feat(backends/python): use tempfile.gettempdir() instead of hardcoded /tmp

[Anai-Guo]

Closes #9601

Problem

The Python backends vllm, vllm-omni, tinygrad, and pocket-tts hardcode /tmp as the directory for scratch files (base64-decoded video/audio buffers, generated image/TTS outputs when no destination is provided). On systems where / is a small or read-only partition (containers with size-limited overlays, locked-down hosts, automation environments where the user can manage data volume but not root), there is no escape hatch: even with LOCALAI_GENERATED_CONTENT_PATH and LOCALAI_UPLOAD_PATH already configurable for the Go side, these Python backends still go to /tmp.

Reporter on #9601:

default /tmp is on the root partition which has limit size; the size limit is because of automation and manageability ... tried out a few parameters to no avail.

Fix

Replace the five hardcoded "/tmp/..." literals with os.path.join(tempfile.gettempdir(), ...). tempfile.gettempdir() honors the standard TMPDIR env var (and TMP/TEMP on Windows), so users can simply set e.g. TMPDIR=/data/tmp and these scratch files relocate accordingly. Behavior is unchanged when TMPDIR is unset (Linux default is still /tmp).

This is a deliberately minimal change — it lets the existing tempfile standard-library mechanism solve the issue without inventing a new LocalAI-specific config knob. It composes cleanly with the existing LOCALAI_GENERATED_CONTENT_PATH / LOCALAI_UPLOAD_PATH options (which target different code paths on the Go side).

Sites changed

File	Line (orig)	Before	After
backend/python/vllm/backend.py	605	f"/tmp/vl-{timestamp}.data"	os.path.join(tempfile.gettempdir(), f"vl-{timestamp}.data")
backend/python/tinygrad/backend.py	671	request.dst or "/tmp/tinygrad_image.png"	request.dst or os.path.join(tempfile.gettempdir(), "tinygrad_image.png")
backend/python/pocket-tts/backend.py	207	output_path = "/tmp/pocket-tts-output.wav"	output_path = os.path.join(tempfile.gettempdir(), "pocket-tts-output.wav")
backend/python/vllm-omni/backend.py	121	f"/tmp/vl-{timestamp}.data"	os.path.join(tempfile.gettempdir(), f"vl-{timestamp}.data")
backend/python/vllm-omni/backend.py	141	f"/tmp/audio-{timestamp}.wav"	os.path.join(tempfile.gettempdir(), f"audio-{timestamp}.wav")

Plus an import tempfile added to each file (next to the existing stdlib imports). All 4 modified files parse cleanly with python -m py_compile.

Why not switch to NamedTemporaryFile?

The existing code uses time.time() * 1000-based filenames and explicit os.remove(p) cleanup; switching to NamedTemporaryFile would be a larger behavior change (collision-safe naming, RAII cleanup) that's worth doing on its own merit but is out of scope for this /tmp-relocation fix.

Test plan

• python -m py_compile on all 4 modified files passes
• grep -n '"/tmp\|f"/tmp' backend/python/{vllm,tinygrad,pocket-tts,vllm-omni}/backend.py returns no hits after the change
• Behavior (manual): with TMPDIR=/data/tmp set, vllm video base64 fallback writes /data/tmp/vl-<ts>.data instead of /tmp/vl-<ts>.data. With TMPDIR unset, behavior is identical to before on Linux (/tmp/...).

🤖 Generated with Claude Code