 on win32 (access violation in backward otherwise). The trainer unwraps .module only when the wrapper was applied.
import fcntl (3 modules)	lpips_cache.py, scripts/dataset_utils.py, utils/reasoner/distributed.py	win_compat.flock (passthrough on Linux, msvcrt byte-range lock on Windows). In reasoner/distributed.py fcntl backs SIOCGIFADDR/SIOCGIFCONF interface discovery (multi-node only); imported defensively, helpers return "no interfaces" when absent.
os.symlink for the rank-0 job alias	scripts/_train.py	Falls back to a directory junction, then a pointer file (symlinks need Developer Mode or SeCreateSymbolicLinkPrivilege).
multistorageclient at module scope	utils/easy_io/transient_retry.py	Optional import (no Windows wheel or sdist; its xattr dependency needs POSIX xattrs). The backends package already guards it.
import triton at module scope	model/generator/moe/native_kernels.py	Optional import with a loud failure on kernel launch. Reached transitively by the config graph even for dense, non-MoE models; Triton has no official Windows wheel.

Also included, cross-platform: CheckpointDirHf drops a trivial --include * because the hf CLI parses a bare * as a positional filename, fetches only README.md, and still exits 0, which fails much later in the tokenizer.

Validation

• End to end (Cosmos3 Nano LoRA SFT, evaluate, export) on Windows 11 x64 (RTX PRO 6000) and Windows 11 on ARM64 (RTX Spark), against the PR Add TAO-compatible Cosmos3 fine-tuning support #164 base.
• This rebase onto main (after Add reproducible Cosmos3 Nano and Edge fine-tuning #277) compiles and passes ruff format --check; the same end-to-end run on main is still pending, which is why this is a draft.
• Not covered: multi-node on Windows, MoE backbones on Windows, Windows CI.

Notes for reviewers

• sys.platform == "win32" guards keep every behavioural change off the Linux path.
• The pre-existing I001 import-order findings in touched files are unchanged by this PR.
• The two remaining top-level Triton imports (mot/maskless_kernels.py, tokenizers/dc_ae) were not on the dense Nano path during validation and are left as is.

🤖 Generated with Claude Code