Cilla — DeepSeek on Radeon PRO V620

An experimental V620-focused source fork of Project Maya, descended from Strata. Named after Cilla, the boss in New World's Lazarus Instrumentality expedition. Independent community software; not affiliated with Amazon Games, AMD, DeepSeek, or upstream maintainers. No game artwork is redistributed.

The focus is DeepSeek-V4.1-Flash-1017 on Linux/HIP (gfx1030): multi-GPU layer splitting, HIP MMQ prompt experts, conversation slots and prefix-checkpoint reuse. The executable keeps its upstream name, maya-ds4.

Experimental source release, not a turnkey appliance. The exported engine revision was used on a four-V620 system. That does not qualify every motherboard, ROCm version, pack, context length or optional execution path. No performance guarantee is made. Read Cilla notes and limitations.

Included

Engine, packer, HTTP server, tests, and upstream documentation; mixed-quant pack overrides, Engram support, and the pinned seed-checkpoint fix. The Cilla branch retains public Maya ancestry but excludes private deployment commits, machine configuration, logs, weights, and credentials. Upstream licenses and credits remain intact.

The upstream README describes broader Maya functionality and upstream measurements, not qualification of this fork. Provenance identifies the snapshot and dependency pin.

Requirements

• Linux x86-64, an AMD kernel driver, /dev/kfd and render-node access.
• A coherent ROCm/HIP SDK with gfx1030 rocBLAS libraries. The original build used ROCm 7.2.4; other versions need verification.
• CMake >=3.24, C++20 compiler, Git, Ninja or Make. Python 3.11 is the tested release environment; other Python versions are not qualified by this release.
• Separately obtained, compatible model weights and enough host RAM/VRAM for the exact pack. One 32-GiB V620 does not by itself hold this large model. Pinned Engram plus copied CPU experts can consume hundreds of GB of host RAM.

Do not copy another machine's kernel, P2P, NUMA or swap changes blindly.

Build

Start with a clean checkout of the Cilla branch:

git clone --branch release/cilla-v620 https://github.com/crispyging/cilla-v620.git cilla
cd cilla

Create the tested Python environment, then fetch llama.cpp revision 3cf03257f219afbe7334045ff7c6a06ac68c627d, the exact source pin required by this engine and its MMQ patch. These commands download dependency source, not model weights, and assume third_party/llama.cpp does not already exist:

python3.11 -m venv .venv
.venv/bin/python -m pip install numpy jinja2 regex pyyaml tqdm psutil 'jsonschema>=4.18,<5'
git init third_party/llama.cpp
git -C third_party/llama.cpp remote add origin https://github.com/ggml-org/llama.cpp.git
git -C third_party/llama.cpp fetch --depth=1 origin 3cf03257f219afbe7334045ff7c6a06ac68c627d
git -C third_party/llama.cpp checkout --detach FETCH_HEAD
git -C third_party/llama.cpp rev-parse HEAD

The final command must print that exact revision. Use the matching ROCm development headers, HIP compiler/runtime, device libraries, hipBLAS and rocBLAS packages, including gfx1030 rocBLAS assets. Ensure your account can access /dev/kfd and the required /dev/dri/renderD* nodes (usually through the host's render/video groups). These instructions do not install drivers or change permissions for you.

Set ROCM_PATH to your actual SDK. Configure a new build directory:

export ROCM_PATH=/opt/rocm
export HIP_PATH="$ROCM_PATH"
export HIP_PLATFORM=amd
export LD_LIBRARY_PATH="$ROCM_PATH/lib:$ROCM_PATH/lib64:$ROCM_PATH/llvm/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
cmake -S . -B build-v620 \
  -DCMAKE_BUILD_TYPE=Release \
  -DSTRATA_ENABLE_HIP=ON -DSTRATA_ENABLE_CUDA=OFF \
  -DSTRATA_PREFILL_MMQ=ON -DSTRATA_NATIVE_EXPERTS=ON \
  -DCMAKE_HIP_ARCHITECTURES=gfx1030 \
  -DCMAKE_HIP_COMPILER="$ROCM_PATH/llvm/bin/clang++" \
  -DCMAKE_PREFIX_PATH="$ROCM_PATH" \
  -DSTRATA_GGML_DIR="$PWD/third_party/llama.cpp" \
  -DSTRATA_BUILD_TESTS=OFF -DSTRATA_WERROR=OFF \
  -DCMAKE_EXE_LINKER_FLAGS=-rdynamic
cmake --build build-v620 --target maya-ds4 -j 4

Use modest parallelism initially. CMake applies MMQ fused-SwiGLU and byte-stride changes to a build-local copy, not the downloaded dependency. Check for Cilla: MMQ byte-stride edit applied; patch failure or missing rewrite seams now stop configuration rather than silently falling back. No private patch script is needed.

Pack and serve

Copyable launch commands: Launching Cilla covers the HTTP server, a separate opt-in four-V620 reference config, the engine-only protocol command, and eight-card adaptation caveats. It does not require any private deployment launcher.

The command below is an example for a complete, supported deepseek41 GGUF set, not a guaranteed recipe for arbitrary GGUFs. A versioned family example is Maya-S at revision 758b04c62d2572f4381d067b32d4686a1c4275b3: six DeepSeek-V4.1-Flash-1017-Maya-S-IQ2_XXS-...gguf shards. This pins artifact identity, not a fresh qualification of that checkpoint on your hardware. Keep all shards from the same revision; check their model license and memory requirements separately.

See tools/ds4/pack_gguf.py --help and upstream DeepSeek docs. A text-only pack without a drafter must explicitly opt out of sibling sidecar discovery:

.venv/bin/python tools/ds4/pack_gguf.py \
  --model /path/to/gguf-directory --out /path/to/new-pack \
  --drafter none --vision none

Not every third-party mixed GGUF works directly. Dense Q5_K/Q6_K may need conversion, and mixed experts need valid per-tensor overrides. tools/ds4/requant_dense_q8.py is an advanced conversion utility requiring a shard-index JSON, not an automatic end-to-end preparation pipeline. Borrowed drafters/vision towers need separate compatibility and output gates.

Copy examples/cilla.local.example.json to cilla.local.json. Replace executable, pack, tokenizer, working-directory and log paths with your own absolute paths, and set the GPU list. The example uses one GPU index, one conversation and 32K configured context, with pinned Engram disabled. It is a configuration starting point, not a claim that a pack fits one V620.

.venv/bin/python serve/server.py --engine strata --config cilla.local.json --host 127.0.0.1 --port 8099

Keep it on loopback. The inherited UI exposes management functions; API-key authentication alone does not protect the entire service. Use an authenticated tunnel/proxy for remote access, not direct internet exposure. Never commit your local config or keys.

Wait for the engine-ready log message before inference: /v1/models alone is not readiness. Test a deterministic answer, growing histories and prefix reuse before long-context or performance tests.

CPU-only checks

With Python 3.11 and the pinned llama.cpp source installed, install the declared test dependencies and put CMake/Ninja from the venv on PATH:

.venv/bin/python -m pip install -r requirements-test-cilla.txt
export PATH="$PWD/.venv/bin:$PATH"
.venv/bin/python tools/check_hip_compat.py
.venv/bin/python -m unittest discover -s serve -p 'test_*.py' -t .
.venv/bin/python -m unittest discover -s tools -p 'test_*.py'
.venv/bin/python -m unittest discover -s tools/ds4 -p 'test_*.py'

Hardware/model-gated tests may skip. These checks do not qualify HIP numerics or full-model reliability. See Cilla notes for optional four-GPU tuning and known experimental paths.

Credits and license

Project Maya by Peasant Smith and contributors; Strata by Niko1221 and contributors. Cilla builds on their work. See LICENSE, NOTICE, and third-party licenses. Source licensing does not grant rights to model weights, game artwork or trademarks. Remove secrets and private infrastructure details from issue reports.