Misc. bug: HIP WIN11 loading Qwen3.6-35B-A3B-UD-Q6_K.gguf around 120k context crashes server and even whole system

[vevi33]

Name and Version

b8850 llama.cpp server
RX 7800XT (latest driver)
R7 7700x
32gb DDR5 6000 MHz RAM

Operating systems

Windows

Which llama.cpp modules do you know to be affected?

llama-server

Command line

-m "unsloth\Qwen\Qwen3.6-35B-A3B-UD-Q6_K.gguf"
      --flash-attn on
      --ctx-size 120000
      --fit on
      --fit-target 384
      --threads 16
      --parallel 1
      --no-mmap
      --mlock
      --cache-ram 2048
      --ctx-checkpoints 12
      --temp 0.75
      --repeat-penalty 1.08
      --repeat-last-n 384
      --min-p 0.05
      --top-k 30
      --alias Qwen3.6-35B_L
      --chat-template-kwargs "{\"preserve_thinking\":true}"
      --reasoning on

Problem description & steps to reproduce

Something really odd happens with the b8850 and this quant. When I load it up with 100k context I have 15.5 gb VRAM usage and around 16GB ram usage. But when I increase the context to 120k my whole ram is filled up (32gb) and even page file, even had a whole system freeze because of it. It never happened with any previous models or in older versions.

The crash is 0xc0000005 = Windows STATUS_ACCESS_VIOLATION

First Bad Commit

No response

Relevant log output

sched_reserve: reserving ...
sched_reserve: resolving fused Gated Delta Net support:
sched_reserve: fused Gated Delta Net (autoregressive) enabled
sched_reserve: fused Gated Delta Net (chunked) enabled
sched_reserve: ROCm0 compute buffer size = 659.77 MiB
sched_reserve: ROCm_Host compute buffer size = 242.52 MiB
sched_reserve: graph nodes = 3729
sched_reserve: graph splits = 74 (with bs=512), 52 (with bs=1)
sched_reserve: reserve took 17654.00 ms, sched copies = 1
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
[WARN] <Qwen3.6-35B_L> ExitError >> exit status 0xc0000005, exit code: 3221225477
[INFO] <Qwen3.6-35B_L> process exited but not StateStopping, current state: starting
[WARN] metrics skipped, HTTP status=502, path=/v1/chat/completions

[vevi33]

VULKAN:

load_tensors: loading model tensors, this can take a while... (mmap = false, direct_io = false)
llama.cpp/llama.cpp/src/llama-mmap.cpp:736: GGML_ASSERT(addr) failed
[WARN] <Qwen3.6-35B_L> ExitError >> exit status 0xc0000409, exit code: 3221226505
[INFO] <Qwen3.6-35B_L> process exited but not StateStopping, current state: starting
[WARN] metrics skipped, HTTP status=502, path=/v1/chat/completions
[INFO] Request 127.0.0.1 "POST /v1/chat/completions HTTP/1.1" 502 78 "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/14

At least it fails during load and it doesn't fill up the whole ram so not risking a system hang.
Q5_XL works perfectly with much larger context, but even in a few days ago (not sure which version, sorry) Q6_XL worked well with 100k context. Now the Q6 crashes with 100k.

[vevi33]

Removing --fit-target 512 or any low values solves it. I just wanted to make sure that it fills up my VRAM, so the issue is related to this flag.

[chejh-amd]

Good find narrowing it to --fit-target. There's a known integer overflow issue with --fit-target on Windows (#20308) — that one specifically affects values above 4095 MiB, so yours at 384 might be hitting a different edge case in the same fitting logic.

Since removing the low --fit-target fixes it, a safer approach if you still want to maximize VRAM utilization is to leave --fit on (default) without specifying --fit-target, or use a higher target like 1024 (the default).