| --- |
| license: mit |
| base_model: hongyuw/ft-bitvla-bitsiglipL-224px-libero_object-bf16 |
| tags: |
| - gguf |
| - robotics |
| - bitvla |
| --- |
| |
| # BitVLA LIBERO-Object I2S for Robot.cpp |
|
|
| Four GGUF components for Robot.cpp's BitVLA runtime. The source is the author's |
| [BitVLA checkpoint](https://huggingface.co/hongyuw/ft-bitvla-bitsiglipL-224px-libero_object-bf16/tree/7fd7a4f2cdb968a4d05759ff6ffb37667abc84f6), |
| revision `7fd7a4f2cdb968a4d05759ff6ffb37667abc84f6`. This is the fine-tuned version without robotics pre-training. |
| The normalization key is `libero_object_no_noops`. |
|
|
| ## Files and precision |
|
|
| - `bitvla-llm-i2s.gguf` |
| - `bitvla-vision-i2s.gguf` |
| - `bitvla-proprio-i2s.gguf` |
| - `bitvla-action-i2s.gguf` |
|
|
| The converter applies BitVLA WeightQuant in BF16, then packs 366 ternary matrices |
| as `i2s-matrix-v1`. Other tensors retain BF16/F32 storage. I2S is a Robot.cpp-specific |
| layout; these files require its patched gguf/ggml runtime, not a standalone llama.cpp CLI. |
| Use all four components from the same checkpoint. |
|
|
| ## Run |
|
|
| Use [Robot.cpp with BitVLA support](https://github.com/Robot-cpp/robot.cpp/tree/cf0712c2751fdd8519de22f822b7baa1b08b5937) and apply its llama.cpp patches before building. Use a model-server binary built from this revision or a later compatible revision. |
|
|
| ```bash |
| GGUF_DIR=/path/to/bitvla-libero-object-i2s |
| /path/to/model-server \ |
| --model-type bitvla \ |
| --llm "${GGUF_DIR}/bitvla-llm-i2s.gguf" \ |
| --mmproj "${GGUF_DIR}/bitvla-vision-i2s.gguf" \ |
| --state-proj "${GGUF_DIR}/bitvla-proprio-i2s.gguf" \ |
| --action-expert "${GGUF_DIR}/bitvla-action-i2s.gguf" \ |
| --threads 8 --host 127.0.0.1 --port 5555 |
| ``` |
|
|
| For these LIBERO checkpoints, send two 224×224 HWC RGB uint8 images in the `full_image`, then `wrist_image` order, |
| eight raw proprio values, and a nonempty task instruction. Resize/crop images |
| according to the source checkpoint's evaluation preprocessing before sending them. |
| The runtime normalizes images/state and returns an 8×7 denormalized action chunk. |
| Robot-specific action postprocessing remains the client's responsibility. |
|
|
| ## Validation |
|
|
| Release checks verify all four GGUF headers and SHA256 hashes. Uploaded file sizes |
| and SHA256 hashes are compared with the local conversion outputs. This model card |
| does not claim a new success-rate, numerical-equivalence, or latency result. |
|
|
