[vince62s]

The workspace opens generation settings by default, prefers compatible DFlash2 with INT8 packed drafter projections, and raises the Chat output allowance to 16,384 tokens in local/RunPod templates. Explicit user choices are retained, with MTP/Off fallback when DFlash2 is unavailable. The proposal count continues to follow the checkpoint block size; server startup preparation is unchanged.

Benchmark now compares three fixed workloads (Python code, English story, Chinese-to-English translation) using six independently selectable profiles:

• Base with native vocabulary heads.
• MTP with native or INT4 vocabulary heads.
• DFlash2 with BF16 or INT8 internal drafter projections and native vocabulary heads.
• DFlash2 with INT8 internal projections and INT4 vocabulary heads.

All profiles use text-only greedy decoding with thinking disabled, identical prompts within each workload and a configurable common output cap (default 512) and 1–5 repeats (default 3). Each workload/profile pair receives one discarded warmup; profiles are grouped to avoid repeated reloads between workloads. Native dtype, target/drafter precision and draft counts remain distinct.

The results table retains previous comparisons until page reload and reports medians, first-token latency, acceptance and completion status. JSON export includes exact prompts/settings, hardware, context/packing/compile metadata, separate warmups and raw measured runs. Early stops, incorrect actual speculation modes, cancellation and errors cannot produce a complete fixed-length median. CUDA OOM stops the suite, retaining available results. Model switching/preparation is excluded from decode timing. The final engine remains loaded until another request or unload.

Validation:

• Full CPU suite: 811 tests, OK (43 skipped).
• Browser helper/runner tests: warmup exclusion, fixed requests, retained comparisons, unavailable profiles, early EOS, actual-mode checks, OOM and cancellation.
• Real RTX 5090 browser smoke tests: six complete MTP/INT4 and DFlash2/INT8+INT4 rows across all three workloads at a 128-token cap and one measured repeat; repeated successfully. Report download and its contents verified. No browser console errors.
• Black, flake8, Prettier, JavaScript syntax and whitespace checks passed.
• Model unloaded after testing; preview server remains available on localhost:5018.

Screenshot is a short validation run, not the default 512-token, three-repeat performance comparison.