Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -43,3 +43,17 @@ ModelOpt 0.44 max-calibration through a per-expert-fidelity pipeline (each route
|
|
| 43 |
M3 experts use the GPT-OSS-style clamped SwiGLU: `(clamp(up,±7)+1) · clamp(gate,max=7) · σ(1.702·gate)`. Serving stacks must implement this exactly — kernels that silently fall back to plain SiLU·up will degrade quality on all 57 MoE layers. Verify your MoE backend honors `swiglu_limit`/`alpha` (or probe it numerically at startup).
|
| 44 |
|
| 45 |
Quantized with [TensorRT Model Optimizer](https://github.com/NVIDIA/TensorRT-Model-Optimizer) via a streaming per-expert calibration pipeline. Base model by [MiniMax](https://huggingface.co/MiniMaxAI); MIT license per the base model.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
M3 experts use the GPT-OSS-style clamped SwiGLU: `(clamp(up,±7)+1) · clamp(gate,max=7) · σ(1.702·gate)`. Serving stacks must implement this exactly — kernels that silently fall back to plain SiLU·up will degrade quality on all 57 MoE layers. Verify your MoE backend honors `swiglu_limit`/`alpha` (or probe it numerically at startup).
|
| 44 |
|
| 45 |
Quantized with [TensorRT Model Optimizer](https://github.com/NVIDIA/TensorRT-Model-Optimizer) via a streaming per-expert calibration pipeline. Base model by [MiniMax](https://huggingface.co/MiniMaxAI); MIT license per the base model.
|
| 46 |
+
|
| 47 |
+
## Acknowledgements
|
| 48 |
+
|
| 49 |
+
- **[Luke Alonso](https://github.com/lukealonso)** — the streaming per-expert quant-toolkit calibration pipeline this conversion is built on, the **b12x** SM120 kernel stack used for serving, and the open-source calibration corpora (`deep_calib`, `diverse_calib`, `agentic_coding_calib*`) included here under `calibration/data/`.
|
| 50 |
+
- **[MiniMax](https://huggingface.co/MiniMaxAI)** — the base model, MiniMax-M3.
|
| 51 |
+
|
| 52 |
+
## Files in this repo
|
| 53 |
+
|
| 54 |
+
- weight shards + index, full VL sidecars (image/video processors, chat template, tokenizer), `hf_quant_config.json`
|
| 55 |
+
- `calibration/m3_merged_amax_gatefix.safetensors` — amax statistics that regenerate this export in ~25 min with **no recalibration**
|
| 56 |
+
- `calibration/data/*.jsonl` — Luke's calibration corpora (~225 MB), to recalibrate from scratch
|
| 57 |
+
- `msa_golden/` — SM100 kernel golden oracles (provenance for the SM120 MSA port)
|
| 58 |
+
|
| 59 |
+
Full reproduction recipe & code: <https://github.com/brandonmmusic-max/minimax-m3-nvfp4-recipe>
|