#2747 Qwen3.6A3B fails with out of memory with recommended configs while Qwen3.8-Flash-Next unsloth Q4 runs fine

本文

Hi,  
I have been able to run all provided models from iq3_s to unsloth-ud-q4_k_xl models with nearly 300K context window on my System of 64GB of system ram and RTX3080 10G pretty smoothly(iq3 runs around 40ish 50ish tokens per second while unsloth-ud-q4-k-xl runs around 10~12 tokens which is completely fine to me)

However today when I updated to t he latest version, and tried to run Qwen3.6-A3B with Q8, with 32K context size it completely fails to load and gives me this error : 
```
❯ ./setup.sh --vision cpu 
Strata - Qwen3.8-Flash-Next on a normal PC (a GPU + system RAM + CPU)

  1) qwen3.6-35b-a3b-q8_0
  2) qwen3.8-flash-next-unsloth-ud-q4_k_xl
  3) qwen3.8-flash-next-iq3_s
  4) qwen3.8-flash-next-iq3_s
  5) qwen3.8-flash-next-unsloth-ud-q4_k_xl
  6) qwen3.8-flash-next-unsloth-ud-q4_k_xl
  7) orcarouter-qwen3.8-flash-next-uncensored-iq3_xxs
  8) orcarouter-qwen3.8-flash-next-uncensored-iq3_xxs
  9) orcarouter-qwen3.8-flash-next-uncensored-iq3_xxs
  10) qwen3.8-flash-next-iq3_s
  11) qwen3.8-flash-next-iq3_s
  12) install another model / change settings
Which one? [1]: 12

=== Step 1: checking your PC ===
  [ok] GPU: NVIDIA GeForce RTX 3080, 10.0 GB VRAM, compute capability 8.6, driver 595.99.02
  [!]  less than 12 GB of VRAM: Strata will run, but most experts stay on the CPU and it will be slow
  [ok] RAM: 62 GB
  [ok] CPU: 12th Gen Intel(R) Core(TM) i7-12700K (AVX2)
  [ok] PCIe: 4.0 x16

=== Step 2: your choices ===
  1) Qwen3.8-Flash-Next   Qwen; GSQ-RCO quants by ISTA-DASLab - the original model
  2) Swift 1.5            UkisAI's fine-tune of Qwen3.8-Flash-Next - thinks much shorter (-63% thinking tokens, 1.8x sooner answers by its authors' numbers)
  3) Qwen3.8-Flash-Next Coder ISTA-DASLab's coding version - half the experts (code, tools, images kept): needs ~32 GB of RAM, faster; weaker outside coding
  4) Qwen3.8-Flash-Next (Unsloth) Unsloth's 4- to 8-bit quantizations - UD-IQ4_XS, UD-Q4_K_XL, UD-Q5_K_XL, UD-Q6_K_XL or Q8_0 (94 to 188 GB downloads): every expert in RAM from ~80 GB (UD-IQ4_XS) to ~136 GB (Q8_0) of RAM, else part of them read from the SSD
  5) Qwen3.8-Flash-Next (NVFP4) NVFP4 packs by sergqwer (strata-nvfp4) - 4-bit NVFP4 experts, NVIDIA cards; a 128 GB download, ~80 GB of RAM to keep every expert in RAM   [runs, but UD-IQ4_XS is faster on this card; 128 GB download, ~80 GB of RAM keeps every expert in RAM]
  6) Qwen3.6-35B-A3B      Qwen; GGUFs with the MTP layer by Unsloth - 35B mixture of experts with 3B active (256 experts, top-8): a smaller cousin of Qwen3.8-Flash-Next with the same tokenizer   [new]
Which model? [1]: 6

  Qwen3.6-35B-A3B - Qwen; GGUFs with the MTP layer by Unsloth: 35B mixture of experts with 3B active (256 experts, top-8): a smaller cousin of Qwen3.8-Flash-Next with the same tokenizer
  [ok] model: Qwen3.6-35B-A3B
  Its license: Apache 2.0: https://huggingface.co/Qwen/Qwen3.6-35B-A3B

  1) UD-Q4_K_XL 4-bit (Unsloth Dynamic), the smallest; download 22.9 GB
  2) UD-Q6_K_XL 6-bit (Unsloth Dynamic), closer to the full model; download 32.6 GB
  3) Q8_0       8-bit, the closest to the full model; download 37.8 GB
  Recommended for 62 GB of RAM: Q8_0 (the largest that fits; any size can be chosen)
Which size? [3]: 3
  [ok] size: Q8_0

  Context length = how much text the model can see at once. Its cache takes 1.4 GB of VRAM per 32K tokens, room the experts would use:
  1) 8K tokens
  2) 32K tokens   (recommended for your GPU)
  3) 64K tokens
  4) 128K tokens
  5) 200K tokens
  6) 256K tokens
Context? [2]: 2
  [ok] context: 32768 tokens
  [!]  images are not available with Qwen3.6-35B-A3B yet: off

=== Step 3: Python packages ===
  [ok] numpy, jinja2, regex, pyyaml, tqdm, requests, cmake, ninja, pillow, psutil already installed

=== Step 4: the Strata engine ===
  [ok] llama.cpp 3cf0325 (gguf-py, ggml, mtmd)
  [ok] engine already built for this PC
  [ok] engine: /media/hossein/SSD/strarta_stuff/Strata/engine/strata

=== Step 5: downloading Qwen3.6-35B-A3B Q8_0 ===
  [ok] Qwen3.6-35B-A3B-Q8_0.gguf already downloaded
  [ok] model file: /media/hossein/SSD/strarta_stuff/Strata-data/models/qwen36-Q8_0/Qwen3.6-35B-A3B-Q8_0.gguf

=== Step 6: preparing the model for Strata ===
  [ok] tokenizer: /media/hossein/SSD/strarta_stuff/Strata-data/packs/qwen36-q8_0/tokenizer

=== Step 7: writing the start script ===
  the earlier run config is kept as strata-qwen36-q8_0.json.bak
  [ok] start script: run-qwen36-q8_0.sh

All set.
  API (OpenAI):     http://127.0.0.1:8080/v1   (any API key; model name: anything)
  API (Anthropic):  http://127.0.0.1:8080/v1/messages
  Next time:        just run ./setup.sh (or run-qwen36-q8_0.sh) - it starts right away
  [ok] GPU: GPU 0 (NVIDIA GeForce RTX 3080, 10 GB)

  ----------------------------------------------------------------------------------------------------
  Starting qwen3.6-35b-a3b-q8_0: it loads about 38 GB into RAM and locks part of it for the GPU.
  While it does, YOUR PC CAN BE SLOW OR STOP RESPONDING FOR 1-3 MINUTES (longer the first time after a
  restart). That is normal: please wait and don't close this window - the browser opens when it is ready.
  Later, closing this window stops the model.
  ----------------------------------------------------------------------------------------------------
  Settings (strata-qwen36-q8_0.json): --max-context 32768 --prefill auto --mtp 2; server
    127.0.0.1:8080, gpu 0
loading the model (the first start takes a minute or two) ...
[strata] starting the engine: reading the model's weights ...
[strata] still starting (20 s) - please wait ...
[strata] still starting (40 s) - please wait ...
[strata] still starting (60 s) - please wait ...
[strata] still starting (80 s) - please wait ...
[strata] still starting (100 s) - please wait ...
[strata] still starting (120 s) - please wait ...
[strata] still starting (140 s) - please wait ...
[strata] still starting (160 s) - please wait ...
[strata] still starting (180 s) - please wait ...
[strata] still starting (200 s) - please wait ...
[strata] still starting (220 s) - please wait ...
[strata] still starting (240 s) - please wait ...
Traceback (most recent call last):
  File "/media/hossein/SSD/strarta_stuff/Strata/serve/server.py", line 6976, in <module>
    sys.exit(main())
             ^^^^^^
  File "/media/hossein/SSD/strarta_stuff/Strata/serve/server.py", line 6753, in main
    engine = StrataEngine(exe, engine_args(cfg) + (effort_end or []), cwd=cfg.get("cwd"), log=cfg.get("log"),
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/media/hossein/SSD/strarta_stuff/Strata/serve/server.py", line 889, in __init__
    raise EngineDied(why + (f" (see {log})" if log else "") + note +
EngineDied: the engine exited before it was ready (see /media/hossein/SSD/strarta_stuff/Strata/strata-qwen36-q8_0.log)
the engine process ended: exit code 1
the engine log's last lines:
  [strata] 2026-10-16 12:45:24 engine started: /media/hossein/SSD/strarta_stuff/Strata/engine/strata --serve --native /media/hossein/SSD/strarta_stuff/Strata-data/models/qwen36-Q8_0/Qwen3.6-35B-A3B-Q8_0.gguf --max-context 32768 --prefill auto --mtp 2
  strata-q36: /media/hossein/SSD/strarta_stuff/Strata-data/models/qwen36-Q8_0/Qwen3.6-35B-A3B-Q8_0.gguf: 753 tensors (layout checked), 4.40 GiB on the GPU in 1.1 s (the routed experts are placed after the engine's buffers)
  strata-q36: prompt experts on MMQ (int8)
  strata-q36: context 32768 cells, prompt chunk 2048, MTP block loaded, 1.28 GiB of VRAM free
  strata-q36: cudaMalloc(&E.stage[b], max_layer + kMmqTail) failed: out of memory (/media/hossein/SSD/strarta_stuff/Strata/src/qwen35/q36_engine.cu:2110)
```

I usually run Qwen3.6-A3B with Q8 and 262K context window using llama.cpp without any issues, so to get an out of memory issue when the much larger models fit seems like a bug to me. 

Am I doing something wrong or is this a bug?