{ "experience": { "tested_on": "2026-10-13 Asia/Shanghai", "strata": "0.1.43", "strata_commit": "7695859aa13afb248bdb4d36f9b2e95707db079d", "engine_sha256_verified_current": "6cc0dda09c0327810d590c4e15d610e8a7a8f3ab05fcfc7376da77b18054e42e", "model": "qwen3.8-flash-next-iq2_xs", "config": { "context": 32768, "prefill": 256, "parallel": 1, "kv": "int8", "vision": false, "reasoning_effort": "none", "temperature": 0, "moe_instrumentation": false }, "requests": [ { "session": "chinese-dialogue", "turn": 1, "http_status": 200, "sse_done": true, "finish_reason": "stop", "ttft_content_seconds": 15.966090703004738, "e2e_seconds": 46.25897292699665, "input_tokens": 126, "output_tokens": 178, "cached_tokens": 0, "engine_decode_tps": 5.9, "error": null }, { "session": "chinese-dialogue", "turn": 2, "http_status": 200, "sse_done": true, "finish_reason": "stop", "ttft_content_seconds": 7.24351614099578, "e2e_seconds": 26.913201265997486, "input_tokens": 366, "output_tokens": 90, "cached_tokens": 304, "engine_decode_tps": 4.6, "error": null }, { "session": "chinese-dialogue", "turn": 3, "http_status": 200, "sse_done": true, "finish_reason": "stop", "ttft_content_seconds": 6.6795737909997115, "e2e_seconds": 17.38428501599992, "input_tokens": 535, "output_tokens": 84, "cached_tokens": 457, "engine_decode_tps": 7.8, "error": null }, { "session": "chinese-dialogue", "turn": 4, "http_status": 200, "sse_done": true, "finish_reason": "stop", "ttft_content_seconds": 9.724326389987255, "e2e_seconds": 23.52638425299665, "input_tokens": 713, "output_tokens": 121, "cached_tokens": 619, "engine_decode_tps": 8.8, "error": null }, { "session": "pelican", "turn": 1, "http_status": 200, "sse_done": true, "finish_reason": "stop", "ttft_content_seconds": 2.1709411649935646, "e2e_seconds": 69.80984544899547, "input_tokens": 22, "output_tokens": 1881, "cached_tokens": 0, "engine_decode_tps": 27.8, "error": null }, { "session": "pelican", "turn": 2, "http_status": 200, "sse_done": true, "finish_reason": "stop", "ttft_content_seconds": 10.708527315000538, "e2e_seconds": 30.618319273999077, "input_tokens": 2002, "output_tokens": 1881, "cached_tokens": 1904, "engine_decode_tps": 94.4, "error": null }, { "session": "pelican", "turn": 3, "http_status": 200, "sse_done": true, "finish_reason": "stop", "ttft_content_seconds": 13.127202471005148, "e2e_seconds": 31.064240987005178, "input_tokens": 4039, "output_tokens": 1881, "cached_tokens": 3882, "engine_decode_tps": 104.7, "error": null }, { "session": "pelican", "turn": 4, "http_status": 200, "sse_done": true, "finish_reason": "stop", "ttft_content_seconds": 3.646410918998299, "e2e_seconds": 4.31951247600955, "input_tokens": 5955, "output_tokens": 4, "cached_tokens": 5920, "engine_decode_tps": 5.7, "error": null }, { "session": "pelican-fresh-edit", "turn": 1, "http_status": 200, "sse_done": true, "finish_reason": "stop", "ttft_content_seconds": 11.591988202999346, "e2e_seconds": 45.92012439398968, "input_tokens": 2014, "output_tokens": 2049, "cached_tokens": 0, "engine_decode_tps": 59.7, "error": null } ], "completion_check": true, "fresh_edit_geometry_checks": { "wheel_circles_unchanged": true, "left_foot_on_lower_crank_endpoint": true, "right_foot_retained": true, "left_leg_reaches_foot": true, "wing_reaches_handlebar_coordinate": true, "dusk_gradient_present": true, "sun_lowered": true }, "quality_scope": "9 requests, each task once; not reliability, general quality, quantization, hardware or performance benchmark" }, "prefill_retest": { "captured_at_utc": "2026-10-13T02:21:08.467928+00:00", "strata_version": "0.1.43", "configuration": { "gpu": "RTX 3090 24GB", "ram": "31.3 GiB", "cpu": "Xeon Gold 5318Y, 16 guest vCPU", "compile_cuda": "12.8.93", "context_tokens": 32768, "quantization": "IQ2_XS", "parallel": 1, "reasoning_effort": "none", "temperature": 0, "max_tokens": 64, "prefill": 256, "PLE_enabled": true, "source_changed": false }, "baseline": { "prefill": "auto, effective maximum 8192", "4k_failures": 3, "4k_runs": 3, "16k_failures": 3, "16k_runs": 3, "failure": "watchdog abort after a step failed to make progress; exit -6" }, "retest_requests": [ { "id": "4k_repro_0", "prompt_tokens": 3939, "completion_tokens": 64, "kv_reused_tokens": 0, "ttft_seconds": 252.73470411200105, "e2e_seconds": 267.7908317620022, "prefill_tokens_per_second": 15.6, "decode_tokens_per_second": 4.2, "http_status": 200, "sse_done": true, "finish_reason": "length", "error": null }, { "id": "4k_repro_1", "prompt_tokens": 3939, "completion_tokens": 64, "kv_reused_tokens": 0, "ttft_seconds": 6.263960215001134, "e2e_seconds": 15.20159733699984, "prefill_tokens_per_second": 632.2, "decode_tokens_per_second": 7.2, "http_status": 200, "sse_done": true, "finish_reason": "length", "error": null }, { "id": "16k_repro_0", "prompt_tokens": 15716, "completion_tokens": 64, "kv_reused_tokens": 0, "ttft_seconds": 225.86787449899566, "e2e_seconds": 235.10131674599688, "prefill_tokens_per_second": 69.6, "decode_tokens_per_second": 6.9, "http_status": 200, "sse_done": true, "finish_reason": "length", "error": null } ], "resource_change": { "resident_expert_ram_gib_before": 18.79, "resident_expert_ram_gib_after": 15.28, "prompt_borrowed_gpu_slots_before": 2812, "prompt_borrowed_gpu_slots_after": 199 }, "io_evidence": "results/prefill256-16k-io-summary.json", "io_assessment": { "observed": "Blocked thread stack filemap_fault -> wait_on_page_bit_common, ongoing major faults and disk reads, no per-request expert blob reads from files", "supported_inference": "Storage-backed table/page fetches are a major latency bottleneck; the PLE n-gram table is the main source-level suspect", "not_proven": [ "Physical backend storage type or host QoS", "Exclusive attribution of all cold/warm improvement to disk I/O", "Production stability, long-output performance, concurrent capacity" ] }, "test_scope": [ "Two approximately 4K inputs and one approximately 16K input, official synthetic word-list pattern", "Second 4K body largely overlaps the first, nonce differs; the 16K input has overlapping corpus too", "No manual OS cache flush; model-loading time excluded", "All responses end at the requested 64-token cap; semantic completeness is not claimed", "No KV reuse in these requests" ], "health_after_loaded": true, "task": "Synthetic English word list; request a five-sentence classification; 64-token visible-output cap; no thinking output" }, "prefill_io_window": { "interval_seconds": 30.163772276006057, "process_read_bytes": 359022592, "process_major_faults": 3038, "process_cpu_seconds": 19.22, "vda_completed_reads": 3103, "vda_read_iops": 102.87174865287983, "vda_mean_read_ms": 9.864969384466646, "vda_read_ms_delta": 30611, "vda_read_sectors_delta": 701600, "attribution_scope": "Guest virtual block device statistics, not a measurement of the physical backend; no synthetic disk load.", "field_definition_source": "https://docs.kernel.org/admin-guide/iostats.html" }, "artifact_note": "Different test batches are separate. Timeline is a separate instrumented run; experience and prefill timings use standard engine.", "disk_space": { "captured_date": "2026-10-13", "method": "du -B1 -s, allocated filesystem bytes; each category counted once", "allocated_bytes": { "gguf_and_metadata": 68026126336, "mtp": 6927724544, "expert_packs": 37002997760, "model_data_total": 111956856832, "entire_test_installation_with_logs": 112917946368 }, "unit": "GB uses 1e9 bytes", "suggested_free_bytes": 130000000000, "budget_scope": "recommendation for this IQ2_XS installation with headroom; not an official minimum" }, "prefill_answer_examples": "prefill-retest-answers.json" }