ggml_new_tensor_impl: not enough space in the scratch memory pool (needed 546644800, available 536870912) Segmentation fault

[vmajor]

Expected Behavior

This happens (so far) only with these models:
Wizard-Vicuna-30B-Uncensored.ggmlv3.q8_0.bin
WizardLM-30B-Uncensored.ggmlv3.q8_0.bin
based-30b.ggmlv3.q8_0.bin

Larger 65B models work fine. It could be something related to how these models are made, I will also reach out to @ehartford

llama-cpp-python 0.1.59 installed with OpenBLAS

CMAKE_ARGS="-DLLAMA_OPENBLAS=on" FORCE_CMAKE=1 pip install llama-cpp-python --force-reinstall --upgrade --no-cache-dir

I was running my usual code on the CPU and restarting it to tweak the results when this error came up. I made no code changes, only to context length, I reduced it as it was exceeding the 2048 token limit.

processed_output = self.llm(
            context + "\n### Instruction: \n" + instruction + "\n### Input: \n" + input_text + output,
            max_tokens=400,
            stop=None,
            temperature=0.7,
            repeat_penalty=1.1,
            top_k=80,
            top_p=0.5,
            echo=True,
        )

Current Behavior

llama.cpp: loading model from /home/****/models/Wizard-Vicuna-30B-Uncensored-GGML/Wizard-Vicuna-30B-Uncensored.ggmlv3.q8_0.bin
llama_model_load_internal: format     = ggjt v3 (latest)
llama_model_load_internal: n_vocab    = 32000
llama_model_load_internal: n_ctx      = 2048
llama_model_load_internal: n_embd     = 6656
llama_model_load_internal: n_mult     = 256
llama_model_load_internal: n_head     = 52
llama_model_load_internal: n_layer    = 60
llama_model_load_internal: n_rot      = 128
llama_model_load_internal: ftype      = 7 (mostly Q8_0)
llama_model_load_internal: n_ff       = 17920
llama_model_load_internal: n_parts    = 1
llama_model_load_internal: model size = 30B
llama_model_load_internal: ggml ctx size =    0.13 MB
llama_model_load_internal: mem required  = 35267.28 MB (+ 6248.00 MB per state)
.
llama_init_from_file: kv self size  = 6240.00 MB
AVX = 1 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 |
Processing all summaries...
ggml_new_tensor_impl: not enough space in the scratch memory pool (needed 546644800, available 536870912)
Segmentation fault

Environment and Context

wsl2
python 3.10.9

• Physical (or virtual) hardware you are using, e.g. for Linux:

$ lscpu
AMD Ryzen 9 3900XT 12-Core Processor

• Operating System, e.g. for Linux:

$ uname -a
5.15.68.1-microsoft-standard-WSL2+ #2 SMP

$ python3 3.10.9
$ make GNU Make 4.3
$ g++(Ubuntu 11.3.0-1ubuntu1~22.04.1) 11.3.0

To me it is 100% reproducible after several inference runs with Wizard-Vicuna-30B-Uncensored.ggmlv3.q8_0.bin

[ehartford]

Standing by!

[vmajor]

It seems to be a llama.cpp issue. I found it mentioned regarding starcoder models too. I think you can carry on :)

[vmajor]

Update. The issue resolves on reboot so there is some memory leak in the code.

This error would likely resolve if I restart wsl2, but it is messy for me to do that due to needing to remount my ext4 partitions, and I do not think that particular data point is as significant.

[eshaanagarwal]

Update. The issue resolves on reboot so there is some memory leak in the code.

This error would likely resolve if I restart wsl2, but it is messy for me to do that due to needing to remount my ext4 partitions, and I do not think that particular data point is as significant.

Hey but for a commercial application we can't afford to have it like this right ? This is happening with me on groovy 1.3 . Even that only on some OS like RHEL. It's really frustrating and difficult to deal with this

[gjmulder]

Hey but for a commercial application we can't afford to have it like this right ? This is happening with me on groovy 1.3 . Even that only on some OS like RHEL. It's really frustrating and difficult to deal with this

1. Facebook did not release their llama models for commercial application.
2. Did you pay for a license for any of the models or llama inference code you are using?

[eshaanagarwal]

Hi I am using gpt-4-all gpt-j 1.3 groovy which has Apache license.

[vmajor]

Another update, my guanaco-65B-GGML-q6_K.bin model just failed with the same error. So it is not just 30B models that are affected.

ggml_new_tensor_impl: not enough space in the scratch memory pool (needed 1143972864, available 1073741824)
Segmentation fault

[ibidani]

In hope to help isolate the bug I tried to reproduce the issue since version 0.1.55.
The first release that I experience the issue is 0.1.76(0.1.75 wasn't tested, isn't available on pypi) and didn't see it on 0.1.74
Could it be related to this change?
v0.1.74...v0.1.76#diff-9184e090a770a03ec97535fbef520d03252b635dafbed7fa99e59a5cca569fbcR200

Environment

python -V
Python 3.10.12
uname -a
Linux Idan-PC 5.15.90.1-microsoft-standard-WSL2 #1 SMP Fri Jan 27 02:56:13 UTC 2023 x86_64 x86_64 x86_64 GNU/Linux

Model: nous-hermes-13b.ggmlv3.q4_0.bin

[jbbqqf]

Hi @vmajor — drive-by triage note, not a maintainer.

The repro here uses .ggmlv3.q8_0.bin model files on llama-cpp-python 0.1.59. Both pieces are now obsolete:

• llama.cpp dropped the GGML format in favor of GGUF in August 2023 (upstream PR GGUF ggml-org/llama.cpp#2398). The current llama_cpp/llama.py only loads GGUF model files (see the "loading the first file of a sharded GGUF loads all remaining shard files" block).
• Current README only documents GGUF model paths and installs.
• Your own comment in this thread already pointed at llama.cpp as the source ("It seems to be a llama.cpp issue.")

Since the scratch-pool error path was internal to the old GGML code, it can't be reproduced against a current build. If you (or anyone subscribed) still hit something analogous with a GGUF model on a recent release (0.3.x), a fresh issue with the new repro would be much easier to action than this one.

Otherwise this is probably safe to close as obsolete.

---

Disclosure: I'm Jean-Baptiste Braun (data engineer, GCP/Python). This comment was drafted with the help of an AI coding assistant (Claude), then reviewed and edited by me. Posting from a small open-source backlog-triage pass — not affiliated with the project. Happy to delete if unwelcome.