Verification logs (patched branch, Arch Linux, 2026-05-31)

Installed from editable checkout: pip install -e /home/jim/Projects/unsloth (branch fix/studio-poll-cli-model-load, includes #5900 + #5901 + #5902).

Command:

unsloth studio run \
  -m unsloth/Qwen3.6-27B-MTP-GGUF \
  --gguf-variant UD-IQ2_XXS \
  --max-seq-length 8192 \
  --no-mmproj \
  --port 8889 --host 127.0.0.1 --silent

#5902 — --no-mmproj honored; no mmproj download on CLI load

Before (repro): Downloading mmproj: unsloth/Qwen3.6-27B-MTP-GGUF/mmproj-F16.gguf

After (patched):

{"timestamp": "2026-05-31T14:14:14.354872Z", "event": "Vision-capable GGUF loaded without a usable mmproj; image input will be disabled for this session"}
{"timestamp": "2026-05-31T14:14:14.355179Z", "event": "Appending user extra args to llama-server: ['--no-mmproj']"}
{"timestamp": "2026-05-31T14:14:14.355224Z", "event": "Starting llama-server: ... --no-mmproj"}

(No Downloading mmproj line. llama-server command ends with --no-mmproj, not --mmproj.)

#5902 — UI reload inherits llama_extra_args

{"timestamp": "2026-05-31T14:15:33.870718Z", "event": "Inheriting llama_extra_args from previous load (same model, shadow-stripped): ['--no-mmproj']"}
{"timestamp": "2026-05-31T14:15:35.282062Z", "event": "Starting llama-server: ... --no-mmproj"}
{"timestamp": "2026-05-31T14:16:01.740749Z", "event": "Loaded GGUF model via llama-server: unsloth/Qwen3.6-27B-MTP-GGUF"}

Reload request: POST /api/inference/load with {"model_path":"unsloth/Qwen3.6-27B-MTP-GGUF","gguf_variant":"UD-IQ2_XXS"} (no llama_extra_args field).

#5900 / #5901 — no Gemma auto-load; Qwen stays active

Before (repro):

{"timestamp": "2026-05-31T13:28:02.197259Z", "event": "Detected remote GGUF repo 'unsloth/gemma-4-E2B-it-GGUF', variant=UD-Q4_K_XL, vision=True"}

After (patched session): no Detected remote GGUF repo 'unsloth/gemma-4-E2B-it-GGUF' inference load. Only background helper pre-cache:

{"timestamp": "2026-05-31T14:14:13.453443Z", "event": "Pre-caching helper GGUF: unsloth/gemma-4-E2B-it-GGUF/gemma-4-E2B-it-UD-Q4_K_XL.gguf"}

Status after CLI load:

{"active_model":"unsloth/Qwen3.6-27B-MTP-GGUF","model_identifier":"unsloth/Qwen3.6-27B-MTP-GGUF","gguf_variant":"UD-IQ2_XXS","is_vision":false}

Chat completion stayed on Qwen:

{"model":"unsloth/Qwen3.6-27B-MTP-GGUF","choices":[{"delta":{"content":"The user is asking me to reply with exactly \"OK\"..."}}]}

POST /api/inference/load count in session: 1 CLI load + 1 intentional same-model reload test — no Gemma load.

Unit tests

pytest studio/backend/tests/test_llama_server_args.py — 165 passed

Pushed 6d56af4 to address the two P1/P2 Codex concerns:

• P1: waitForModelReady() is no longer entered when checkpointEmpty is true but no UI load is in progress. Normal empty-session starts go straight to autoLoadSmallestModel() (which calls tryAdoptServerActiveModel() first) instead of spinning for 120 s.
• P2: listModels() / listLoras() are now fetched and committed to the store before the CLI-load poll starts, so the model selector is never blocked for a full minute during an idle session.

@danielhanchen @rolandtannous — would you be able to review when you get a chance? CI is showing action_required on all workflows; maintainer approval to run the workflows would be appreciated.

Merge + Codex follow-up (c5e6e70)

Merged latest main and addressed the remaining Codex threads:

• Merge conflict (chat-adapter.ts): kept autoLoadSmallestModel(abortSignal) so abort/cleanup still works.
• P1 — wait before mmap / HF download (chat-adapter.ts + inference.py): /api/inference/status now reports in-flight GGUF loads while _serial_load_lock is held; adoptInFlightServerLoad waits on status.loading or any non-null load-progress phase (adaptive poll — idle sessions still return immediately).
• P2 — multimodal reset during poll (use-chat-model-runtime.ts): skip the no-active-model reset branch when userSelectedDuringPoll.
• P2 — adopt race (apply-inference-status-to-store.ts): re-read params.checkpoint after getInferenceStatus() before calling setCheckpoint.

PR is MERGEABLE; waiting on pre-commit.ci.

Pushed b709b8f with two changes:

1. extra_args_disable_mmproj() now matches the version on fix(studio): inherit llama_extra_args and honor --no-mmproj #5902: it recognises the --no-mmproj-auto alias and mirrors llama-server's last-wins parsing for the --mmproj-auto / --no-mmproj / --no-mmproj-auto boolean group, with tests for both. Identical content on both branches so they merge cleanly in either order.
2. Restored the original shorter wording for three comments in use-chat-model-runtime.ts that were rewritten into longer versions with the same meaning, which shrinks the diff.

The polling design here is solid: evidence-gated waiting via /load-progress and status.loading means idle sessions still auto-load with zero delay, and the post-poll checkpoint re-check protects user selections.

Synced with latest main and cleared open review threads.

• Merged origin/main into this branch; PR is mergeable again.
• Resolved remaining Codex/Gemini review threads (including items already addressed in @danielhanchen's follow-up commits).

Waiting on maintainer review/approval. pre-commit.ci is the only automated gate visible from fork PRs; GitHub Actions still require maintainer approval for first-time contributors.

Addressed in 010142e: the capability-clear branch now gates on checkpointAfterPoll instead of the stale pre-await isExternalSelectionActive / poll-only userSelectedDuringPoll, so a local selection during in-flight refresh no longer clears multimodal flags.