Skip to content

Linux Vulkan b10018-1.3.0: Gemma 4 31B output is corrupted; upstream llama.cpp works #61

Description

@hsjlyj

Summary

The official Linux x64 Vulkan release b10018-1.3.0 produces corrupted Gemma 4 31B output: control tokens followed by gibberish/repetition. The same GGUF produces correct output with a separate pre-existing upstream llama.cpp binary on the same host/GPU.

This also makes the bundled Gemma 4 MTP path unusable: with the HauhauCS target + provided draft-mtp head, draft acceptance was 0.00026 (4 accepted / 15318 generated).

Environment

  • Atomic release: llama-turboquant-linux-x64-vulkan.tar.gz, b10018-1.3.0 (build 10408, commit 5bc5c248d)
  • OS: Ubuntu 22.04.5 LTS x86_64
  • GPU: NVIDIA GeForce RTX 5090, driver 580.119.02
  • Vulkan device is detected and model fully offloads successfully.

Reproduction

Model:

llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF

File:

gemma-4-31B-it-uncensored-heretic-Q4_K_M.gguf

Atomic command (MTP disabled to isolate the base decode path):

atomic-llama-cli \
  -m gemma-4-31B-it-uncensored-heretic-Q4_K_M.gguf \
  -ngl 99 -ctk turbo4 -ctv turbo4 -fa on -c 4096 -t 14 \
  -n 32 -p '请只回答:测试成功' -st \
  --no-display-prompt --no-warmup -co off \
  --reasoning off --reasoning-format none

Actual result

<|channel>thought
<channel|>H lC/ or single ly single l orP wife/ deSまだまだ deH ly single l or single lL,C// or

The server API shows the same corruption. It also logs:

W load: control-looking token: 212 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
W load: control-looking token: 50 '<|tool_response>' was not control-type; this is probably a bug in the model. its type will be overridden
W load: special_eog_ids contains '<|tool_response>', removing '</s>' token from EOG list

I reproduced the same behavior with a second Gemma 4 31B target (HauhauCS/Gemma4-31B-QAT-Uncensored-HauhauCS-Balanced-MTP) both with and without its MTP drafter, and with the embedded template and --chat-template gemma.

Control result

On the same host, GPU and target GGUF, a pre-existing upstream llama-cli build (b1-876a432) returns the expected text:

<|channel>thought
<channel|>测试成功

It runs at about 45 tok/s in this small test. Therefore this does not appear to be a damaged download, a target-model issue, or an MTP-only issue.

Expected result

Atomic's Linux Vulkan release should generate coherent content from the same Gemma 4 GGUF, including when TurboQuant KV cache is enabled. If there is a known Gemma 4 incompatibility in this release/backend, please document it and consider disabling the affected configuration rather than returning silent/corrupted output.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

VulkanbugSomething isn't working

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions