トピック / Security
Security
Auto topic: Security
Issues
- #5893 [Engine] API key changes when unloading and reloading a model, invalidating existing API clients
- #6438 [Engine] Strata 2.17 + Lazarus 4.8 + Lazarus MCP Assistent not working ?
- #6489 [Engine] 16K cold prefill regression on a single RTX 4090 D (sm_89, IQ3_S): 0.1.55 → 0.1.60/0.1.63 (−21 %), unrelated to --prefill, --expert-cache or the route tail skip
- #6273 [Engine] 16 GB cards: guard the ~1 GB VRAM floor during KV growth and document the per-format context limit
- #6493 [Engine] serve loop: GENI prompt can write past the position table; --cvec offsets unchecked (static analysis)
- #6409 [Engine] Qwen3.8-Next-Flash on Strix Halo AI Max 395 (128GB UMA): tuning for 2 concurrent streams at >100 tok/s each @ 200k context
- #6478 [Engine] Expert-cache hit rate and decode throughput regress from engine 0.1.52 to 0.1.62/0.1.63 on two 11 GB Turing cards (Qwen3.8-Flash-Next GSQ-RCO IQ3_XXS)
- #6448 [Engine] Scope: plan an integrated Intel GPU (Arc B390, Panther Lake) as one pool of memory with the system in the SYCL engine
- #5990 [Engine] V100 (sm_70) on Linux: 13× Xid 79 + PCIe CmpltTO across engines 0.1.35–0.1.59.1 — not version-specific
- #6019 [Engine] Rare non-finite logits (#879) with native IQ2_XS and IQ3_XXS packs on sm_86, engine 0.1.59.1 (none with Q2_0 on the same flags)
- #5807 [Engine] DFlash2 drafting confidence is mis-tuned for a 16 GB Ada card: `CONF_STEPS` chooses 0.8, the measured optimum is 0.5, worth +16.7 % decode
- #5778 [GUI] Scope opt-in side questions, deeper research and private workflow skills
- #5654 [Engine] Dual RTX 5060 Ti over OCuLink: Xid 79 during tune auto (0.1.58, Linux driver 615.71.09 open)
- #5653 [Engine] SYCL (Arc): systematic instruction-following gap vs CUDA on identical weights — 2 % vs 77 % on paired probe
- #5608 [Engine] RX 9070 XT gfx1201: repeatable GPU page faults during IQ3_S inference on HIP 0.1.57/0.1.58
- #5559 [Engine] 0.1.58 hip-rx9070 (gfx1201): strata-host crashes repeatedly (GPF in libhsa-runtime64, exit -6/-11) under 35K-92K prompts; 0.1.57 runs clean
- #5812 [TUI] strata launch: moving the Access-step highlight makes an API key a local launch keeps
- #6062 [Engine] Prometheus /metrics: request token histograms, finished_reason, conversation-cache gauges, telemetry key
- #6322 [Engine] v0.2.16 source build on gfx1201 segfaults before any output with system ROCm 7.16; STRATA_HIP_QUEUE_PRIME=0 and the official rx9070 package are unaffected
- #6285 [Engine] serve: a tool call that does not match the declared schema reaches the client unchanged and is dropped
- #6276 [Engine] Re-land the conversation-cache disk-spill tier (closed #5399) with the disk-restore erase-order fix
- #6270 [Engine] vision: every picture fails with 'ERR cannot write <hash>.sve' on PCs where Windows blocks file creation by newly-landed unsigned binaries (work dir is under %TEMP%)
- #6103 [Other] Publish the Docker images for v0.2.14 and v0.2.15 (ghcr still stops at 0.2.13-cuda13)
- #6258 [Benchmark] AMD Radeon 880M (gfx1150, Linux) / Strata 0.2.12. swift-qwen38-flash-next-iq2xs-abliterated-IQ2_XS
- #5795 [Engine] ROCm build fails due to: `__shared__ alignas(16) int4 ...`
- #6146 [Benchmark] AMD Radeon 890M (gfx1150, Windows) / Strata 0.2.13. qwen3.8-flash-next-unsloth-ud-iq2_xs
- #6104 [Benchmark] RTX 3070 Ti + RTX 3070 (8 GB each, x16/x4): R-resident prefill, LM cache loans, and 0.1.60 migration
- #5857 [Benchmark] 2× RTX 3060 12 GB, vfio-passthrough KVM VM — qwen3.8-flash-next iq3_s, layer split
- #6180 [Other] Control-app /v1/models strips meta.n_ctx/status/architecture the engine exposes — breaks context-window auto-detection for OpenAI-compatible clients
- #5836 [Engine] Native /v1/models still has no context window on 0.2.13 — engine reports n_ctx, the host gateway drops it
- #5546 [Engine] One card or two: print the prompt-length crossover at start-up
- #5575 [Engine] test_security.py: test_ipv6_loopback fails instead of skipping when ::1 connect is blocked
- #5569 [Engine] IQ3_S grouped GEMM: half2-SIMD dequantization + a decode tile, -24%/-36% on sm_70 (bit-exact)
Pull requests
- #6520 [Engine] Inkling-2-Mini from the official NVFP4/MXFP8 checkpoint, expert offload and CPU experts for 8-16 GB cards
- #3553 [Engine] let a HIP helper read pinned expert rows in place
- #3552 [Engine] skip file-tier experts while a helper GPU is attached
- #3551 [Engine] add an opt-in CUDA and HIP hybrid build
- #4551 [Engine] session-file v3: incremental append saves under --session-save-append
- #3800 [Engine] Token mask for JSON response formats: MTP drafts survive constrained decoding
- #5436 [Engine] opt-in STRATA_ARENA_NUMA: place the expert arena per NUMA node on two-socket Linux
- #6102 [Engine] use runtime expert count in layer-major prefill
- #6485 [Engine] Add experimental gfx90a HIP build
- #2965 setup: a separate opt-in layers 8-33 projection vector for the original Qwen3.8-Flash-Next (any quantization)
- #3957 [Engine] serve: a long conversation survives an engine restart - parked conversations saved at a graceful stop and loaded at the next start, to files or (Windows) kept in RAM with no disk writes (opt-in STRATA_PARK_DIR, STRATA_PARK_RAMLIKE=1; layer split included)
- #1614 [Engine] shared KV pool for all lanes (--kv-pool-tokens): at 2 x 256k the first token comes 2-5x sooner and lane moves ~10x faster; off = identical to main
- #6453 [Engine] SYCL: an integrated Intel GPU is planned as one pool of memory with the system (Arc B390)
- #6034 [Engine] Parked conversations spill to a disk tier instead of being dropped
- #6009 [Docs] speed projection: a request that switches reads its prompt again, the other setting's parked conversation stays (measured on 2x P100)
- #5721 [GUI] Fix MCP tool forwarding in Desktop Serving API
- #5500 [Engine] L2 session cache: cut prefill time by ~95% on returning conversations
- #3348 [GUI] Add an iPhone companion with shared chats and model downloads
- #5201 [Engine] One-read top-k: double the sampling stride past 524,288 cells instead of failing every query
- #3500 [Engine] add GET /v1/models/{id} so Visual Studio can use Strata
- #3416 [Engine] Preserve target probabilities in gated MTP
- #2955 [Engine] f32 -> bf16 conversions in PLE and prefill keep a NaN a NaN (as bf16_bits.hpp does)
- #6360 [Benchmark] 2x RTX 3090 / Qwen3.8-Flash-Next / IQ3_S / standard-v1 draft protocol (MTP + DFlash2)
- #6367 [Benchmark] unlocked RTX 4080 (sm_89) / Qwen3.8-Flash-Next (IQ3_S · Coder IQ1_M · sc117 · UD-Q4_K_XL) / strata-qwen38-thinking-v1-draft.1
- #3756 [Engine] Metal: the control vector (experimental speed projection) on the Mac engine
- #5776 [Other] POST /tokenize: the prompt a request would read, without running it
- #3380 [Other] serve: an all-cached image turn never wakes the elastic encoder
- #3379 [Other] serve: a tool's picture starts the elastic encoder, and its turn counts
- #6140 [Benchmark] ab_engine: compare Strata versions with their own defaults, isolated from the install
- #5374 [GUI] show saved KV format for packaged Qwen3.8 models
- #5303 [Other] packaging: swap the Terminal bundle's stale pip-vendored copies and re-align its SBOM (the last CVE carriers)
- #4312 [TUI] launcher: start the bundled runtime when access(X_OK) is denied (Docker's default seccomp profile)
- #4161 [TUI] readiness: probe /v1/telemetry with the telemetry key when there is no API key
- #5081 [GUI] let the side panel's page preview use pointer lock
- #5884 [GUI] newSession keeps the model the composer shows
- #5359 [GUI] recover chat revisions after an index write failure
- #3267 [Engine] LoRA adapter support with per-request on/off switch
- #5992 [Engine] a GPU wait that runs out is an error for the window (xe spin bound follows the host)
- #5930 [GUI] Return HTTP 400 for malformed inference JSON
- #5925 [GUI] Fix Responses API 404 over LAN sharing
- #6005 [Engine] read tools from Anthropic tool_addition blocks (#5952)
- #5244 [Engine] serve: accept Codex multi_agent_v2 agent_message and multi_agent_call(_output) items on /v1/responses
- #4210 [Engine] serve: /v1/responses accepts Codex's agent_message input items (multi-agent v2 workers)
- #4621 [Engine] Responses API: run Codex's client-side tool_search, finds declared where the search answered
- #4758 [Engine] send the stream role with the first response data
- #2870 [Engine] serve: put the streamed usage in its own final chunk when include_usage is requested
- #4834 [Engine] Anthropic errors on /v1/messages use Anthropic's error shape, so the Vercel AI SDK shows Strata's message
- #3527 [Engine] Metal: release old prompt scratch buffers when the workspace grows
- #4141 [Engine] metal: fix wrong prompt results with K-quant packs: the gate/up grid follows the tile (carries #2903)
- #1492 [Engine] save and restore sessions without an MTP drafter
- #5690 [Benchmark] add synthetic two-request reproduction for shared Qwen repetition (#606)
- #5486 [Benchmark] AMD Radeon RX 7600 XT 16 GB (gfx1102), CachyOS Linux, Ryzen 5 3600
- #5446 [Benchmark] 2x RX 6800 XT (x8 + x8) on Windows / UD-Q4_K_XL / Strata 0.2.11 / engine 0.1.57
- #5442 [Benchmark] RTX 5080 / Qwen3.8-Flash-Next-GSQ-RCO / IQ3_S / strata-qwen38-thinking-v1-draft.1
- #5431 [Benchmark] RTX 4070 Ti SUPER 16 GB / Qwen3.8-Flash-Next IQ2_XS / 0.2.11 engine 0.1.57
- #4476 [Benchmark] redact credentials by whole word; a value-less flag no longer hides the next option
- #6142 [Engine] Gateway scheduling on top of the L2 session cache: cut prefill time ~90% on returning conversations — bytes-only budgets, cost-gated restore, fail-closed admission
- #6185 [GUI] Use the native Windows folder picker in Strata Desktop
- #3333 [Other] Refresh conversation-cache metrics after refused parking
- #3420 [GUI] Web search and page reading for the web app's Chat (DuckDuckGo default, SearXNG, Brave, Tavily, Serper, Exa, Perplexity)
- #5805 [Engine] serve: read a tool call written without its wrapper inside the reasoning, and unwrap a lone <parameter=arguments>
- #6105 OpenAI streaming: opt-in continuous_usage_stats (vLLM-style cumulative usage on chunks)
- #4113 [Engine] GBNF constrained decoding: stable tool calls + JSON schema, now riding the MTP drafter
- #5592 [Engine] Jetson AGX Thor (sm_110): the aarch64 CUDA engine carries machine code for it, and the Terminal runs it (#4559)
- #4522 [Engine] layer split: trim every card's dense weights on --layer-split auto too (from Strata-DualGPU, #642)
- #5399 [Engine] Parked conversations spill to a disk tier instead of being dropped
- #1493 [Engine] Add opt-in memory relief and supervised tool resource leases
- #2399 [Engine] allow elastic expert cache and growing K/V on Linux HIP
- #5454 [Benchmark] 2x RX 6900 XT (gfx1030), standard protocol on a v0.2.11-based build: 4K / 16K 1,234 / 2,124 (two cards), 1,058 / 1,434 (one card)
- #5203 [Engine] STRATA_SELECT_F16=1: the prompt path's block scores on FP16 tensor cores (opt-in, sm_80+, 1M read selection 95 -> 62 s)
- #4889 [TUI] doctor: say when the old Docker seccomp profile, not the bundled Python, is what fails
- #4316 [Docs] DOCKER.md: name the old-daemon start that looks like a card or image problem
- #4314 [Engine] reject a boolean or a fractional token limit with 400 (Fix #4037)
- #2522 [Engine] #533: elastic VRAM per GPU of a layer split (VRAM <reserve_mib> [device])
- #5685 [Engine] AMD RX660xt (gfx1032) support [Docs] AMD benchmark details corrected so title isn't generic [Benchmark] AMD RX6600xt