[gspark-etri]

Background

Torch PD serving always enabled producer prefix reuse and used chunk2048, independently of the decoder's cache/chunk settings. This prevented an uncached, matching-chunk comparison from the CLI. BF16 GPU PD showed concurrent differences on Qwen3-0.6B, KONI and Gemma; precision/cache/chunk controls under #240 isolate the tested conditions.

Changes

Expose --pd-prefill-no-prefix-cache and --pd-prefill-chunk for the in-process Torch producer. Preserve cache-on/chunk2048 defaults and existing programmatic callers. Reject nonpositive chunks before loading the model. Document the independent producer/decoder policy.

Verification

• Tiny real-model factory regressions: cache opt-out initially failed1/3, chunk control initially failed1/4; final focused set20passed.
• Full project suite:464passed,15skipped,135subtests passed; one existing flex_attention warning (142.43s).
• Two A10080GB PCIe, Torch2.13.0+cu130/Transformers5.15.0, BF16: all four models pass actual GPU0-prefill/GPU1-decode HTTP with the new cache-off/chunk256 flags. Inputs128/2048/8192,32outputs ignoringEOS,decode slots4,concurrency4/8. Concurrent equality48/48; native/PD serial equality12/12;84remote prefills,0producer cache hits; cancel/disconnect/reuse and invalid-request checks pass.
• Runtime: core369c8ff plus this factory patch; pinned HTTP9e442cd; Validate four-model live GPU PD HTTP over long inputs and lifecycle events #240 validator forwards the new flags. Raw selected JSON and exact setup are in opt_prime/examples/pd/http_matrix/evidence241 and the linked experiment record. No KV transport changes.
• Ruff for the new tests and diff checks pass. Scoped sensitive scan clean; repository-wide scan retains four unchanged baseline findings.

Compatibility and limits

Stacked on #239. Explicit policy controls only; no default numerical or speedup promise. Disabling reuse and smaller chunks may cost performance. Vendor/subprocess producer configuration is unchanged. Original mismatches are retained under #240; broader numerical quality, performance and RNGD/GPU overlap remain separate work.

Related Issue

Closes #241; supports #240.