Upload folder using huggingface_hub

docs/features/llamafile_build.md

@@ -109,7 +109,7 @@ you're polishing a patch you haven't yet captured under
 ## `package` subcommand details
 
 ```
-build:llamafile:package [--with-model PATH] [--with-args PATH] [--arch ID] [--cpu-only]
+build:llamafile:package [--with-model PATH] [--with-args PATH] [--arch ID] [--cpu-only] [--win-variant]
 ```
 
 - `--with-model PATH` — GGUF to embed. Filename is preserved; the
@@ -124,6 +124,12 @@ build:llamafile:package [--with-model PATH] [--with-args PATH] [--arch ID] [--cp
 - `--cpu-only` — skip embedding any DSO. Produces a binary that
   works on any host but always runs on CPU. Useful for CI smoke
   on machines with no GPU.
+- `--win-variant` — package the bare thin APE (no embedded DSO) as
+  `…-win-x86_64.llamafile.exe` and hard-fail if it exceeds 4 GiB:
+  Windows refuses executables >4 GB, so this is the only variant
+  that can run there (CPU out of the box; GPU via side-load DSO on
+  Linux or llamafile's on-the-fly nvcc DLL compile on Windows).
+  MANIFEST records `arch: "win-x86_64"`, `backends: []`.
 
 Embedding uses the vendored zipalign at
 `vendors/sources/llamafile/o/third_party/zipalign/zipalign`. For DSOs