Problems people reported
Community · 600 items
- [Engine] strata-inkling 0.1.65: Inkling-2-Mini SQ-15 on a 12 GB RTX 3060 — OOM after load (default backend), empty answers + segfault in gemm_sk_q8_1 (hybrid backend)
- [Engine] Model files are hashed once at install but never verified at load; damaged or swapped shards are served silently
- [Other] Setup picks model tier, size and image encoder from total VRAM, ignoring memory already held
- [GUI] Host shutdown exits without draining the chat-save queue, dropping the last edit
- [Engine] Per-layer VRAM expert-cache admission is measured 2.97% vs 21-70% but left off by default
- [Engine] Speculative decoding drafts argmax for sampled requests, so almost every drafted token is rejected
- [Engine] SSD conversation tier permanently deletes a snapshot when the read-back RAM gate refuses it
- [Engine] Single-GPU start with a too-large K/V cache exits instead of falling back to KV streaming
- [Engine] 0.1.65 outputs random multilingual tokens on NVFP4 Qwen3.8-Flash-Next with the same config that works on 0.1.64
- [Engine] Periodically rebalance the resident expert cache as free VRAM changes
- [Engine] Large-context compaction re-reads history from token 0 when the snapshot cannot be parked
- [Engine] Allow an opt-in resident expert cache to lend VRAM to vision
- [Engine] Engine 0.1.65 fails to start after update on Windows 11
- Hy4-Flash: prompt/KV cache always resets to 0 in multi-turn agentic loops (--session-reuse 1 missing + strict memcmp on BPE re-tokenization)
- [Engine]Hy4-Flash SQ-64 Windows CUDA13 engine 0.1.64 repeatedly crashes
- [Engine] hip-rx9070 (gfx1201): Inkling-2-mini prompts of 64+ tokens fail in prefill with NaN logits; --set chunk=<64 works
- [Engine] Linux mixed AMD GPUs (gfx1201 R9700 + gfx1100 RX 7900 XTX) on 0.1.65: rx7900 engine rejects GPU 0 (gfx1201)
- Docker: inkling-2-mini setup crashes with AttributeError ('int' object has no attribute 'get') - install_inkling2 expects GPU dicts but the docker preset passes int indices
- [Engine] Hy4-Flash SQ-128 at 128K on Strix Halo (hip-halo, 128 GiB unified): OOM-killed after [load]; --vram-experts-gb and HY4_VRAM_CAP_GB ignored
- [Engine] Inkling-2-Mini SQ-15 never starts on RTX 3060 12 GB: CUDA error 2 (out of memory) at ink_cuda.cuh:54 even after auto-sizing
- [Engine] strata-inkling 0.1.65 (gfx1101): non-finite logits for every prompt of 64+ tokens in serve mode
- [Engine] Multi-GPU: Inkling 2 Mini loads on a single GPU only (main/helper setting ignored), crashes when context fills VRAM
- [Engine] strata-inkling: GGUF --native never reaches the CUDA backend ("Invalid .strata magic")
- [Engine] Inkling-2-Mini on RTX 3090 24 GB: long prompts OOM at ink_cuda.cuh:54 unless --max-context is small (SQ-20 and SQ-15)
- [Engine] strata-inkling: --bench with --prompt-len >= 32768 without --max-context triggers CUDA error 700/715 and NVRM Xid 13 on RTX 3090 (sm_86)
- [Engine] strata-inkling: --kv k8v8 and --kv k8v4 SIGSEGV immediately after load on Linux RTX 3090 (sm_86)
- [Other] serve: web UI static assets randomly fail (connection reset / empty reply) with parallel connections; `Bad file descriptor` in serve.log
- Tensor-derived geometry fallback for DeepSeek V4.1 Flash GGUF compatibility
- [Engine] Inkling-2-Mini --gpus 2 crashes right after load on Windows (2x RTX 5070 Ti): access violation in strata-inkling.exe, server restart loop
- [Engine] Inkling-2-Mini: multi-turn chats never reuse the prefix, so every turn re-reads the whole conversation (hidden thinking is in the slot but never in the next prompt)
- [GUI] helpers on one card: three read-only helpers ran 163, 266 and 351 minutes, with no cost warning and no ceiling
- [Engine] 16 GB cards: guard the ~1 GB VRAM floor during KV growth and document the per-format context limit
- [Other] the shipped Harness 0.2.1-alpha.3 has no model-switch pressure check, added upstream in dsh-v0.2.1-alpha.4
- [Engine] 384K compaction re-reads 264K tokens (456 s): content-addressed, position-independent segment cache for near-zero-reread compaction
- [Engine] Compaction re-reads 331K-342K tokens five times (19.5-20.1 min each): the cache budget is the blocker (PR #4338)
- [Engine] 0.1.65 has broke 0.1.65-windows-x86_64-cuda-cuda13
- [Benchmark] hipBLASLt tuning table by default on gfx1100 (RX 7900 XTX, IQ3_S): speed, decode and logits
- [Benchmark] RTX PRO 4500, EPYC 73F3, 256GB
- [Engine] The default MTP `draft_vocab.bin` contains only 0.4% of the Arabic tokens: Arabic decode accepts 42% of drafts; adding the missing tokens gives 74% (+38% tok/s)
- [Engine] Inkling-2-Mini (.strata): DSpark speculative decoding never engages (drafting: off, drafts_offered: 0)
- [Engine] sm_86 Windows 16 GB: engine 0.1.65 decodes 30-40% slower than 0.1.56/0.1.57 (RTX 3080 Ti Laptop)
- [Engine] Inkling-2-Mini SQ-15 on Metal (M6, 20.5 GiB): every reply is token id 0 ("!"), finish_reason length
- [Engine] serve Inkling-2-Mini via serve/server.py — SIGSEGV on any prompt above ~40 tokens (--bench is fine)
- [Docs] settings-register: 41 HIP knobs land without a status in 0.2.19 — the pristine tag fails tools/test_settings_register.py
- [Engine] model is unable to distingush - and /
- [Benchmark] Quick Agent benchmark silently changes workload after the 0.2.16 layout move
- [Docs] Please keep in sync README & MODELS
- [Engine] Inkling-2-Mini SQ-20 on RTX 3080 12 GB: wrong tokens with experts on the host and pf_buf=0MB (--cpu is correct)
- Deepseek RE-ENABLE NOW!!
- [GUI] Update engine smoke-test passes --engram; engine 0.1.65 rejects it ("unknown argument"), so every auto-update reports failure
- [GUI] Win 0.2.16 with V100+5070: panel downloads cuda13 engine "for V100" (no sm_70), launches on 5070; manual cuda12 works on V100 but UI still reports the 5070
- [Engine] Inkling SQ-10 slow on my Windows 11 / RX 9060 XT 16GB / 64GB RAM setup; mismatched API model IDs silently use Qwen
- [Engine] Inkling 2 Mini: empty reply with finish_reason "length" when tools are present
- [GUI] Regression: default GPU selection for 8060S (gfx1151) changed to R9700 (gfx1201) between 0.2.12 and 0.2.19
- Engine 0.1.65 (NVFP4 GPTQ pack): EXP_MODE on = 0 drafts accepted for every drafter; --prefill 32768 = NaN at 24k
- strata-inkling 0.1.65 crashes with SIGILL on CPUs without AVX2 (Xeon E5 v2)
- [Engine] 0.1.65 HIP gfx1200: illegal memory access in batched prefill (Inkling-2 GGUF, NVFP4→SF4 experts); short prompts are fine
- [Engine] Inkling-2-Mini SQ-10 segfaults in CPU expert computation on RTX 4070 SUPER
- [Engine] Windows SYCL: the Level Zero V2 adapter fails at startup with UR_RESULT_ERROR_UNKNOWN; the legacy adapter works (Arc Pro B70, oneAPI 2026.2)
- [Engine] Inkling-2-Mini setup stops on a PC with two NVIDIA cards: AttributeError 'int' object has no attribute 'get' in install_inkling2
- [Engine] Hy4-Flash SQ-64 on RX 9070 XT 16 GB: 18-24 tok/s decode against 62 tok/s on the model page
- [TUI] Inkling-2-Mini SQ-15 planned as full-VRAM on a 16 GB Radeon runs out of memory at about 8.6K tokens of a 32K context
- [Engine] Inkling-2-Mini: replies start with a thinking span that the server discards (hidden tokens, empty answers at small max_tokens)
- Inkling 2 Mini (0.1.65, Windows CUDA, 16 GB): empty answers from hidden thinking; experts forced into VRAM (SQ-15 pages, hybrid crashes)
- [Benchmark] R9700 + Threadripper PRO 3975WX: Hy4-Flash at ~30 tok/s (SQ-64, 4K) vs. 62 tok/s in the 0.2.18 notes
- [Engine] gfx1100 (RX 7900 GRE): 0.2.16 cold-prefill gap is driver pauses, not kernels — 4 "Linux moved memory" pauses/25min vs 0 on 0.1.57
- [Engine] sm_75 inkling path: --kv k8v8/k8v4 segfault, CUDA 700 when prompt+gen fills --max-context, DSpark OOM at 131K
- [Engine] inkling2 Metal: ink_moe_gu_situ fails to compile on M1 Max, MTLCompilerService aborts with "cannot select
- [GUI] Please let app show release version.
- [Engine] 0.2.19 sm_120 defaults break the FP4-Local (nvfp4_local_pack) Qwen3.8-Flash-Next pack: EXP_MODE=8 gives garbled output, PREFILL_NVFP4=w4a4 gives non-finite logits (#879) on long prompts
- [Engine] Linux stable feed has no HIP gfx1150/igpu engine package for 0.2.19
- [Engine] Inkling-2-Mini (CUDA, RTX 5090): prefill hangs for prompts above ~20-27k tokens, GPU idle, no token generated
- [Benchmark] 70-minute soak on a two-card layer split (RTX 3060 + RTX 5060 Ti, 0.2.18): no errors, no slowdown, temperature-0 answers vary
- [Benchmark] DeepSeek-V4.1-Flash IST-DAS 2.25 bpw on RTX 5060 Ti — 256K/512K/1M parameter sweep - strata 0.2.13.1 - NUMA RAM 330GB
- [Engine] strata-hy4: dual-GPU expert-parallel support (--gpus 2)