feat: load models and LoRA adapters from a file descriptor

[ykhrustalev]

Overview

You can now load a model, or a LoRA adapter, from an already-open file descriptor
instead of a file path:

• llama_model_load_from_fd(fd, offset, length, params)
• llama_adapter_lora_init_from_fd(model, fd, offset, length)

The reason is Android. An app ships the GGUF inside its APK (stored with
android:noCompress), and there is no real file to open - all you get is a
descriptor for the APK plus the byte range where the model sits. Copying the
model out to disk first would double the storage use, so we read it in place.

Both functions sit on top of a new ggml call, gguf_init_from_fd(), which is a
small pread reader on top of the existing gguf_init_from_callback(). Two
things follow from that:

• The descriptor is never closed, duplicated, or seeked. The caller keeps full
  control of it, and the same descriptor can serve more than one load at once.
• Offsets are counted from the start of the GGUF, not from the start of the file.
  So alignment and bounds checks work the same no matter where the GGUF was
  placed inside the container.

Limits on this path: no mmap, no direct IO, no split models, and it returns
NULL on Windows.

One unrelated fix rides along: GGML_ABORT messages now also go through
GGML_LOG_ERROR, so an app sees them in its own log callback (logcat on Android)
instead of only on stderr.

Additional information

This is a port of an internal change first written against da426cb25
(2026-02-24). Upstream has changed a lot since then - it gained the
callback-based GGUF reader and the FILE * source in the model loader - so most
of the original patch was no longer needed and was dropped. What is left is about
290 lines instead of the original 900.

How it was tested:

• tests/test-fd.cpp checks gguf_init_from_fd at offsets 0, 1, 7, 31, 32 and
  4096, rejects a length that cuts into the header or the tensor data, and
  confirms the descriptor position is untouched.
• A second entry, test-fd-model, reuses the model that the test suite already
  downloads. It loads the same model both ways and compares the full logits, then
  does the same for a LoRA adapter. To be sure this test can actually fail, I
  removed the container offset from the read path and confirmed both cases break.
• Ran on macOS arm64 and on a Galaxy S24 Ultra (arm64-v8a, NDK 29). 10/10 both.
• Not built for Windows. The _WIN32 branches were only read, not compiled.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. The feature and its first implementation are mine,
  from our internal fork. An assistant did the port to current master, including
  the redesign onto gguf_init_from_callback() with pread, and wrote the tests.

[CISC]

This has been discussed before, but pretty much boils down to #20402 (review)

[JohannesGaessler]

My opinion is that we should not extend the APIs with explicit file descriptor support.

[ykhrustalev]

replaced with #28993