Instructions to use Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4")

# Load model directly
from transformers import AutoTokenizer, LlamaForCausalLMEagle3

tokenizer = AutoTokenizer.from_pretrained("Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4")
model = LlamaForCausalLMEagle3.from_pretrained("Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4

SGLang

How to use Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4 with Docker Model Runner:
```
docker model run hf.co/Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4
```

ZixiQi commited on 9 days ago

Commit

e6b2bbf

verified ·

1 Parent(s): 1cc6023

Simplify model card

Browse files

Files changed (1) hide show

README.md +2 -63

README.md CHANGED Viewed

@@ -9,71 +9,10 @@ tags:
   - draft-model
   - gqa
   - vllm
-  - torchspec
-  - minimax
   - nvfp4
   - quantized
 ---
-# MiniMax-M3-EAGLE3-GQA-NVFP4 (W4A4)
-**NVFP4 W4A4** quantization of [`Inferact/MiniMax-M3-EAGLE3-GQA`](https://huggingface.co/Inferact/MiniMax-M3-EAGLE3-GQA) —
-the 1-layer **grouped-query-attention (GQA)** EAGLE3 speculative-decoding draft for
-[`MiniMaxAI/MiniMax-M3`](https://huggingface.co/MiniMaxAI/MiniMax-M3), served with
-**[vLLM](https://github.com/vllm-project/vllm)** and trained with **[TorchSpec](https://github.com/lightseekorg/TorchSpec)**.
-Only the draft's **MLP** (`layers.0.mlp.{gate,up,down}_proj`, `intermediate_size=18432`) is quantized —
-**both weights and activations to FP4** (NVFP4, group size 16). Everything else (`embed_tokens`,
-`lm_head`, `fc`, `self_attn`, all norms) stays bf16. Produced with NVIDIA TensorRT Model Optimizer
-(`quant_method: modelopt`, `quant_algo: NVFP4`); W4A4 engages the Blackwell FP4 tensor cores.
-The base draft is a 1-layer dense Llama (`LlamaForCausalLMEagle3`) on MiniMax-M3's `hidden_size=6144` /
-`vocab_size=200064`; at serve time it shares the target's embedding and LM head (EAGLE3). See the
-[base model card](https://huggingface.co/Inferact/MiniMax-M3-EAGLE3-GQA) for training data and bf16
-acceptance numbers, and `config.json` here for the full architecture.
-## Why MLP-only
-Profiling the draft's decode path shows the MLP is the bulk of the draft's *compressible* compute; the
-`lm_head` (full 200k vocab) and `self_attn` are the larger remaining bf16 ops. Quantizing the MLP
-weights+activations to FP4 cuts its weight bandwidth ~4× and runs the GEMM on the FP4 tensor cores, while
-keeping the output-distribution-sensitive `lm_head`/`fc`/attention in bf16 to preserve acceptance.
-## Calibration
-The activation `input_scale` for each MLP linear was calibrated by capturing `max|input|` on the **bf16**
-draft while serving MiniMax-M3 (TP4 + EP, `num_speculative_tokens=3`, `--enforce-eager`) over a short
-MT-Bench drive, aggregated by max across TP ranks. Calibration used the `MiniMaxAI/MiniMax-M3-MXFP8`
-target (its hidden states are representative for capturing MLP-input amax). NVFP4 sets
-`input_scale = amax / (fp4_max·fp8_e4m3_max) = amax / 2688`.
-| MLP input | amax | input_scale |
-|---|---|---|
-| gate/up (post-attention-norm output) | 40.75 | 0.01515997 |
-| down (SiLU(gate)·up) | 119.50 | 0.04445685 |
-## Serving (vLLM)
-```bash
-vllm serve MiniMaxAI/MiniMax-M3 \
-  --speculative-config '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA-NVFP4","num_speculative_tokens":3}'
-```
-> Serving a **quantized** EAGLE3 draft requires a vLLM whose eagle3 first layer constructs its `qkv_proj`
-> under the real `self_attn.qkv_proj` module path, so the quant config's exclude-matching keeps the (bf16)
-> attention out of quantization. Otherwise the excluded attention is wrongly quantized and the 2×hidden
-> fused-QKV weight fails to load.
-## Files
-| File | Notes |
-|---|---|
-| `model.safetensors` | FP4-packed MLP (`weight` U8, `weight_scale` F8_E4M3 per-16-group, `weight_scale_2` F32, `input_scale` F32); bf16 elsewhere |
-| `config.json` | Original EAGLE3 config + `quantization_config` (`quant_algo: NVFP4`, group 16) |
-| `hf_quant_config.json` | ModelOpt quant metadata; `exclude_modules: [embed_tokens, fc, self_attn, lm_head]` |
-## Status
-Acceptance-rate and decode-latency evaluation vs the bf16 draft is **pending**. Prior W4A4 MLP
-quantization of sibling MiniMax-M3 EAGLE3 GQA drafts was quality-neutral (acceptance / accuracy within
-run-to-run noise), so no quality regression is expected.

   - draft-model
   - gqa
   - vllm
   - nvfp4
   - quantized
 ---
+# MiniMax-M3-EAGLE3-GQA-NVFP4
+W4A4 NVFP4 MLP-quantized version of [`Inferact/MiniMax-M3-EAGLE3-GQA`](https://huggingface.co/Inferact/MiniMax-M3-EAGLE3-GQA).