hal0-brain-sft-ROCmFPX-GGUF

HAL0 BRAIN — ROCmFPX GGUF

Advanced reasoning · Tool calling · Platform management

HAL0 BRAIN is a mini-agent administrator trained on the hal0 system: its services, profiles, slots, model lifecycle, diagnostics, and tool workflows. This repository packages the verified F16 reference plus two agent-oriented ROCmFPX quants for structured reasoning, tool calling, and platform management.

Files

The Agent presets deliberately retain higher-precision tensors where structured behavior is most sensitive. The filenames are descriptive, but the format claims above come from ROCmFPX-aware GGUF inspection—not filename inference.

Provenance

• Base model: Hal0ai/hal0-brain-sft
• Verified F16 GGUF source: Hal0ai/hal0-brain-sft-GGUF at revision 6b190df6e816cc806f7fa7ae3de7248f5551e00b
• F16 source SHA-256: ed9d28c4eac1d7c291bc80d9410c243a3d28e655921ccaf90f2b6619aa24d2c3
• Quantizer: charlie12345/ROCmFPX at commit 61f2f2d7bc4955e9bca821095ef69125837133b5
• Architecture: Llama-compatible GGUF, approximately 1.08B parameters
• License: Apache-2.0

Both quantized files were produced independently from F16. Neither was requantized from another low-bit artifact.

Runtime compatibility

The Q4 and Q8 files contain custom GGML tensor types and require a ROCmFPX-capable runner. Current stock llama.cpp rejects the Q4 tensor type ID 100 and Q8 tensor type ID 103 as invalid/unknown. LM Studio, Ollama, or another stock llama.cpp package must not be assumed compatible.

hal0 pins the unified runner by immutable digest:

ghcr.io/hal0ai/hal0-rocmfpx@sha256:fd6b02a720e633e402e929e19eedefff52aeec18e5de8f43e525689e523985f3

That runner contains the MiniCPM5 pre-tokenizer mapping required by this model and supports both ROCm/HIP and Vulkan ROCmFPX paths.

Download

Install the Hugging Face CLI and authenticate only if your environment requires it:

pip install -U huggingface_hub

F16:

hf download Hal0ai/hal0-brain-sft-ROCmFPX-GGUF \
  hal0-brain-sft-F16.gguf --local-dir .

ROCmFP4 Coherent Agent:

hf download Hal0ai/hal0-brain-sft-ROCmFPX-GGUF \
  hal0-brain-sft-Q4_0_ROCMFP4_COHERENT.gguf --local-dir .

ROCmFP8 Agent:

hf download Hal0ai/hal0-brain-sft-ROCmFPX-GGUF \
  hal0-brain-sft-Q8_0_ROCMFPX_AGENT.gguf --local-dir .

Portable hal0 long-context profile:

hf download Hal0ai/hal0-brain-sft-ROCmFPX-GGUF \
  chat-long-context.hal0profile.json --local-dir .

Run with ROCmFPX

ROCm/HIP:

llama-cli \
  -m hal0-brain-sft-Q8_0_ROCMFPX_AGENT.gguf \
  -dev ROCm0 -ngl 999 -fa on -c 8192 -b 512 -ub 512 --jinja

Vulkan:

llama-cli \
  -m hal0-brain-sft-Q4_0_ROCMFP4_COHERENT.gguf \
  -dev Vulkan0 -ngl 999 -fa on -c 8192 -b 512 -ub 512 --jinja

OpenAI-compatible server:

llama-server \
  -m hal0-brain-sft-Q8_0_ROCMFPX_AGENT.gguf \
  --host 0.0.0.0 --port 8080 \
  -dev Vulkan0 -ngl 999 -fa on -c 8192 -b 512 -ub 512 --jinja

Choose ROCm0 or Vulkan0 according to the devices reported by the ROCmFPX runner. Long context and large batches consume substantial memory; tune them for the host.

Tool-calling contract

hal0's catalog declares the hal0-function-xml prompt contract. A complete tool call has this shape:

<function name="memory_add">
  <param name="text">Remember that the brain slot uses the ROCmFPX runner.</param>
</function>

The hal0 integration validates the tool schema, allows at most one complete tool call per turn, and stops after </function>.

Portable hal0 profile

chat-long-context.hal0profile.json is a checksummed, portable hal0.profile schema-1 envelope. It includes:

-fa on -ctk q8_0 -ctv q8_0 -b 2048 -ub 512 --parallel 1
--no-mmap --no-context-shift --poll 100 --poll-batch 1 --metrics --no-webui

These are performance-oriented defaults, not a hardware guarantee. Reduce context/batch settings or use another profile when the host cannot sustain them.

Dashboard import

Open Profiles, choose Import, select chat-long-context.hal0profile.json, review the dry-run result, and commit the import. If chat-long-context already exists, choose another name rather than overwriting it implicitly.

REST API dry-run and import

Dry-run validates the envelope and checksum without changing the profile catalog:

PROFILE=chat-long-context.hal0profile.json
jq -n --slurpfile envelope "$PROFILE" \
  '{envelope:$envelope[0],name:"chat-long-context",dry_run:true}' |
  curl --fail-with-body -sS http://127.0.0.1:8080/api/profiles/import \
    -H 'content-type: application/json' --data-binary @-

Commit only after the dry-run reports valid: true, checksum_ok: true, and no unwanted collision:

PROFILE=chat-long-context.hal0profile.json
jq -n --slurpfile envelope "$PROFILE" \
  '{envelope:$envelope[0],name:"chat-long-context",dry_run:false}' |
  curl --fail-with-body -sS http://127.0.0.1:8080/api/profiles/import \
    -H 'content-type: application/json' --data-binary @-

Validation record

• ROCmFPX-aware inspection confirmed every file type and tensor count in the table.
• ROCmFPX CPU reference execution loaded and benchmarked both custom artifacts.
• The catalog runner image includes the required MiniCPM5 tokenizer mapping and ROCmFPX ROCm/Vulkan kernels.
• Stock llama.cpp e8e6c7af2456fd50bb62f7a2bbd642e6fb14ae77 rejected Q4 type 100 and Q8 type 103 before tensor loading.

Related

• hal0
• ROCmFPX
• Hal0ai model collection