Skip to content

laptop crash with latest #779

Closed as not planned
Closed as not planned
laptop crash with latest#779

Description

@XeonG
XeonG
opened on Oct 4, 2026
Image

win11, 16vram, 64gb ram usually 10gb free ram still with this running, vram is a different story.

strata generate: PCIe probe: 24.0 GB/s host->device (best of 24.0 24.0 24.0 24.0) -> pcie_frac 0.55 (default 0.55)
strata generate: this CPU has no AVX-512: the expert kernels run on AVX-2 (multi-token for the i-quant gate/up rows)
strata generate: native pack: E:\Strata-data\packs\swift-iq3_xxs experts (largest blob 2.33 MB), token embedding IQ3_S in mapped host memory (260 MiB)
strata generate: 1466 MiB of weights loaded from E:\Strata-data\packs\swift-iq3_xxs (302 canonical tensors skipped: served natively)
strata generate: 300 native projection matrices, 1781.11 MiB of weights
strata generate: the SSD is kept awake while rows are read: one page of the table after 100 ms without a read, until 60 s after the last request (STRATA_SSD_KEEPALIVE=0 turns it off)
strata generate: profile E:\Strata\data\expert-profile.bin: 24576 ranked pairs, built for 24576 slots
strata generate: KV streaming: 32768 of 131072 cells per QSA layer in VRAM, the K/V in 1.55 GiB of pinned RAM
strata generate: PLE on, table 320001536 rows of E:\Strata-data\models\swift-IQ3_XXS\Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf
strata mtp: draft layer loaded, 835 MiB of VRAM (experts 675, dense 111), files read in 0.20 s (3885 MiB/s)
strata generate: experimental native Q5_K head, 437043200 bytes
strata generate: GPU 0: NVIDIA GeForce RTX 4090 Laptop GPU, compute capability 8.9
strata generate: expert arena read unbuffered (2 of 32 probe reads from the file cache; 14.5 GiB available, 70.7 GiB of files)
strata generate: expert arena: VirtualLock stopped at 0 of 11452 MiB (error 87); cudaHostRegister of the whole arena FAILED (out of memory); 36 slices pinned (28 GiB); large pages (2097152 B)
strata generate: loaded 39.97 GiB at 1.42 GiB/s
strata generate: expert cache auto: 9.36 GiB free, 700 MiB reserved (+184 MiB for the draft head) -> 3915 slots
strata generate: expert cache 5235 slots, 8.49 GiB of VRAM; policy is
strata generate: the GPU computes the experts in the cache; it rounds differently from the CPU,
                 so a reply can differ slightly from a run without the cache (same quality:
                 bench/results/2026-09-27-cache-parity).
                 PROFILE, ranked by routing frequency, no eviction.
strata generate: pre-filled 5235 of 5235 slots from the profile; slot 0 verified
strata generate: R4 hit path ON - resident experts are computed on the GPU
strata generate: 23 expert-pool workers + the host thread
strata generate: session is up (engine 0.1.39)
strata generate: token graph hit path: 5235 resident experts, decided on the device
strata serve: prompt chunk auto: 8192 tokens, a 96-slot ring
strata serve: the prompt path borrows 2178 CUDA0 cache slots (3.50 GiB)
strata hc: CUDA0: the hyper-connection read runs as staged (the norm per token and stream, the down projection's activations staged ahead by cp.async); checked bit for bit against the plain read on this card (STRATA_HC_SPLIT=1 or 0 for the earlier ones)
strata verify: window up to 6 tokens, 69.7 MiB of device buffers
strata mtp: draft head over 106299 tokens (178.4 MiB)
strata serve: 444 MiB of VRAM free with everything loaded
strata generate: PCIe probe: 23.9 GB/s host->device (best of 23.9 23.7 23.9 23.9) -> pcie_frac 0.55 (default 0.55)
strata generate: this CPU has no AVX-512: the expert kernels run on AVX-2 (multi-token for the i-quant gate/up rows)
strata generate: native pack: E:\Strata-data\packs\swift-iq3_xxs experts (largest blob 2.33 MB), token embedding IQ3_S in mapped host memory (260 MiB)
strata generate: 1466 MiB of weights loaded from E:\Strata-data\packs\swift-iq3_xxs (302 canonical tensors skipped: served natively)
strata generate: 300 native projection matrices, 1781.11 MiB of weights
strata generate: the SSD is kept awake while rows are read: one page of the table after 100 ms without a read, until 60 s after the last request (STRATA_SSD_KEEPALIVE=0 turns it off)
strata generate: profile E:\Strata\data\expert-profile.bin: 24576 ranked pairs, built for 24576 slots
strata generate: KV streaming: 32768 of 131072 cells per QSA layer in VRAM, the K/V in 1.55 GiB of pinned RAM
strata generate: PLE on, table 320001536 rows of E:\Strata-data\models\swift-IQ3_XXS\Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf
strata mtp: draft layer loaded, 835 MiB of VRAM (experts 675, dense 111), files read in 0.69 s (1134 MiB/s)
strata generate: experimental native Q5_K head, 437043200 bytes
strata generate: GPU 0: NVIDIA GeForce RTX 4090 Laptop GPU, compute capability 8.9
strata generate: expert arena read unbuffered (1 of 32 probe reads from the file cache; 14.9 GiB available, 70.7 GiB of files)
strata generate: expert arena: VirtualLock stopped at 0 of 11452 MiB (error 87); cudaHostRegister of the whole arena FAILED (out of memory); 36 slices pinned (28 GiB); large pages (2097152 B)
strata generate: loaded 39.97 GiB at 1.43 GiB/s
strata generate: expert cache auto: 9.36 GiB free, 700 MiB reserved (+184 MiB for the draft head) -> 3915 slots
strata generate: expert cache 5235 slots, 8.49 GiB of VRAM; policy is
strata generate: the GPU computes the experts in the cache; it rounds differently from the CPU,
                 so a reply can differ slightly from a run without the cache (same quality:
                 bench/results/2026-09-27-cache-parity).
                 PROFILE, ranked by routing frequency, no eviction.
strata generate: pre-filled 5235 of 5235 slots from the profile; slot 0 verified
strata generate: R4 hit path ON - resident experts are computed on the GPU
strata generate: 23 expert-pool workers + the host thread
strata generate: session is up (engine 0.1.39)
strata generate: token graph hit path: 5235 resident experts, decided on the device
strata serve: prompt chunk auto: 8192 tokens, a 96-slot ring
strata serve: the prompt path borrows 2178 CUDA0 cache slots (3.50 GiB)
strata hc: CUDA0: the hyper-connection read runs as staged (the norm per token and stream, the down projection's activations staged ahead by cp.async); checked bit for bit against the plain read on this card (STRATA_HC_SPLIT=1 or 0 for the earlier ones)
strata verify: window up to 6 tokens, 69.7 MiB of device buffers
strata mtp: draft head over 106299 tokens (178.4 MiB)
strata serve: 444 MiB of VRAM free with everything loaded
strata generate: PCIe probe: 24.0 GB/s host->device (best of 23.8 23.9 23.9 24.0) -> pcie_frac 0.55 (default 0.55)
strata generate: this CPU has no AVX-512: the expert kernels run on AVX-2 (multi-token for the i-quant gate/up rows)
strata generate: native pack: E:\Strata-data\packs\swift-iq3_xxs experts (largest blob 2.33 MB), token embedding IQ3_S in mapped host memory (260 MiB)
strata generate: 1466 MiB of weights loaded from E:\Strata-data\packs\swift-iq3_xxs (302 canonical tensors skipped: served natively)
strata generate: 300 native projection matrices, 1781.11 MiB of weights
strata generate: the SSD is kept awake while rows are read: one page of the table after 100 ms without a read, until 60 s after the last request (STRATA_SSD_KEEPALIVE=0 turns it off)
strata generate: profile E:\Strata\data\expert-profile.bin: 24576 ranked pairs, built for 24576 slots
strata generate: KV streaming: 32768 of 131072 cells per QSA layer in VRAM, the K/V in 1.55 GiB of pinned RAM
strata generate: PLE on, table 320001536 rows of E:\Strata-data\models\swift-IQ3_XXS\Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf
strata mtp: draft layer loaded, 835 MiB of VRAM (experts 675, dense 111), files read in 0.20 s (3878 MiB/s)
strata generate: experimental native Q5_K head, 437043200 bytes
strata generate: GPU 0: NVIDIA GeForce RTX 4090 Laptop GPU, compute capability 8.9
strata generate: expert arena read unbuffered (0 of 32 probe reads from the file cache; 14.8 GiB available, 70.7 GiB of files)
strata generate: expert arena: VirtualLock stopped at 0 of 11452 MiB (error 87); cudaHostRegister of the whole arena FAILED (out of memory); 36 slices pinned (28 GiB); large pages (2097152 B)
strata generate: loaded 39.97 GiB at 1.43 GiB/s
strata generate: expert cache auto: 9.36 GiB free, 700 MiB reserved (+184 MiB for the draft head) -> 3915 slots
strata generate: expert cache 5235 slots, 8.49 GiB of VRAM; policy is
strata generate: the GPU computes the experts in the cache; it rounds differently from the CPU,
                 so a reply can differ slightly from a run without the cache (same quality:
                 bench/results/2026-09-27-cache-parity).
                 PROFILE, ranked by routing frequency, no eviction.
strata generate: pre-filled 5235 of 5235 slots from the profile; slot 0 verified
strata generate: R4 hit path ON - resident experts are computed on the GPU
strata generate: 23 expert-pool workers + the host thread
strata generate: session is up (engine 0.1.39)
strata generate: token graph hit path: 5235 resident experts, decided on the device
strata serve: prompt chunk auto: 8192 tokens, a 96-slot ring
strata serve: the prompt path borrows 2178 CUDA0 cache slots (3.50 GiB)
strata hc: CUDA0: the hyper-connection read runs as staged (the norm per token and stream, the down projection's activations staged ahead by cp.async); checked bit for bit against the plain read on this card (STRATA_HC_SPLIT=1 or 0 for the earlier ones)
strata verify: window up to 6 tokens, 69.7 MiB of device buffers
strata mtp: draft head over 106299 tokens (178.4 MiB)
strata serve: 444 MiB of VRAM free with everything loaded
strata verify: captured the 6-token window (upload no error, sync no error)
strata verify: captured the 4-token window (upload no error, sync no error)
strata verify: captured the 1-token window (upload no error, sync no error)
strata verify: captured the 2-token window (upload no error, sync no error)
strata serve: prompt 53 tokens = 0 reused + 53 read in 895 ms (59.2 tok/s), 33 generated in 603 ms (54.8 tok/s), drafts accepted 20 of 26, 1 checkpoints
strata serve: decode expert cache hit rate: 66.3% (10620 hits / 16022 lookups); 2698 more read by the GPU over PCIe or from another GPU (14.4% of all 18720 routed)
strata serve: KV streaming: 97.90% of 12576 block reads hit VRAM, 1.1 MiB read from RAM
strata verify: captured the 3-token window (upload no error, sync no error)
strata serve: prompt 59 tokens = 48 reused + 11 read in 477 ms (23.0 tok/s), 57 generated in 1053 ms (54.1 tok/s), drafts accepted 31 of 47, 2 checkpoints
strata serve: decode expert cache hit rate: 79.2% (25339 hits / 31976 lookups); 3544 more read by the GPU over PCIe or from another GPU (10.0% of all 35520 routed)
strata serve: KV streaming: 98.98% of 34212 block reads hit VRAM, 1.4 MiB read from RAM
strata serve: prompt 306 tokens = 0 reused + 306 read in 1645 ms (186.0 tok/s), 334 generated in 5313 ms (62.9 tok/s), drafts accepted 213 of 271, 1 checkpoints
strata serve: decode expert cache hit rate: 75.6% (124964 hits / 165201 lookups); 23439 more read by the GPU over PCIe or from another GPU (12.4% of all 188640 routed)
strata serve: KV streaming: 99.65% of 552024 block reads hit VRAM, 7.7 MiB read from RAM
strata serve: suffix drafts: 3 windows, 6 of 9 drafts accepted
strata serve: prompt 306 tokens = 0 reused + 306 read in 1020 ms (300.1 tok/s), 289 generated in 4588 ms (63.0 tok/s), drafts accepted 177 of 223, 1 checkpoints
strata serve: decode expert cache hit rate: 80.6% (116757 hits / 144809 lookups); 16471 more read by the GPU over PCIe or from another GPU (10.2% of all 161280 routed)
strata serve: KV streaming: 99.61% of 453240 block reads hit VRAM, 7.2 MiB read from RAM
strata serve: suffix drafts: 2 windows, 1 of 3 drafts accepted
strata verify: captured the 5-token window (upload no error, sync no error)
strata serve: prompt 28034 tokens = 0 reused + 28034 read in 17605 ms (1592.4 tok/s), 385 generated in 6897 ms (55.8 tok/s), drafts accepted 262 of 333, 3 checkpoints
strata serve: decode expert cache hit rate: 65.9% (120848 hits / 183313 lookups); 35567 more read by the GPU over PCIe or from another GPU (16.2% of all 218880 routed)
strata serve: KV streaming: 97.83% of 2834544 block reads hit VRAM, 247.8 MiB read from RAM
strata serve: suffix drafts: 10 windows, 31 of 37 drafts accepted
strata generate: PCIe probe: 24.0 GB/s host->device (best of 24.0 24.0 23.8 23.9) -> pcie_frac 0.55 (default 0.55)
strata generate: this CPU has no AVX-512: the expert kernels run on AVX-2 (multi-token for the i-quant gate/up rows)
strata generate: native pack: E:\Strata-data\packs\swift-iq3_xxs experts (largest blob 2.33 MB), token embedding IQ3_S in mapped host memory (260 MiB)
strata generate: 1466 MiB of weights loaded from E:\Strata-data\packs\swift-iq3_xxs (302 canonical tensors skipped: served natively)
strata generate: 300 native projection matrices, 1781.11 MiB of weights
strata generate: the SSD is kept awake while rows are read: one page of the table after 100 ms without a read, until 60 s after the last request (STRATA_SSD_KEEPALIVE=0 turns it off)
strata generate: profile E:\Strata\data\expert-profile.bin: 24576 ranked pairs, built for 24576 slots
strata generate: KV streaming: 32768 of 131072 cells per QSA layer in VRAM, the K/V in 1.55 GiB of pinned RAM
strata generate: PLE on, table 320001536 rows of E:\Strata-data\models\swift-IQ3_XXS\Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf
strata mtp: draft layer loaded, 835 MiB of VRAM (experts 675, dense 111), files read in 0.71 s (1111 MiB/s)
strata generate: experimental native Q5_K head, 437043200 bytes
strata generate: GPU 0: NVIDIA GeForce RTX 4090 Laptop GPU, compute capability 8.9

{
 "exe": "E:\\Strata\\engine\\strata.exe",
 "args": [
  "--pack",
  "E:\\Strata-data\\packs\\swift-iq3_xxs",
  "--native",
  "E:\\Strata-data\\models\\swift-IQ3_XXS\\Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf",
  "--ple-gguf",
  "E:\\Strata-data\\models\\swift-IQ3_XXS\\Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf",
  "--expert-profile",
  "E:\\Strata\\data\\expert-profile.bin",
  "--expert-cache",
  "auto",
  "--prefill",
  "auto",
  "--spec",
  "4",
  "--mtp",
  "E:\\Strata-data\\mtp\\rt",
  "--max-context",
  "131072",
  "--kv",
  "int8",
  "--kv-resident",
  "32768",
  "--spec-min-p",
  "0.5"
 ],
 "cwd": "E:\\Strata",
 "tokenizer": "E:\\Strata-data\\packs\\swift-iq3_xxs\\tokenizer",
 "model_name": "swift-1.5-iq3_xxs",
 "log": "E:\\Strata\\strata-swift-iq3_xxs.log",
 "lib_dirs": [
  "E:\\Strata\\.venv\\Lib\\site-packages\\nvidia\\cu13\\bin\\x86_64"
 ],
 "port": 8080,
 "gpu": 0,
 "gpus_asked": true,
 "open_browser": false
}

Before adding the new open_browser flag, the start process had opened firefox triggering a display crash

log file doesn't show model launch parameters, its not very clear when the model is launched or any timestamp, just minor stuff that would make it better.

And lastly this #765 think most problems I've face in the last 2 days of checking newer updates have been around vram usage and it just not working without additional settings trying to get the most out of it while not causing it crash out. For desktops this just isn't feasible I'd rather it release vram memory for other apps and claw it back when those apps are closed even if it means tk/s drops. It would be nice if it could manage itself a bit better with what vram is available while not hogging it to the point nothing else can run and it's just crashing anyway.

Activity

  1. BGonnermann commented on Oct 4, 2026

    @BGonnermann
    BGonnermann
    on Oct 4, 2026
    Contributor

    For the desktop crash, first test --vram-reserve-mib 3072. The log shows only 444 MiB free after startup; a larger reserve reduces the expert cache and leaves more room for Firefox. Keep --kv-resident 32768 and --prefill auto.

    For runtime sharing, enable "vram_elastic": true. Send /v1/vram {"reserve_mib":3072} before opening GPU-heavy apps, then {"reserve_mib":null} afterward to reclaim memory. This currently requires manual requests.

    For clearer diagnostics, write the server's launch arguments, timestamps, READY events, and restart reasons into the same log as engine stderr.

  2. Niko1221
    added 3 commits that reference this issue on Oct 6, 2026

    serve: vision temp files, Smart App Control message, engine start lin…

    f2c0b68

    setup: quick fixes and OS items (#657 #737 #670 #798 #775 #779 #841 #881

    cbb88e0

    pinned arena: no VirtualLock on large pages, no MADV_HUGEPAGE under T…

    08d3d9f

  3. Niko1221 commented on Oct 6, 2026

    @Niko1221
    Niko1221
    on Oct 6, 2026
    Owner

    0.1.40: the engine log gets a timestamped 'engine started: exe args' line, setup gives a --vram-reserve-mib 1500 tip when nvidia-smi says the card drives a display, and a large-page arena is no longer VirtualLocked. We still need the Kernel-Power 41 BugcheckCode, WHEA events and your config to say more.

  4. Niko1221
    added 3 commits that reference this issue on Oct 6, 2026

    serve: vision temp files, Smart App Control message, engine start lin…

    2ab6304

    setup: quick fixes and OS items (#657 #737 #670 #798 #775 #779 #841 #881

    fb57b6c

    pinned arena: no VirtualLock on large pages, no MADV_HUGEPAGE under T…

    bb19bec
  5. anon761
    added 3 commits that reference this issue on Oct 6, 2026

    serve: vision temp files, Smart App Control message, engine start lin…

    d41f34e

    setup: quick fixes and OS items (Niko1221#657 Niko1221#737 Niko1221#670

    5490c23

    pinned arena: no VirtualLock on large pages, no MADV_HUGEPAGE under T…

    c8a3d91

  6. Niko1221 commented on Oct 7, 2026

    @Niko1221
    Niko1221
    on Oct 7, 2026
    Owner

    Please retest on 0.1.40 with --vram-reserve-mib 3072. 0.1.40 prints a start line with the time and free VRAM, and no longer locks large pages. If Firefox still dies, send that line and the Windows memory figures.

  7. sskver
    added 3 commits that reference this issue on Oct 7, 2026

    serve: vision temp files, Smart App Control message, engine start lin…

    07f67b3

    setup: quick fixes and OS items (Niko1221#657 Niko1221#737 Niko1221#670

    8d57be9

    pinned arena: no VirtualLock on large pages, no MADV_HUGEPAGE under T…

    06a7d2d
  8. MistVVK
    added a commit that references this issue on Oct 8, 2026

    エキスパートのアリーナ: THP が always のときは MADV_HUGEPAGE を頼まない（upstream bb19bec の N…

    55bde35
  9. luckywang97
    added a commit that references this issue on Oct 10, 2026

    Desktop launcher (STRATA-GUI.bat) and an opt-in --low-vram profile fo…

    4027e18
  10. Buddychek
    added 2 commits that reference this issue on Oct 11, 2026

    serve: a start that never reached READY gives the engine's log back (N…

    56e0be8

    docs: the engine taken mid-request, the failed start's log handle (Ni…

    466e64a
  11. MistVVK
    added a commit that references this issue on Oct 11, 2026

    setup とサーバー: 画面も出すカードに VRAM の予約を勧める（upstream 40ae2d1・fb57b6cd8 の Niko…

    50844da

  12. Niko1221 commented on Oct 13, 2026

    @Niko1221
    Niko1221
    on Oct 13, 2026
    Owner

    Thanks for the report. To keep the tracker current we are closing issues older than three days: 0.1.44 is out with a long list of fixes, and many older reports are covered by it or by 0.1.43.1.

    If this still happens on 0.1.44, please leave a comment here with your engine log (and the version line it prints), and we will reopen it. A new issue that links this one works too.

  13. Niko1221
    closed this as not plannedon Oct 13, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions