{ "experiment": "strata-smoke", "arm_id": "strata_q2_0_32k", "started_utc": "2026-10-05T05:54:05Z", "model": "Q2_0", "family": "qwen", "context_tokens": 32768, "pass_threshold_tok_s": 100.0, "config": { "exe": "/opt/strata/engine/strata", "args": [ "--pack", "/data/packs/q2_0", "--native", "/data/models/Q2_0/Qwen3.8-Flash-Next-GSQ-RCO-Q2_0-00001-of-00002.gguf", "--ple-gguf", "/data/models/Q2_0/Qwen3.8-Flash-Next-GSQ-RCO-Q2_0-00002-of-00002.gguf", "--expert-profile", "/opt/strata/data/expert-profile.bin", "--expert-cache", "auto", "--prefill", "auto", "--spec", "4", "--spec-min-p", "0.5", "--mtp", "/data/mtp/rt", "--max-context", "32768", "--kv", "int8" ], "cwd": "/opt/strata", "tokenizer": "/data/packs/q2_0/tokenizer", "model_name": "qwen3.8-flash-next-q2_0", "log": "/opt/strata/strata-q2_0.log", "lib_dirs": [ "/usr/local/cuda/bin", "/usr/local/cuda/lib64" ], "port": 8080, "gpu": 0, "gpus_asked": true, "host": "127.0.0.1", "open_browser": false }, "environment": { "memtotal_kb": 65765244, "memavailable_kb": 49840204, "memlock_soft": "unlimited", "memlock_hard": "unlimited", "nproc": 16, "nvidia_smi_rc": 0, "nvidia_smi": "NVIDIA GeForce RTX 3090, 24576 MiB, 306 MiB, 595.91.07" }, "calibrate": { "rc": 0, "tail": " ...\n PCIe share 0.00: 127.6 tok/s\n PCIe share 0.20: 125.9 tok/s\n PCIe share 0.35: 118.6 tok/s\n PCIe share 0.36: 114.5 tok/s\n PCIe share 0.55: 110.8 tok/s\n PCIe share 0.75: 110.3 tok/s\n draft floor 0.30: 115.9 tok/s\n draft floor 0.50: 119.4 tok/s\n draft floor 0.70: 120.0 tok/s\n Measuring with 7 CPU workers (restarts the engine) ...\n[strata] starting the engine: reading the model's weights ...\n[strata] loading the experts into RAM (about 38 GB) and locking part of them for the GPU.\n YOUR PC CAN BE SLOW OR STOP RESPONDING FOR 1-3 MINUTES NOW - this is normal.\n Please wait and don't close this window; the browser opens when it is ready.\n[strata] still starting (23 s) - please wait ...\n[strata] still starting (43 s) - please wait ...\n[strata] still starting (63 s) - please wait ...\n[strata] still starting (83 s) - please wait ...\n[strata] still starting (103 s) - please wait ...\n[strata] still starting (123 s) - please wait ...\n[strata] still starting (143 s) - please wait ...\n[strata] still starting (163 s) - please wait ...\n[strata] still starting (183 s) - please wait ...\n[strata] still starting (203 s) - please wait ...\n[strata] still starting (223 s) - please wait ...\n[strata] still starting (243 s) - please wait ...\n[strata] still starting (263 s) - please wait ...\n[strata] experts loaded: 31.64 GiB at 1.93 GiB/s (267 s so far)\n[strata] filling the GPU's expert cache (13465 experts, 17.34 GiB of VRAM) ...\n[strata] almost ready ...\n 7 workers: 130.1 tok/s\n Measuring with 5 CPU workers (restarts the engine) ...\n[strata] starting the engine: reading the model's weights ...\n[strata] loading the experts into RAM (about 38 GB) and locking part of them for the GPU.\n YOUR PC CAN BE SLOW OR STOP RESPONDING FOR 1-3 MINUTES NOW - this is normal.\n Please wait and don't close this window; the browser opens when it is ready.\n[strata] still starting (24 s) - please wait ...\n[strata] still starting (44 s) - please wait ...\n[strata] still starting (64 s) - please wait ...\n[strata] still starting (84 s) - please wait ...\n[strata] still starting (104 s) - please wait ...\n[strata] still starting (124 s) - please wait ...\n[strata] still starting (144 s) - please wait ...\n[strata] still starting (164 s) - please wait ...\n[strata] still starting (184 s) - please wait ...\n[strata] still starting (204 s) - please wait ...\n[strata] still starting (224 s) - please wait ...\n[strata] still starting (244 s) - please wait ...\n[strata] still starting (264 s) - please wait ...\n[strata] experts loaded: 31.64 GiB at 1.77 GiB/s (266 s so far)\n[strata] filling the GPU's expert cache (13465 experts, 17.34 GiB of VRAM) ...\n[strata] almost ready ...\n 5 workers: 128.9 tok/s\n Measuring with 4 CPU workers (restarts the engine) ...\n[strata] starting the engine: reading the model's weights ...\n[strata] loading the experts into RAM (about 38 GB) and locking part of them for the GPU.\n YOUR PC CAN BE SLOW OR STOP RESPONDING FOR 1-3 MINUTES NOW - this is normal.\n Please wait and don't close this window; the browser opens when it is ready.\n[strata] still starting (23 s) - please wait ...\n[strata] still starting (43 s) - please wait ...\n[strata] still starting (63 s) - please wait ...\n[strata] still starting (83 s) - please wait ...\n[strata] still starting (103 s) - please wait ...\n[strata] still starting (123 s) - please wait ...\n[strata] still starting (143 s) - please wait ...\n[strata] still starting (163 s) - please wait ...\n[strata] still starting (183 s) - please wait ...\n[strata] still starting (203 s) - please wait ...\n[strata] still starting (223 s) - please wait ...\n[strata] still starting (243 s) - please wait ...\n[strata] experts loaded: 31.64 GiB at 2.16 GiB/s (263 s so far)\n[strata] filling the GPU's expert cache (13465 experts, 17.34 GiB of VRAM) ...\n[strata] almost ready ...\n 4 workers: 126.0 tok/s\n [ok] tuned for this PC: --pcie-frac 0.00, --spec-min-p 0.70 (130.1 tok/s)\n", "config_after": { "exe": "/opt/strata/engine/strata", "args": [ "--pack", "/data/packs/q2_0", "--native", "/data/models/Q2_0/Qwen3.8-Flash-Next-GSQ-RCO-Q2_0-00001-of-00002.gguf", "--ple-gguf", "/data/models/Q2_0/Qwen3.8-Flash-Next-GSQ-RCO-Q2_0-00002-of-00002.gguf", "--expert-profile", "/opt/strata/data/expert-profile.bin", "--expert-cache", "auto", "--prefill", "auto", "--spec", "4", "--spec-min-p", "0.5", "--mtp", "/data/mtp/rt", "--max-context", "32768", "--kv", "int8" ], "cwd": "/opt/strata", "tokenizer": "/data/packs/q2_0/tokenizer", "model_name": "qwen3.8-flash-next-q2_0", "log": "/opt/strata/strata-q2_0.log", "lib_dirs": [ "/usr/local/cuda/bin", "/usr/local/cuda/lib64" ], "port": 8080, "gpu": 0, "gpus_asked": true, "host": "127.0.0.1", "open_browser": false } }, "health": { "status": "ok", "max_context": 32768, "model": "qwen3.8-flash-next-q2_0", "images": false, "api_key": false, "loaded": true, "service": "strata" }, "status": { "service": "strata", "model": "qwen3.8-flash-next-q2_0", "loaded": true, "auto_load": true, "structured_output": { "formats": [ "json_object", "json_schema" ], "method": "prompt_and_validate", "constrained_decoding": false, "stream_buffered": true }, "engine": "0.1.39", "started": 1791181047, "uptime_s": 3, "cache_max_tokens": 32768, "context": { "native": 32768, "max_positions": 32768 }, "concurrency": { "serving": 1, "requested": 1 }, "dialects": [ "/v1/chat/completions", "/v1/messages", "/v1/responses" ], "vision": { "enabled": false, "available": false, "error": null }, "activity": { "requests": 0, "in_flight": 0, "last_request_at": null }, "last_timings": null, "vram": { "elastic": false }, "machine": { "at": 1791181051, "gpu": { "name": "NVIDIA GeForce RTX 3090", "used_mib": 24095.0, "total_mib": 24576.0, "util_pct": 0, "temp_c": 38, "power_w": 99.758 }, "ram": { "used_gib": 44.2, "total_gib": 62.7 } } }, "models": { "object": "list", "data": [ { "id": "qwen3.8-flash-next-q2_0", "object": "model", "status": { "value": "loaded" }, "meta": { "n_ctx": 32768 }, "architecture": { "input_modalities": [ "text" ], "output_modalities": [ "text" ] } } ] }, "prompt_fit": { "target_tokens": 30000, "fitted_tokens": 30341, "prompt_chars": 97552, "max_tokens": 2048, "reasoning_budget_tokens": 1024, "corpus_chars": 348127, "kv_cache_warm_for_this_prompt": true }, "measurement": { "prompt_tokens": 30341, "completion_tokens": 1916, "completion_token_source": "usage.completion_tokens", "time_to_first_token_s": 0.16, "prefill_tok_s": null, "prefill_note": "not measured: the prompt-fitting probes already sent this exact prompt, so the engine's KV cache was warm and time-to-first-token is a cache hit, not a prefill. Generation rate is unaffected.", "generation_seconds": 17.209, "generation_tok_s": 111.28, "generation_window": "first streamed token (reasoning or content) to last", "usage": { "prompt_tokens": 30341, "completion_tokens": 1916, "total_tokens": 32257, "prompt_tokens_details": { "cached_tokens": 30336 } }, "streamed_chunks": 1907, "answer_chars": 6930, "reasoning_chars": 1612, "answer_head": "Strata\u2019s engine is built around a practical split between the graphics card and the rest of the computer. The GPU is used for the parts of inference that benefit most from fast, parallel execution: dense matrix operations, attention kernels, quantized expert kernels, speculative decoding, and the portions of the model that can be kept resident in VRAM. The system RAM is used as a larger, slower but more flexible backing store for the model\u2019s many expert weights, for KV-cache data when the context is long, and for temporary buffers that cannot comfortably fit in the GPU\u2019s limited memory. The SS", "reasoning_head": "We need answer user's request. Need produce final answer only. Need use only document above. Need detailed technical summary of how engine divides work between GPU and system RAM, why, and what user must provide to run. At least 400 words continuous prose. Need likely no citations? Could reference docs. Need ensure only from provided docs. Need discuss GPU VRAM, RAM, experts, KV cache, mmap, resident, low-RAM, SSD, PLE, model download, requirements, API key, etc. Need continuous prose, not bullets? User says continuous prose. We can write paragraphs. Need at least 400 words. Need be detailed t" }, "problems": [], "passed": true, "ended_utc": "2026-10-05T06:18:10Z" }