Skip to content

OOM error with 16GB VRAM and 64GB RAM #60

Closed
Closed
OOM error with 16GB VRAM and 64GB RAM#60

Description

@shanoor
shanoor
opened on Sep 28, 2026

I tried to run swift1.5 iq2_xs and it doesn't load, I get an OOM error despite the machine having 16GB VRAM and 64GB RAM. I think it might related to the fact that Windows only gives 32GB of shared gpu memory to my GPU (I don't know why).

Log :

load_hparams: Qwen-VL models require at minimum 1024 image tokens to function correctly on grounding tasks
load_hparams: if you encounter problems with accuracy, try adding --image-min-tokens 1024
load_hparams: more info: https://github.com/ggml-org/llama.cpp/issues/16842

strata-vision: warmed up at 1024 image tokens
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: D:\llm\Strata-data\packs\swift-iq2_xs experts (largest blob 1.51 MB), token embedding type 23 in mapped host memory (322 MiB)
strata generate: 1466 MiB of weights loaded from D:\llm\Strata-data\packs\swift-iq2_xs (302 canonical tensors skipped: served natively)
strata generate: 300 native projection matrices, 1474.85 MiB of weights
strata generate: KV streaming: 32768 of 65536 cells per QSA layer in VRAM, the K/V in 0.77 GiB of pinned RAM
strata generate: PLE on, table 320001536 rows of D:\llm\Strata-data\models\swift-IQ2_XS\Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ2_XS-00001-of-00002.gguf
strata generate: expert arena: locked 3688 MiB via working-set minimum + VirtualLock; cudaHostRegister of the whole arena FAILED (out of memory); 43 slices pinned (29 GiB); large pages refused for 35456548864 B (GetLargePageMinimum=2097152, VirtualAlloc error 1450); using 4 KB pages
strata generate: loaded 33.02 GiB at 2.08 GiB/s
strata mtp: draft layer loaded, 835 MiB of VRAM (experts 675, dense 111)
strata generate: profile D:\llm\Strata\data\expert-profile.bin: 24576 ranked pairs, built for 24576 slots
strata generate: experimental native Q5_K head, 337715200 bytes
strata generate: expert cache auto: 9.96 GiB free, 700 MiB reserved -> 6598 slots
strata generate: ExpertCache: cudaMalloc(9.28 GiB) failed: out of memory

I'm using Windows 10

Image Image Image

Activity

  1. itguy327 commented on Sep 28, 2026

    @itguy327
    itguy327
    on Sep 28, 2026

    similar issue here

  2. technerd-official commented on Sep 28, 2026

    @technerd-official
    technerd-official
    on Sep 28, 2026

    same here have a single 3090 with 64gb system ram and still cant get it to run. (ExpertCache: cudaMalloc failed: out of memory).

  3. Niko1221
    added a commit that references this issue on Sep 28, 2026

    Expert cache: retry smaller when the allocation fails; setup warns ab…

    6a6de69

  4. Niko1221 commented on Sep 28, 2026

    @Niko1221
    Niko1221
    on Sep 28, 2026
    Owner

    Thanks @shanoor - this is fixed as far as the engine can fix it, in v0.1.19.

    The cause (not the 32 GB shared-memory figure; the engine copes with that by pinning in slices):

    • On Windows, memory on the graphics card also has to fit in Windows' commit, which is RAM plus the page file.
    • Your screenshot shows a commit limit of 64.5 GB with 63.7 GB of RAM, so the page file is about 1 GB or off.
    • With the 33 GB expert arena loaded, the 9.3 GB expert cache no longer fit, although the VRAM was free, and the engine stopped.

    What 0.1.19 does:

    • The engine retries with a smaller cache instead of stopping, and says how much commit Windows has left.
    • Setup warns when the page file is under 4 GB.

    The real fix on your PC: set the page file to System managed, then restart. The setting is under System > About > Advanced system settings > Performance > Advanced > Virtual memory. Then the cache can use all the free VRAM.

    Could you update (git pull, then START-HERE.bat) and tell me whether it starts - with and without the page-file change? If it still fails, the lines the engine prints about the expert cache would help.

  5. Niko1221 commented on Sep 28, 2026

    @Niko1221
    Niko1221
    on Sep 28, 2026
    Owner

    @technerd-official @itguy327 - 0.1.19 retries with a smaller expert cache instead of stopping, and on Windows the real fix is setting the page file to "System managed" (see my comment above). If it still fails after updating, please say which OS you're on and paste the lines starting with strata generate: expert cache.

  6. technerd-official commented on Sep 28, 2026

    @technerd-official
    technerd-official
    on Sep 28, 2026

    @technerd-official @itguy327 - 0.1.19 retries with a smaller expert cache instead of stopping, and on Windows the real fix is setting the page file to "System managed" (see my comment above). If it still fails after updating, please say which OS you're on and paste the lines starting with strata generate: expert cache.

    ah okay thank you, I had to set pagefile to 0 because my SSD is already full with all these huge LLM models. Will try again after I allocate some more space on the SSD.

  7. clhaga
    added a commit that references this issue on Sep 29, 2026

    Dual-GPU fix: each side's NativeDense uploads only its own layers

    adeb27f

  8. Niko1221 commented on Sep 29, 2026

    @Niko1221
    Niko1221
    on Sep 29, 2026
    Owner

    Closing: since 0.1.19 the engine retries with a smaller expert cache instead of stopping, and on Windows the page file needs room for the graphics driver (set it to "System managed"; setup warns when it is small). Thanks @technerd-official and @itguy327 - reopen if it still happens with some page file space.

  9. Niko1221
    closed this as completedon Sep 29, 2026

  10. shanoor commented on Sep 30, 2026

    @shanoor
    shanoor
    on Sep 30, 2026
    Author

    I changed windows page file to system managed, same error. I then tried to force it to use between 4096 and 32768MB and still the same OOM error

    load_hparams: Qwen-VL models require at minimum 1024 image tokens to function correctly on grounding tasks
load_hparams: if you encounter problems with accuracy, try adding --image-min-tokens 1024
load_hparams: more info: https://github.com/ggml-org/llama.cpp/issues/16842

strata-vision: warmed up at 1024 image tokens
strata generate: PCIe probe: 24.1 GB/s host->device -> 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: D:\llm\Strata-data\packs\swift-iq2_xs experts (largest blob 1.51 MB), token embedding type 23 in mapped host memory (322 MiB)
strata generate: 1466 MiB of weights loaded from D:\llm\Strata-data\packs\swift-iq2_xs (302 canonical tensors skipped: served natively)
strata generate: 300 native projection matrices, 1474.85 MiB of weights
strata generate: KV streaming: 32768 of 65536 cells per QSA layer in VRAM, the K/V in 0.42 GiB of pinned RAM
strata generate: PLE on, table 320001536 rows of D:\llm\Strata-data\models\swift-IQ2_XS\Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ2_XS-00001-of-00002.gguf
strata mtp: draft layer loaded, 820 MiB of VRAM (experts 675, dense 111), files read in 0.32 s (2421 MiB/s)
strata generate: expert arena: locked 2951 MiB via working-set minimum + VirtualLock; cudaHostRegister of the whole arena FAILED (out of memory); 44 slices pinned (30 GiB); large pages refused for 35456548864 B (GetLargePageMinimum=2097152, VirtualAlloc error 1450); using 4 KB pages
strata generate: loaded 33.02 GiB at 2.04 GiB/s
strata generate: profile D:\llm\Strata\data\expert-profile.bin: 24576 ranked pairs, built for 24576 slots
strata generate: native head upload: out of memory

    Edit : well, by putting a 64GB limit on the pagefile, it works.

  11. MistVVK
    added a commit that references this issue on Oct 3, 2026

    Expert cache: retry smaller when the allocation fails; setup warns ab…

    d03be0d
  12. ScrambledEggsOnToast
    added 2 commits that reference this issue on Oct 4, 2026

    Charge a committed allowance for the two opaque device classes

    3b9ce43

    Price the guest-scope opaque class inside an enforced cgroup ceiling

    4917e23
  13. Hugua700
    added a commit that references this issue on Oct 6, 2026

    Expert cache: retry smaller when the allocation fails; setup warns ab…

    069b92b
  14. anon761
    added a commit that references this issue on Oct 6, 2026

    Expert cache: retry smaller when the allocation fails; setup warns ab…

    689ce16
  15. Niko1221
    added a commit that references this issue on Oct 10, 2026

    setup: the page-file warning advises a fixed size, and says why when …

    c396dc9
  16. giridhart
    added a commit that references this issue on Oct 15, 2026

    setup: the page-file warning advises a fixed size, and says why when …

    791cc43
  17. DarthCY-K
    added a commit that references this issue on Oct 20, 2026

    page file: count only what is there for sure; a growing page file is …

    2be22ae
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