Instructions to use hexoy/gemma-4-e2b-distilled with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use hexoy/gemma-4-e2b-distilled with Transformers:

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

pipe = pipeline("image-text-to-text", model="hexoy/gemma-4-e2b-distilled", trust_remote_code=True)
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForImageTextToText
model = AutoModelForImageTextToText.from_pretrained("hexoy/gemma-4-e2b-distilled", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use hexoy/gemma-4-e2b-distilled with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "hexoy/gemma-4-e2b-distilled"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "hexoy/gemma-4-e2b-distilled",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/hexoy/gemma-4-e2b-distilled

SGLang

How to use hexoy/gemma-4-e2b-distilled 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 "hexoy/gemma-4-e2b-distilled" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "hexoy/gemma-4-e2b-distilled",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

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 "hexoy/gemma-4-e2b-distilled" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "hexoy/gemma-4-e2b-distilled",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use hexoy/gemma-4-e2b-distilled with Docker Model Runner:
```
docker model run hf.co/hexoy/gemma-4-e2b-distilled
```

hexoy commited on Jul 21

Commit

21509e0

verified ·

1 Parent(s): a63d660

Add experimental rank-8 LoRA comparison

Browse files

Files changed (1) hide show

README.md +21 -23

README.md CHANGED Viewed

@@ -11,13 +11,9 @@ tags:
 # Gemma 4 E2B Monarch 35-MLP
-Private experimental Gemma 4 E2B model with all 35 language-model MLPs
-replaced by two-factor Monarch linear maps. Attention, embeddings, and the
-multimodal components remain structurally unchanged.
-The Monarch factors were initialized by dense SVD projection and trained with
-activation alignment followed by full-model logit distillation. The released
-BF16 model has `3,682,268,704` parameters.
 ## 1. Weight Storage
@@ -26,29 +22,28 @@ BF16 model has `3,682,268,704` parameters.
 | Dense Gemma 4 BF16 | Reference | 9.507 GiB | 9.543 GiB | - |
 | Four-layer Monarch BF16 | Released | 9.122 GiB | 9.122 GiB | 4.05% |
 | 35-layer Monarch BF16 | Released | 6.859 GiB | 6.859 GiB | 27.86% |
 | 35-layer Monarch + 420 INT8 linears | Released | 6.135 GiB | 6.136 GiB | 35.48% |
-BF16 sizes use two bytes per parameter. Quantized sizes were measured from
-physical tensor storage. These values exclude activations, logits, CUDA
-workspaces, and allocator overhead.
 ## 2. TinyHellaSwag
-All rows used the same RTX PRO 6000, fixed batch size `32`, seed `1234`,
-100 official examples, 10-shot prompts, and no chat template.
 | Model | Status | GP-IRT accuracy | Raw accuracy | Runtime | Peak VRAM |
 | --- | --- | ---: | ---: | ---: | ---: |
 | Dense Gemma 4 BF16 | Reference | 39.23% | 29% | 28.50 s | 57.88 GiB |
 | Four-layer Monarch BF16 | Released | 37.04% | 27% | 22.96 s | 57.50 GiB |
 | 35-layer Monarch BF16 | Released | 32.35% | 22% | 23.33 s | 55.23 GiB |
 | 35-layer Monarch + INT8 linears | Released | 30.57% | 21% | 23.14 s | 54.51 GiB |
-TinyHellaSwag has only 100 examples, so small score changes are noisy. A good
-benchmark score does not override failed logit or generation checks.
-Historical auto-batch runtime and peak-VRAM measurements are omitted because
-selected batch sizes differed. The table above uses fixed batch size `32`.
 ## Usage
@@ -65,15 +60,18 @@ model = AutoModelForImageTextToText.from_pretrained(
 )
 ```
-This repository contains custom modeling code. Review it before enabling
-`trust_remote_code=True`.
 ## Reproducibility
-- Model weights: `f897353fca328b1cc5fd2e12d645773ca637f5f0`
-- Current source: [`gemma-distillation`](https://github.com/ratmir-miftachov/gemma-distillation)
-- Full private results and hashes:
-  [`gemma4-monarch-artifacts`](https://huggingface.co/datasets/hexoy/gemma4-monarch-artifacts/tree/main/benchmarks/int8/20260717T110806Z)
 Derived from [google/gemma-4-E2B-it](https://huggingface.co/google/gemma-4-E2B-it).
 See `NOTICE` for the modification summary.

 # Gemma 4 E2B Monarch 35-MLP
+Private experimental Gemma 4 E2B model with all 35 language-model MLPs replaced
+by two-factor Monarch maps. The released BF16 model has `3,682,268,704`
+parameters.
 ## 1. Weight Storage
 | Dense Gemma 4 BF16 | Reference | 9.507 GiB | 9.543 GiB | - |
 | Four-layer Monarch BF16 | Released | 9.122 GiB | 9.122 GiB | 4.05% |
 | 35-layer Monarch BF16 | Released | 6.859 GiB | 6.859 GiB | 27.86% |
+| 35-layer Monarch + LoRA r8 BF16 | Experimental | 6.876 GiB | 6.877 GiB | 27.68% |
 | 35-layer Monarch + 420 INT8 linears | Released | 6.135 GiB | 6.136 GiB | 35.48% |
+BF16 sizes use two bytes per parameter; serialized LoRA size was audited from
+the two safetensor shards. Values exclude activations and CUDA workspaces.
 ## 2. TinyHellaSwag
+All rows used the same RTX PRO 6000, fixed batch size `32`, seed `1234`, 100
+official examples, 10-shot prompts, and no chat template.
 | Model | Status | GP-IRT accuracy | Raw accuracy | Runtime | Peak VRAM |
 | --- | --- | ---: | ---: | ---: | ---: |
 | Dense Gemma 4 BF16 | Reference | 39.23% | 29% | 28.50 s | 57.88 GiB |
 | Four-layer Monarch BF16 | Released | 37.04% | 27% | 22.96 s | 57.50 GiB |
 | 35-layer Monarch BF16 | Released | 32.35% | 22% | 23.33 s | 55.23 GiB |
+| 35-layer Monarch + LoRA r8 BF16 | Experimental | 33.25% | 23% | 24.93 s | 55.25 GiB |
 | 35-layer Monarch + INT8 linears | Released | 30.57% | 21% | 23.14 s | 54.51 GiB |
+Rank-8 LoRA improved the 35-layer source by `0.90` GP-IRT points and one correct
+item, but missed the former `+1.0` point or `+2` item release gate. The
+100-example benchmark is noisy.
 ## Usage
 )
 ```
+The rank-8 result is published separately at
+[`hexoy/gemma-4-e2b-monarch-35mlp-lora-r8`](https://huggingface.co/hexoy/gemma-4-e2b-monarch-35mlp-lora-r8).
+It is a benchmark-recovery experiment: corrected 64-token generation repeated
+phrases for text and remained malformed or inaccurate for images. It is not a
+general text or multimodal recovery release.
 ## Reproducibility
+- BF16 model revision: `f897353fca328b1cc5fd2e12d645773ca637f5f0`
+- Source: [`gemma-distillation`](https://github.com/ratmir-miftachov/gemma-distillation)
+- Private LoRA release evidence:
+  [`gemma4-monarch-artifacts`](https://huggingface.co/datasets/hexoy/gemma4-monarch-artifacts/tree/main/experiments/lora-recovery-r8-release/20260721T134717Z)
 Derived from [google/gemma-4-E2B-it](https://huggingface.co/google/gemma-4-E2B-it).
 See `NOTICE` for the modification summary.