Modelpublic

ModelCloud/Phi-tiny-MoE-instruct-GPTQ-W4-MixedGroup-G32-G128

sourceHugging Facemitupdated 2mo agoView on Hugging Face

0likes36downloads

Model Card

Phi-tiny-MoE-instruct-GPTQ-W4-MixedGroup-G32-G128

This is a GPTQ-quantized version of microsoft/Phi-tiny-MoE-instruct, created with GPTQModel 7.3.4. It is intended as a small mixed-group-size PhiMoE checkpoint and loader integration fixture.

What is special about this checkpoint?

The global configuration is symmetric W4G128. Exact per-module dynamic overrides use W4G32 for attention Q/K/V and every expert gate/up projection. This is mixed group-size quantization; all quantized weights use 4 bits.

Projection family	Modules per layer	Configuration	Modules in 32 layers
Attention Q/K/V	3	W4G32	96
Attention output	1	W4G128	32
16 experts x gate/up	32	W4G32	1,024
16 experts x down	16	W4G128	512
Total	52		1,664

Across the complete model:

—All 32 decoder layers and all 16 experts per layer are quantized.
—1,120 projections use W4G32 dynamic overrides.
—544 projections use the global W4G128 setting.
—The 32 routers and lm_head remain dense by design.
—qweight, qzeros, scales, and g_idx are present for all 1,664 quantized projections.

Quantization details

Setting	Value
Source dtype	BF16
Method / format	GPTQ / `gptq`
Bits	4 everywhere
Global group size	128
Dynamic group size	32
Symmetric	Yes
Activation order (`desc_act`)	No
Group-aware activation ordering	Yes
Sequential quantization	Yes
MSE search	2
Pack dtype / implementation	INT32 / GPU
Calibration samples	257
Calibration concatenation size	4,096 tokens
Observed attention calibration tokens	97,946
Checkpoint layout	One unsharded `model.safetensors` file
Quantization runtime	GPTQModel 7.3.4, Transformers 5.14.1, Torch 2.9.1+cu130

Every expert received a nonzero number of calibration activations during quantization.

Loading

python

from gptqmodel import BACKEND, GPTQModel

model = GPTQModel.load(
    "ModelCloud/Phi-tiny-MoE-instruct-GPTQ-W4-MixedGroup-G32-G128",
    backend=BACKEND.GPTQ_TORCH,
    device="cuda",
)

Validation reloads all 1,664 quantized projections with the expected group sizes and completes a finite-logit forward pass. No downstream quality benchmark is claimed; evaluate the checkpoint for the intended use case.