Compare commits

...

3 commits

Author SHA1 Message Date
danielhanchen
7c7357667e Merge main into fix-mlx-gguf-reload-context
# Conflicts:
#	tests/studio/run_real_mlx_smoke.py
2026-07-07 04:06:08 +00:00
danielhanchen
052f2b6f97 MLX CI: decode llama-cli timeout output before logging
subprocess.TimeoutExpired carries stdout/stderr as raw bytes even under text=True, so
printing them directly rendered a b'...' repr with escaped newlines. Decode (replacing
undecodable bytes) so the timeout diagnostics are readable in the CI log.
2026-07-01 08:45:38 +00:00
danielhanchen
5998279f17 MLX CI: bound llama-cli context so the GGUF reload stops timing out
With the binary finally located and run, the reload timed out after 300s on a 270m
model. The invocation passed no -c, so llama-cli allocated gemma-3's large default
n_ctx; that KV cache plus the Metal graph for a 262144-token vocab blows past the
runner's memory guard and never returns. The sibling llama-server smoke step already
runs this same model family with -c 256 and passes, so bound the context the same way.
Also surface llama-cli's partial stdout/stderr on TimeoutExpired so any future hang is
diagnosable from the CI log rather than opaque.
2026-07-01 08:17:22 +00:00

Diff content is not available