File size: 3,258 Bytes
9ff8b2f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 | ---
license: other
license_name: minimax-license
license_link: https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE
base_model: MiniMaxAI/MiniMax-M2.7
base_model_relation: quantized
library_name: transformers
pipeline_tag: text-generation
language:
- en
- zh
- ru
tags:
- minimax
- minimax-m2
- moe
- mixture-of-experts
- bf16
- dequantized
---
# MiniMax-M2.7 — BF16 (dequantized from FP8)
Plain `bfloat16` weights of [MiniMaxAI/MiniMax-M2.7](https://huggingface.co/MiniMaxAI/MiniMax-M2.7),
reconstructed from the upstream block-FP8 (E4M3, 128×128 blocks) checkpoint
via shard-by-shard blockwise dequantization. **No calibration, no rounding loss
beyond the original FP8→BF16 cast** — every block is materialized exactly:
```
bf16_block = (fp8_block.float() * scale_fp32).bfloat16()
```
## Why this exists
`MiniMaxAI/MiniMax-M2.7` ships natively in FP8. On Ampere and earlier
(e.g. RTX A5000) FP8 tensor cores don't exist and inference engines have
to emulate FP8 through FP16 — paying double bandwidth without the speed
benefit. For further offline quantization (AWQ, GPTQ, RTN INT8, …) you
need plain BF16 weights anyway: `transformers + torch_dtype=bfloat16`
won't materialize the attention projections under the FP8 quant config,
which trips up `llmcompressor`'s GPTQ tracer.
This repo is the missing intermediate: **upstream MiniMax-M2.7 weights in
plain BF16 safetensors**, ready to be fed into any standard quantization
pipeline.
## Contents
- 47 shards `model-NNNNN-of-00047.safetensors`
- rebuilt `model.safetensors.index.json` (no `*.weight_scale_inv` entries)
- `config.json` with the upstream `quantization_config` stripped
- tokenizer + custom modeling `.py` files copied verbatim from the FP8 source
Total ≈ **458 GB**.
## Provenance
Produced on a single 48 GB GPU pod (~30 minutes wall time) using a
~150-line script — see
[`dequant_fp8_blockwise.py`](https://github.com/operationrange/zonatelecom-agent/blob/main/scripts/quant/dequant_fp8_blockwise.py).
Process per shard:
1. open `model-XXXXX-of-00130.safetensors` from the FP8 source
2. for each `*.weight` (FP8 e4m3fn): look up `*.weight_scale_inv` (FP32, 128×128)
3. broadcast scale to weight shape, multiply, cast to BF16
4. drop the scale tensor
5. write `model-NNNNN-of-00047.safetensors` (5 GB shards)
Other tensors (embeddings, layer norms, MoE routers/gates that were already
unquantized in the upstream config's `modules_to_not_convert`) are passed
through with a BF16 cast.
## Quick load
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained(
"operationrange/MiniMax-M2.7-BF16",
torch_dtype="bfloat16",
device_map="auto",
trust_remote_code=True,
)
tok = AutoTokenizer.from_pretrained("operationrange/MiniMax-M2.7-BF16", trust_remote_code=True)
```
Inference at full BF16 needs ≥ ~470 GB combined GPU+CPU memory, so this
checkpoint is mostly intended as a starting point for further compression
(AWQ-INT4, GPTQ-INT8, etc.) rather than direct serving.
## License
Inherits the [MiniMax-M2 license](https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE) from the
upstream model. No weights were modified — only the storage format.
|