Skip to content

[Engine]Hy4-Flash SQ-64 Windows CUDA13 engine 0.1.64 repeatedly crashes #6528

Open

Description

@rxysww

Hy4-Flash SQ-64 Windows CUDA13 engine 0.1.64 repeatedly crashes with STATUS_ILLEGAL_INSTRUCTION (0xC000001D), fault offset 0x102CAD.
Hardware: Intel Core i9-14900HX (AVX2, no AVX-512), RTX 5090 Laptop 24GB, 64GB RAM, Windows 11.
Both Strata Desktop 0.2.18 and Terminal 0.2.18 reproduce the issue. Model verification succeeds, then the engine crashes approximately every 60 seconds during loading and the supervisor restarts it.
Please check for unguarded AVX-512 instructions in the prebuilt strata-hy4.exe and advise whether a portable AVX2 build is available.

Activity

  1. github-actions commented on Nov 21, 2026

    @github-actions

    Thanks for the issue. To keep the tracker readable, every item needs an area prefix in the title and the parts of the issue form filled in. This one is missing:

    • Start the title with its area: [Engine], [GUI], [TUI], [Benchmark], [Docs] or [Other], then a short description. Example: "[Engine] decode stops after 2 minutes".

    Edit the issue (title and description) and this note goes away on its own. If it is still missing parts in 5 days it will be closed, which is not a rejection: fix it and comment /reopen.

    Details: CONTRIBUTING.md

  2. changed the title [-]Hy4-Flash SQ-64 Windows CUDA13 engine 0.1.64 repeatedly crashes[/-] [+][Engine]Hy4-Flash SQ-64 Windows CUDA13 engine 0.1.64 repeatedly crashes[/+] on Nov 21, 2026
  3. couleurs256 commented on Nov 22, 2026

    @couleurs256

    Same thing here, the engine loads the model and when loaded it unloads and restarts in an enldess loops.
    Using strata latest version and engine, windows 11, GUI version.

    Here is a log on the 96 version, the 128 does the same :

    [strata] 2026-11-21 19:58:47 engine started: C:\Users\BS\AppData\Local\Strata\components\shapes\hy4\engine\0.1.64-windows-x86_64-cuda-cuda13\strata-hy4.exe --serve --native C:\Users\BS\Documents\Strata\models\hy4-flash\Hy4-Flash-SQ-96.strata --engram C:\Users\BS\Documents\Strata\models\hy4-flash\Hy4-Flash-Engram-Q8.strata --max-context 65536 --vision C:\Users\BS\Documents\Strata\models\hy4-flash\Hy4-Flash-Vision.strata --vram-experts-gb 10.83 --ram-experts-gb 72.10 --draft dfly2
    [args] --vram-experts-gb 10.83: not used by this runner yet
    [args] --ram-experts-gb 72.10: not used by this runner yet
    [gpu] NVIDIA RTX 4500 Ada Generation sm_89 free 22.7 / 24.0 GiB
    [plan] ctx 65536 (init 16384) chunk 8192: buffers ~3.33 GB, KV+idx 0.34 GB (128K full 1.36 GB), kv_b scratch 0.03 GB, trunk 6.20 GB (embed in RAM) on a 25.8 GB card -> host mask 0; reserve 5.05 GB
    [load] manifest 0.1 s: 1433 trunk tensors
    [load trunk] 10% 0.1 s 4.50 GB/s
    [load trunk] 20% 0.2 s 6.35 GB/s
    [load trunk] 30% 0.3 s 7.17 GB/s
    [load trunk] 40% 0.3 s 7.71 GB/s
    [load trunk] 50% 0.4 s 8.04 GB/s
    [load trunk] 60% 0.4 s 8.53 GB/s
    [load trunk] 70% 0.5 s 8.87 GB/s
    [load trunk] 80% 0.6 s 9.00 GB/s
    [load trunk] 90% 0.6 s 8.99 GB/s
    [load trunk] 100% 0.7 s 9.24 GB/s
    [load] embed_tokens (0.38 GB) in pinned mapped host RAM -> trunk VRAM 6.20 GB
    [load] experts tiered: 0 static VRAM (0.00 GB) + 2609 adaptive VRAM slots (12.76 GB) = 12.76 GB VRAM, 20992 in host RAM computed by the CPU (82.93 GB), 0.2 s
    [load] done: trunk 6.20 GB in 0.7 s, experts 82.93 GB in 0.2 s
    [load] total 1.2 s
    [cpuexp] 20992 experts resident, 82.93 GB in 9.1 s (9.13 GB/s), 626 reads, mlock=no, isa=scalar, kernel=fast(avx2-vnni), act=q8
    [cpuexp] 10 pool threads (pinned, physical cores first), spin 300 us, sched=1, pf=2 rows
    [hybrid] 20992 experts (82.93 GB) in host RAM computed by 10 CPU threads
    [vcache] pre-populated 2609/2609 adaptive VRAM slots (12.76 GB, decay=0.965, max_swaps=4) in 2.69 s
    [hybrid] prefill streaming on (pinned bounce ring): 2 buffer(s) of 384 MiB, 205 groups for 41 layers (70.17 GB/pass, skipped 2609 vcache experts, cpu_frac=0.00, largest 1.71 GB), active for T >= 128 (gather < 4096)
    [tune] HY4_TUNE=834
    [cublas-pf] side-stream double-buffered dequant + cuBLAS enabled (2 x 32 MiB scratch, M >= 32)
    [runner] CUDA Graph capture/replay enabled (HY4_CUDA_GRAPH=1)
    [runner] kv cache 0.53 GB (init 16384 / ctx 65536, chunk 8192), device free 0.2 GiB
    [strata] 2026-11-21 19:59:07 engine started: C:\Users\BS\AppData\Local\Strata\components\shapes\hy4\engine\0.1.64-windows-x86_64-cuda-cuda13\strata-hy4.exe --serve --native C:\Users\BS\Documents\Strata\models\hy4-flash\Hy4-Flash-SQ-96.strata --engram C:\Users\BS\Documents\Strata\models\hy4-flash\Hy4-Flash-Engram-Q8.strata --max-context 65536 --vision C:\Users\BS\Documents\Strata\models\hy4-flash\Hy4-Flash-Vision.strata --vram-experts-gb 10.83 --ram-experts-gb 72.10 --draft dfly2
    [args] --vram-experts-gb 10.83: not used by this runner yet
    [args] --ram-experts-gb 72.10: not used by this runner yet
    [gpu] NVIDIA RTX 4500 Ada Generation sm_89 free 22.7 / 24.0 GiB
    [plan] ctx 65536 (init 16384) chunk 8192: buffers ~3.33 GB, KV+idx 0.34 GB (128K full 1.36 GB), kv_b scratch 0.03 GB, trunk 6.20 GB (embed in RAM) on a 25.8 GB card -> host mask 0; reserve 5.05 GB
    [load] manifest 0.1 s: 1433 trunk tensors
    [load trunk] 10% 0.2 s 4.33 GB/s
    [load trunk] 20% 0.2 s 6.23 GB/s
    [load trunk] 30% 0.3 s 7.21 GB/s
    [load trunk] 40% 0.3 s 7.93 GB/s
    [load trunk] 50% 0.4 s 8.11 GB/s
    [load trunk] 60% 0.4 s 8.58 GB/s
    [load trunk] 70% 0.5 s 8.76 GB/s
    [load trunk] 80% 0.6 s 9.07 GB/s
    [load trunk] 90% 0.6 s 9.23 GB/s
    [load trunk] 100% 0.7 s 9.28 GB/s
    [load] embed_tokens (0.38 GB) in pinned mapped host RAM -> trunk VRAM 6.20 GB
    [load] experts tiered: 0 static VRAM (0.00 GB) + 2609 adaptive VRAM slots (12.76 GB) = 12.76 GB VRAM, 20992 in host RAM computed by the CPU (82.93 GB), 0.2 s
    [load] done: trunk 6.20 GB in 0.7 s, experts 82.93 GB in 0.2 s
    [load] total 1.2 s
    [cpuexp] 20992 experts resident, 82.93 GB in 8.3 s (9.95 GB/s), 626 reads, mlock=no, isa=scalar, kernel=fast(avx2-vnni), act=q8
    [cpuexp] 10 pool threads (pinned, physical cores first), spin 300 us, sched=1, pf=2 rows
    [hybrid] 20992 experts (82.93 GB) in host RAM computed by 10 CPU threads
    [vcache] pre-populated 2609/2609 adaptive VRAM slots (12.76 GB, decay=0.965, max_swaps=4) in 2.74 s
    [hybrid] prefill streaming on (pinned bounce ring): 2 buffer(s) of 384 MiB, 205 groups for 41 layers (70.17 GB/pass, skipped 2609 vcache experts, cpu_frac=0.00, largest 1.71 GB), active for T >= 128 (gather < 4096)
    [tune] HY4_TUNE=834
    [cublas-pf] side-stream double-buffered dequant + cuBLAS enabled (2 x 32 MiB scratch, M >= 32)
    [runner] CUDA Graph capture/replay enabled (HY4_CUDA_GRAPH=1)
    [runner] kv cache 0.53 GB (init 16384 / ctx 65536, chunk 8192), device free 0.2 GiB
    [strata] 2026-11-21 19:59:24 engine started: C:\Users\BS\AppData\Local\Strata\components\shapes\hy4\engine\0.1.64-windows-x86_64-cuda-cuda13\strata-hy4.exe --serve --native C:\Users\BS\Documents\Strata\models\hy4-flash\Hy4-Flash-SQ-96.strata --engram C:\Users\BS\Documents\Strata\models\hy4-flash\Hy4-Flash-Engram-Q8.strata --max-context 65536 --vision C:\Users\BS\Documents\Strata\models\hy4-flash\Hy4-Flash-Vision.strata --vram-experts-gb 10.83 --ram-experts-gb 72.10 --draft dfly2
    [args] --vram-experts-gb 10.83: not used by this runner yet
    [args] --ram-experts-gb 72.10: not used by this runner yet
    [gpu] NVIDIA RTX 4500 Ada Generation sm_89 free 22.7 / 24.0 GiB
    [plan] ctx 65536 (init 16384) chunk 8192: buffers ~3.33 GB, KV+idx 0.34 GB (128K full 1.36 GB), kv_b scratch 0.03 GB, trunk 6.20 GB (embed in RAM) on a 25.8 GB card -> host mask 0; reserve 5.05 GB
    [load] manifest 0.1 s: 1433 trunk tensors
    [load trunk] 10% 0.1 s 4.60 GB/s
    [load trunk] 20% 0.2 s 6.15 GB/s
    [load trunk] 30% 0.3 s 6.88 GB/s
    [load trunk] 40% 0.3 s 7.70 GB/s
    [load trunk] 50% 0.4 s 8.05 GB/s
    [load trunk] 60% 0.4 s 8.45 GB/s
    [load trunk] 70% 0.5 s 8.71 GB/s
    [load trunk] 80% 0.6 s 8.96 GB/s
    [load trunk] 90% 0.6 s 9.04 GB/s
    [load trunk] 100% 0.7 s 9.38 GB/s
    [load] embed_tokens (0.38 GB) in pinned mapped host RAM -> trunk VRAM 6.20 GB
    [load] experts tiered: 0 static VRAM (0.00 GB) + 2609 adaptive VRAM slots (12.76 GB) = 12.76 GB VRAM, 20992 in host RAM computed by the CPU (82.93 GB), 0.2 s
    [load] done: trunk 6.20 GB in 0.7 s, experts 82.93 GB in 0.2 s
    [load] total 1.2 s
    [cpuexp] 20992 experts resident, 82.93 GB in 8.2 s (10.11 GB/s), 626 reads, mlock=no, isa=scalar, kernel=fast(avx2-vnni), act=q8
    [cpuexp] 10 pool threads (pinned, physical cores first), spin 300 us, sched=1, pf=2 rows
    [hybrid] 20992 experts (82.93 GB) in host RAM computed by 10 CPU threads
    [vcache] pre-populated 2609/2609 adaptive VRAM slots (12.76 GB, decay=0.965, max_swaps=4) in 2.57 s
    [hybrid] prefill streaming on (pinned bounce ring): 2 buffer(s) of 384 MiB, 205 groups for 41 layers (70.17 GB/pass, skipped 2609 vcache experts, cpu_frac=0.00, largest 1.71 GB), active for T >= 128 (gather < 4096)
    [tune] HY4_TUNE=834
    [cublas-pf] side-stream double-buffered dequant + cuBLAS enabled (2 x 32 MiB scratch, M >= 32)
    [runner] CUDA Graph capture/replay enabled (HY4_CUDA_GRAPH=1)
    [runner] kv cache 0.53 GB (init 16384 / ctx 65536, chunk 8192), device free 0.3 GiB

  4. user37708 commented on Nov 22, 2026

    @user37708

    I am having this same issue: model loading (reading from disk, filling memory) and then unloading again, repeating in loops. I am also seeing exit code 3221225501 in the logs. Had this issue on both 0.2.18 and 0.2.19, engine 0.1.64 and 0.1.65, using s96 version of the model.

    OS: Windows 11
    CPU: Intel Core i9-14900K
    RAM: 96GB (2x48) DDR5
    GPU_0 (slot pcie 5.0 x16): 3090Ti
    GPU_1 (slot pcie 4.0 x4, chipset): 3090Ti

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

    needs-formatTitle or template does not follow CONTRIBUTING.md yet

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions