[gspark-etri]

Background

Fixed-length repeated-capital fixtures do not establish normal chat stopping or answer quality. The four-model goal needs diverse prompts, shared exact input IDs and explicit separation of token parity from factual correctness.

Changes

• Add shared four-question/three-context chat fixtures, GPU/ATOM direct runners and RNGD producer/handoff checks with actual EOS stopping.
• Union generation-config and tokenizer EOS: KONI's config specifies128001 but its tokenizer uses128009 as turn end. The original policy continued into synthetic dialogue; retain that failed-policy evidence.
• Serialize common IDs and compare input hashes; preserve every answer, stop reason, timing and repeat.
• All three RNGD producer families accept saved cases. Gemma right-pads to a compiled bucket but exports only real prompt KV and uses hidden state n-1; CPU full-model oracle covers this padding property. GPU and ATOM handoff checkers honor exported EOS IDs.
• Update coverage/tracking with current KONI/Gemma long-prefix evidence and remaining requirements.

Verification

Base b131963, native evidence e16c256, RNGD handoff implementation5e19f0e, medium-context evidence c645fff. A10080GB BF16 SDPA; exclusive ATOM B1/context9216, Rebel0.11/Optimum0.11.0.post1. Gemma global+sliding corrected artifact; Qwen k3, KONI k1. Checkpoint SHAs and commands in the experiment record.

-192 completed runs:4models ×2backends ×12cases ×2repeats. All terminate at EOS;96 repeat pairs exact.

• GPU/ATOM exact cases: Qwen3-0.6B12/12, Qwen3-4B12/12, KONI8B12/12, Gemma3-1B11/12. Input lengths23–6428 tokens, depending on tokenizer.
• Shared wrong answers are explicitly retained: Qwen4B short17+25→52; Qwen0.6B long Korean-capital→Incheon; Gemma middle-context17+25→40. Longest Gemma arithmetic differs: GPU30, ATOM25. Expected arithmetic answer42. Token parity is not a quality pass.
• RNGD short chat handoff:4models ×4cases ×2decoder backends =32/32 native/imported EOS continuations match. All48prefill repetitions finite/exact. Exact fixture IDs and true KV lengths verified;52/56/72/64slots read back. Qwen/KONI original/portable importer logits exact, Gemma native-state restoration exact. Inputs23–143tokens, buckets128/512; Additional1526–1643token cases (width2048) also pass32/32 decoder comparisons and48finite/exact prefill runs, with exact fixture IDs/true KV lengths and all slot checks. Approximately6K semantic RNGD cases remain open.
• Selected CPU equation/cache/export/EOS suite20passed. Four EOS regression tests pass, including no extra decode after first/later EOS and differing config/tokenizer EOS. Scoped Ruff and diff checks pass.
• Changed-file sensitive scan clean; full tree retains7 unchanged baseline findings outside this work.
• Initial chat-template BatchEncoding serialization failure corrected with explicit return_dict=False. All hardware jobs terminal and owned ATOM allocation released.

Compatibility and limits

Research runner only: production HTTP stopping was not changed. Direct native GPU/ATOM graphs and short and medium-context RNGD full-prefix handoff, static B1; no production PD serving measurement or genuine producer/transfer overlap. Timings are observations, not controlled end-to-end performance claims. Four questions and repetitive context are diagnostic fixtures, not a general quality benchmark. Broader numerical analysis, long RNGD semantic inputs, production cache/cancellation and compute/transfer overlap remain open.

Related Issue

Related #227 and #201. Stacked on #226 (exp/225-gemma-global-causal).

Generated with Codex; maintainer review and merge required.

Long semantic results (14d9772; all four models)

KONI and Qwen3-0.6B pass16/16 GPU/ATOM comparisons at approximately6K input tokens. Gemma at6423–6428 tokens passes3/4 GPU and4/4 ATOM comparisons, Qwen3-4B also passes4/4 on each decoder, for31/32 total across all four models. The original queue completed with48producer runs and exited0. All tested52/56/64/72KV slot readbacks are exact.

Preserved failure: long arithmetic yields30 on nativeGPU BF16/FP32 but25 with RNGD prefill on either decoder; nativeATOM also25. Both are wrong. The same Gemma layer equations executed on GPU choose3 as first token at BF16/FP32, while RNGD chooses2. Intermediate KV differences grow by layer. Cause is unresolved and tracked in #233; no transport-correctness or quality-pass inference is made from native/imported agreement. Reusable check_gemma_equations.py, exact commands, output JSON and the failed comparison are included.

Full-prefix host staging only; live production PD, overlap and sustained serving performance remain outside the demonstrated results.