Misc. bug: xmfp6 and turbo3 sharing same ID:42 this preventing from loading xmfp6 models

[ovadmani-sudo]

Name and Version

version: 9459 (07ac3ce)
built with Clang 21.1.8 for Linux x86_64

Thank you.
ovadmani

Operating systems

Linux

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

llama-server

Command line

#!/bin/bash                                                                                                                            
                                                                                                                                       
# Define the exact absolute path based on your verified folder structure                                                               
BUILD_DIR="/home/ovadm/beellama.cpp/rocm-build"                                                                                        
                                                                                                                                       
# 1. Map both local build libraries and global ROCm targets cleanly                                                                    
export LD_LIBRARY_PATH="$BUILD_DIR/bin:/opt/rocm/lib:/opt/rocm/lib64:/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH"                       
                                                                                                                                       
# 2. Hardware Architecture Overrides for Ryzen AI Max+ 395 (Strix Halo)                                                                
export HSA_OVERRIDE_GFX_VERSION=11.5.1                                                                                                 
export HCC_AMDGPU_TARGET=gfx1151                                                                                                       
export PYTORCH_ROCM_ARCH=gfx1151                                                                                                       
export GGML_HIP_NO_PINNED=0                                                                                                            
export AMD_DISABLE_GFXOFF=1                                                                                                            
# 3. Unified Memory and Power Constraints Optimization                                                                                 
export AMD_DISABLE_GFXOFF=1                                                                                                            
export HIP_FORCE_POINTER_MAPPING=1                                                                                                     
                                                                                                                                       
echo "Starting BeeLlama Server with DFlash speculation on gfx1151..."                                                                  
                                                                                                                                       
# 4. Invoke the binary using its absolute location path                                                                                
ARGS=(
   # --verbose
    #-m /home/ovadm/models/bartowski--google_gemma-4-31B-it-GGUF/google_gemma-4-31B-it-IQ4_NL.gguf
   # -m /home/ovadm/models/google--gemma-4-31B-it/gemma-4-31b-q6_k.gguf
    -m /home/ovadm/models/google--gemma-4-31B-it/gemma-4-31b-mxfp6.gguf
   --alias local_model
   #--mmproj /home/ovadm/models/mmproj-BF16.gguf
   -np 1   #second app
   --kv-unified  # not splitting context
   --spec-type dflash
   #--ctx-size-draft 8
   --spec-draft-model /home/ovadm/models/Anbeeld--gemma-4-31B-it-DFlash-GGUF/gemma4-31b-it-dflash-Q8_0.gguf
   #--spec-dflash-cross-ctx  1024
   #--spec-draft-ngl 99   #draft engine on cpu 
   #--spec-draft-threads 32
   #--spec-draft-threads-batch 32
   -ngl 99 #brain on gpu
   -t 32
   --draft-max 8
   --draft-min 0
   #--no-spec-dm-adaptive
   --spec-dflash-default
   --spec-draft-p-min 0.25
   --flash-attn on
   --cache-type-k q8_0
   --cache-type-v q8_0
   -b 2048
   -ub 2048
   -c  65536
   --port 9091
   --no-mmap
   --mlock
   --metrics
   --perf 
   --presence-penalty 0
   --temp 0.3
   --top-p 1.0
   --top-k 20
   --min-p 0
   -to 1800
   #--jinja 
   #--reasoning-format deepseek 
   #--chat-template-kwargs '{"enable_thinking":true}'
   --webui-mcp-proxy
)
sudo pkill -9 llama-server
sleep 1
"$BUILD_DIR/bin/llama-server" "${ARGS[@]}" 2>&1 | tee >(stdbuf -oL socat - UDP4-SENDTO:127.0.0.1:5001)  | grep --line-buffered -iE "error|failed|exception|critical|fatal"

**********************************************************************************

 ./gemma-31b-dflsh.sh 
Starting BeeLlama Server with DFlash speculation on gfx1151...
gguf_init_from_file: failed to open GGUF file '/home/ovadm/models/google--gemma-4-31B-it/gemma-4-31b-mxfp6.gguf' (No such file or directory)
llama_model_load: error loading model: llama_model_loader: failed to load model from /home/ovadm/models/google--gemma-4-31B-it/gemma-4-31b-mxfp6.gguf
llama_model_load_from_file_impl: failed to load model
common_fit_params: encountered an error while trying to fit params to free device memory: failed to load model
gguf_init_from_file: failed to open GGUF file '/home/ovadm/models/google--gemma-4-31B-it/gemma-4-31b-mxfp6.gguf' (No such file or directory)
llama_model_load: error loading model: llama_model_loader: failed to load model from /home/ovadm/models/google--gemma-4-31B-it/gemma-4-31b-mxfp6.gguf
llama_model_load_from_file_impl: failed to load model
common_init_from_params: failed to load model '/home/ovadm/models/google--gemma-4-31B-it/gemma-4-31b-mxfp6.gguf'
srv    load_model: failed to load model, '/home/ovadm/models/google--gemma-4-31B-it/gemma-4-31b-mxfp6.gguf'

Problem description & steps to reproduce

load any xmfp6 model (xfmp4 can only do moe)

First Bad Commit

No response

Relevant log output

Logs

[Anbeeld]

MXFP6 is not merged into llama.cpp master yet, so I'm deferring this until then.