Eval bug: Gemma4 fails with NotImplemented: map: filter-mapping not implemented

Name and Version

llama-cli --version ggml_cuda_init: found 1 ROCm devices (Total VRAM: 81920 MiB): Device 0: Radeon 8060S Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 81920 MiB version: 8683 (d0a6dfeb2) built with GNU 15.2.1 for Linux x86_64

Operating systems

Linux

GGML backends

HIP

Hardware

AMD 395+ AI (Strix Halo)

Models

https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF with Q4_K_M

Problem description & steps to reproduce

I'm testing gemma4 (https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF with Q4_K_M) with opencode. The model loads correctly. When I issue a prompt via opencode, I see the following message in the logs of llama-server:

[36203] srv operator(): got exception: {"error":{"code":500,"message":"NotImplemented: map: filter-mapping not implemented","type":"server_error"}}

The request fails with this.

I run strix-halo toolboxes in nightly with the image of today.

Serverstart:

#!/bin/sh
llama-server \
  --models-preset ./models.ini \
  --host 0.0.0.0 \
  --port 8080 \
  --models-max 3 \
  --cont-batching \
  --jinja \
  --metrics \
  --kv-unified

Modelfile:

[*]
slots = 4

slot-save-path = /cache/prompt-cache/

threads = 12
flash-attn = on
mlock = off
mmap = off
fit = off
warmup = off
batch-size = 4096
ubatch-size = 512
cache-type-k = q8_0
cache-type-v = q8_0
jinja = true
direct-io = on
cache-prompt = true
#cache-reuse = 256
#cache-ram = 8192
...

[reasoner2]
model = /home/XXX/llm/models/gemma4/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf
chat-template-file = /home/XXX/llm/models/gemma4/template2.jinja
alias = reasoner2
ctx-size = 131072
n-predict = -1
mlock = true
jinja = enabled
...

I did not found any other information on this so far, any ideas?

First Bad Commit

No response

Relevant log output

[36203] srv    operator(): got exception: {"error":{"code":500,"message":"NotImplemented: map: filter-mapping not implemented","type":"server_error"}}