GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
akalakaaaa
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
akalakaaaa
Model Tree
unsloth/Qwen2.5-Coder-7B-Instruct-bnb-4bit
Input Modalities
Output Modalities
Supported Functionality
4858886f896bd05db823557fc5cbc4ac28342af7.6e2bdf60320db7f016104b41fd8b0d7762fb3c408f866137431653e7b2573479.q_proj, k_proj, v_proj, o_proj in all 28 layers.adapter_config.json to replace an author's local absolute base path with the public upstream ID and pinned revision.Response supervised fine tuning of EDA-accepted teacher responses and actual-error repair trajectories, rather than logits distillation. The source dataset contains 60 training records; four complete responses exceeded the 1,024-token training limit and were dropped, leaving 56 training records (32 generation, 22 simulation repair, 2 elaboration repair). Validation uses 13 records. These are synthetic project-generated DS2_ curriculum tasks; only these records entered this training run.
AdamW, learning rate 0.0001, weight decay 0.01, microbatch 1, gradient accumulation 4, seed 1808, gradient clipping norm 1.0. See training_config.json for recorded settings and data manifest hashes. Training data and the entire training pipeline are not included; this is a weights, inference and training-metadata release, not a complete retraining reproduction package.
Results below were audited on 2026-09-09, using Windows, NVIDIA CUDA, bitsandbytes NF4 and Vivado 2018.3. Denominators are fixed at 156 public development cases per model, including failed extraction and compile attempts.
| Model | Compile and elaborate | Simulation / full pass | Full pass rate |
|---|---|---|---|
| Unchanged NF4 base | 133/156 | 50/156 | 32.05% |
| Base + step-80 LoRA | 132/156 | 61/156 | 39.10% |
There were 18 gains, 7 regressions, net +11 cases (+7.05 percentage points). This checkpoint was explored using development probes; these are public benchmarks, not a hidden test or a generalization guarantee. Public benchmark material may also have appeared in the base model's pretraining. Generated RTL still requires independent tool validation.
evaluation/historical156.json contains paired per-case pass/failure stages and the original evaluation-manifest hash. It omits prompts and solution code. The historical truncated_outputs field reports eight non-EOS responses for each model; it does not independently prove that all eight reached the output limit. Reported generation and EDA stage sums exclude model loading, manual waits and full process overhead, so they are not end-to-end wall-clock time.
No new generation or EDA evaluation was run for publication. This release has not established AMD/ROCm, vLLM, Vivado 2025.2/2026.1 or official competition acceptance.
The byte-identical base used here is also mirrored at akalakaaaa/Qwen2.5-Coder-7B-Instruct-NF4-RTL-base-v1.
Install a suitable NVIDIA CUDA-enabled PyTorch first, then use the historical library pins in requirements-inference.txt in a separate environment. Download the base to ./base and this adapter to ./adapter:
bash
python adapter/inference.py --base ./base --adapter ./adapter --offline --max-new-tokens 512For the public upstream base and an already downloaded local adapter:
bash
python inference.py --base unsloth/Qwen2.5-Coder-7B-Instruct-bnb-4bit --revision 4858886f896bd05db823557fc5cbc4ac28342af7 --adapter . --max-new-tokens 512Do not apply this adapter to AWQ/GGUF weights or to the project's newer Qwen3.8-27B model. This distribution leaves the adapter separate from the quantized base and does not merge or requantize weights.
Apache-2.0. See LICENSE, NOTICE and provenance.json for Qwen/Unsloth attribution and packaging changes. Retain their attribution when redistributing.
python adapter/inference.py --base ./base --adapter ./adapter --offline --max-new-tokens 512python inference.py --base unsloth/Qwen2.5-Coder-7B-Instruct-bnb-4bit --revision 4858886f896bd05db823557fc5cbc4ac28342af7 --adapter . --max-new-tokens 512