gemma-3n-E2B-it-quantized.w8a8

251
2
2.0B
license:mit
by
RedHatAI
Image Model
OTHER
2B params
New
251 downloads
Early-stage
Try on Hugging FaceAdd to Compare
Edge AI:
Mobile
Laptop
Server
5GB+ RAM
Mobile
Laptop
Server
Quick Summary

Model Overview - Model Architecture: gemma-3n-E2B-it - Input: Audio-Vision-Text - Output: Text - Model Optimizations: - Weight quantization: INT8 - Activation q...

Device Compatibility

Mobile
4-6GB RAM
Laptop
16GB RAM
Server
GPU
Minimum Recommended
2GB+ RAM

Training Data Analysis

🟡 Average (4.3/10)

Researched training datasets used by gemma-3n-E2B-it-quantized.w8a8 with quality assessment

Specialized For

general
science
multilingual
reasoning

Training Datasets (3)

common crawl
🔴 2.5/10
general
science
Key Strengths
  • •Scale and Accessibility: At 9.5+ petabytes, Common Crawl provides unprecedented scale for training d...
  • •Diversity: The dataset captures billions of web pages across multiple domains and content types, ena...
  • •Comprehensive Coverage: Despite limitations, Common Crawl attempts to represent the broader web acro...
Considerations
  • •Biased Coverage: The crawling process prioritizes frequently linked domains, making content from dig...
  • •Large-Scale Problematic Content: Contains significant amounts of hate speech, pornography, violent c...
wikipedia
🟡 5/10
science
multilingual
Key Strengths
  • •High-Quality Content: Wikipedia articles are subject to community review, fact-checking, and citatio...
  • •Multilingual Coverage: Available in 300+ languages, enabling training of models that understand and ...
  • •Structured Knowledge: Articles follow consistent formatting with clear sections, allowing models to ...
Considerations
  • •Language Inequality: Low-resource language editions have significantly lower quality, fewer articles...
  • •Biased Coverage: Reflects biases in contributor demographics; topics related to Western culture and ...
arxiv
🟡 5.5/10
science
reasoning
Key Strengths
  • •Scientific Authority: Peer-reviewed content from established repository
  • •Domain-Specific: Specialized vocabulary and concepts
  • •Mathematical Content: Includes complex equations and notation
Considerations
  • •Specialized: Primarily technical and mathematical content
  • •English-Heavy: Predominantly English-language papers

Explore our comprehensive training dataset analysis

View All Datasets

Code Examples

Deploymentpythonvllm
from vllm.assets.image import ImageAsset
from vllm import LLM, SamplingParams

# prepare model
llm = LLM(
    model="RedHatAI/gemma-3n-E2B-it-quantized.w8a8",
    trust_remote_code=True,
    max_model_len=4096,
    max_num_seqs=2,
)

# prepare inputs
question = "What is the content of this image?"
inputs = {
    "prompt": f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n",
    "multi_modal_data": {
        "image": ImageAsset("cherry_blossom").pil_image.convert("RGB")
    },
}

# generate response
print("========== SAMPLE GENERATION ==============")
outputs = llm.generate(inputs, SamplingParams(temperature=0.2, max_tokens=64))
print(f"PROMPT  : {outputs[0].prompt}")
print(f"RESPONSE: {outputs[0].outputs[0].text}")
print("==========================================")
Deploymentpythonvllm
from vllm.assets.image import ImageAsset
from vllm import LLM, SamplingParams

# prepare model
llm = LLM(
    model="RedHatAI/gemma-3n-E2B-it-quantized.w8a8",
    trust_remote_code=True,
    max_model_len=4096,
    max_num_seqs=2,
)

# prepare inputs
question = "What is the content of this image?"
inputs = {
    "prompt": f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n",
    "multi_modal_data": {
        "image": ImageAsset("cherry_blossom").pil_image.convert("RGB")
    },
}

# generate response
print("========== SAMPLE GENERATION ==============")
outputs = llm.generate(inputs, SamplingParams(temperature=0.2, max_tokens=64))
print(f"PROMPT  : {outputs[0].prompt}")
print(f"RESPONSE: {outputs[0].outputs[0].text}")
print("==========================================")
Deploymentpythonvllm
from vllm.assets.image import ImageAsset
from vllm import LLM, SamplingParams

# prepare model
llm = LLM(
    model="RedHatAI/gemma-3n-E2B-it-quantized.w8a8",
    trust_remote_code=True,
    max_model_len=4096,
    max_num_seqs=2,
)

# prepare inputs
question = "What is the content of this image?"
inputs = {
    "prompt": f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n",
    "multi_modal_data": {
        "image": ImageAsset("cherry_blossom").pil_image.convert("RGB")
    },
}

# generate response
print("========== SAMPLE GENERATION ==============")
outputs = llm.generate(inputs, SamplingParams(temperature=0.2, max_tokens=64))
print(f"PROMPT  : {outputs[0].prompt}")
print(f"RESPONSE: {outputs[0].outputs[0].text}")
print("==========================================")
Deploymentpythonvllm
from vllm.assets.image import ImageAsset
from vllm import LLM, SamplingParams

# prepare model
llm = LLM(
    model="RedHatAI/gemma-3n-E2B-it-quantized.w8a8",
    trust_remote_code=True,
    max_model_len=4096,
    max_num_seqs=2,
)

# prepare inputs
question = "What is the content of this image?"
inputs = {
    "prompt": f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n",
    "multi_modal_data": {
        "image": ImageAsset("cherry_blossom").pil_image.convert("RGB")
    },
}

# generate response
print("========== SAMPLE GENERATION ==============")
outputs = llm.generate(inputs, SamplingParams(temperature=0.2, max_tokens=64))
print(f"PROMPT  : {outputs[0].prompt}")
print(f"RESPONSE: {outputs[0].outputs[0].text}")
print("==========================================")
Deploymentpythonvllm
from vllm.assets.image import ImageAsset
from vllm import LLM, SamplingParams

# prepare model
llm = LLM(
    model="RedHatAI/gemma-3n-E2B-it-quantized.w8a8",
    trust_remote_code=True,
    max_model_len=4096,
    max_num_seqs=2,
)

# prepare inputs
question = "What is the content of this image?"
inputs = {
    "prompt": f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n",
    "multi_modal_data": {
        "image": ImageAsset("cherry_blossom").pil_image.convert("RGB")
    },
}

# generate response
print("========== SAMPLE GENERATION ==============")
outputs = llm.generate(inputs, SamplingParams(temperature=0.2, max_tokens=64))
print(f"PROMPT  : {outputs[0].prompt}")
print(f"RESPONSE: {outputs[0].outputs[0].text}")
print("==========================================")
Deploymentpythonvllm
from vllm.assets.image import ImageAsset
from vllm import LLM, SamplingParams

# prepare model
llm = LLM(
    model="RedHatAI/gemma-3n-E2B-it-quantized.w8a8",
    trust_remote_code=True,
    max_model_len=4096,
    max_num_seqs=2,
)

# prepare inputs
question = "What is the content of this image?"
inputs = {
    "prompt": f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n",
    "multi_modal_data": {
        "image": ImageAsset("cherry_blossom").pil_image.convert("RGB")
    },
}

# generate response
print("========== SAMPLE GENERATION ==============")
outputs = llm.generate(inputs, SamplingParams(temperature=0.2, max_tokens=64))
print(f"PROMPT  : {outputs[0].prompt}")
print(f"RESPONSE: {outputs[0].outputs[0].text}")
print("==========================================")
Deploymentpythonvllm
from vllm.assets.image import ImageAsset
from vllm import LLM, SamplingParams

# prepare model
llm = LLM(
    model="RedHatAI/gemma-3n-E2B-it-quantized.w8a8",
    trust_remote_code=True,
    max_model_len=4096,
    max_num_seqs=2,
)

# prepare inputs
question = "What is the content of this image?"
inputs = {
    "prompt": f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n",
    "multi_modal_data": {
        "image": ImageAsset("cherry_blossom").pil_image.convert("RGB")
    },
}

# generate response
print("========== SAMPLE GENERATION ==============")
outputs = llm.generate(inputs, SamplingParams(temperature=0.2, max_tokens=64))
print(f"PROMPT  : {outputs[0].prompt}")
print(f"RESPONSE: {outputs[0].outputs[0].text}")
print("==========================================")
Deploymentpythonvllm
from vllm.assets.image import ImageAsset
from vllm import LLM, SamplingParams

# prepare model
llm = LLM(
    model="RedHatAI/gemma-3n-E2B-it-quantized.w8a8",
    trust_remote_code=True,
    max_model_len=4096,
    max_num_seqs=2,
)

# prepare inputs
question = "What is the content of this image?"
inputs = {
    "prompt": f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n",
    "multi_modal_data": {
        "image": ImageAsset("cherry_blossom").pil_image.convert("RGB")
    },
}

# generate response
print("========== SAMPLE GENERATION ==============")
outputs = llm.generate(inputs, SamplingParams(temperature=0.2, max_tokens=64))
print(f"PROMPT  : {outputs[0].prompt}")
print(f"RESPONSE: {outputs[0].outputs[0].text}")
print("==========================================")
Deploymentpythonvllm
from vllm.assets.image import ImageAsset
from vllm import LLM, SamplingParams

# prepare model
llm = LLM(
    model="RedHatAI/gemma-3n-E2B-it-quantized.w8a8",
    trust_remote_code=True,
    max_model_len=4096,
    max_num_seqs=2,
)

# prepare inputs
question = "What is the content of this image?"
inputs = {
    "prompt": f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n",
    "multi_modal_data": {
        "image": ImageAsset("cherry_blossom").pil_image.convert("RGB")
    },
}

# generate response
print("========== SAMPLE GENERATION ==============")
outputs = llm.generate(inputs, SamplingParams(temperature=0.2, max_tokens=64))
print(f"PROMPT  : {outputs[0].prompt}")
print(f"RESPONSE: {outputs[0].outputs[0].text}")
print("==========================================")
Deploymentpythonvllm
from vllm.assets.image import ImageAsset
from vllm import LLM, SamplingParams

# prepare model
llm = LLM(
    model="RedHatAI/gemma-3n-E2B-it-quantized.w8a8",
    trust_remote_code=True,
    max_model_len=4096,
    max_num_seqs=2,
)

# prepare inputs
question = "What is the content of this image?"
inputs = {
    "prompt": f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n",
    "multi_modal_data": {
        "image": ImageAsset("cherry_blossom").pil_image.convert("RGB")
    },
}

# generate response
print("========== SAMPLE GENERATION ==============")
outputs = llm.generate(inputs, SamplingParams(temperature=0.2, max_tokens=64))
print(f"PROMPT  : {outputs[0].prompt}")
print(f"RESPONSE: {outputs[0].outputs[0].text}")
print("==========================================")
Deploymentpythonvllm
from vllm.assets.image import ImageAsset
from vllm import LLM, SamplingParams

# prepare model
llm = LLM(
    model="RedHatAI/gemma-3n-E2B-it-quantized.w8a8",
    trust_remote_code=True,
    max_model_len=4096,
    max_num_seqs=2,
)

# prepare inputs
question = "What is the content of this image?"
inputs = {
    "prompt": f"<|user|>\n<|image_1|>\n{question}<|end|>\n<|assistant|>\n",
    "multi_modal_data": {
        "image": ImageAsset("cherry_blossom").pil_image.convert("RGB")
    },
}

# generate response
print("========== SAMPLE GENERATION ==============")
outputs = llm.generate(inputs, SamplingParams(temperature=0.2, max_tokens=64))
print(f"PROMPT  : {outputs[0].prompt}")
print(f"RESPONSE: {outputs[0].outputs[0].text}")
print("==========================================")
Creationpythontransformers
import requests
import torch
from PIL import Image
from transformers import AutoProcessor, Gemma3nForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.utils import dispatch_for_generation

# Load model.
model_id = "google/gemma-3n-E2B-it"
model = Gemma3nForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "flickr30k"
DATASET_SPLIT = {"calibration": "test[:512]"}
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048


# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {key: torch.tensor(value) for key, value in batch[0].items()}

dampening_frac=0.01

# Recipe
recipe = [
    GPTQModifier(
        targets="Linear",
        scheme="W8A8",
        ignore=[
            "re:.*embed_audio.*",
            "re:.*embed_vision.*",
            "re:.*audio_tower.*",
            "re:.*vision_tower.*",
            "re:.*altup.*",
            "re:.*lm_head.*",
            "re:.*laurel.*",
            "re:model\.language_model\.layers\.\d+\.per_layer_input_gate",
            "re:model\.language_model\.layers\.\d+\.per_layer_projection",
            "model.language_model.per_layer_model_projection",
        ],
        dampening_frac=dampening_frac
    ),
]

SAVE_DIR = f"{model_id.split('/')[1]}-quantized.{recipe[0].scheme}"

# Perform oneshot
oneshot(
    model=model,
    tokenizer=model_id,
    dataset=DATASET_ID,
    splits=DATASET_SPLIT,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
    data_collator=data_collator,
    # gemma3n has broken weight offloading which is required by the sequential pipeline
    pipeline="basic",
    # gemma3n does not support untying word embeddings
    tie_word_embeddings=True,
    output_dir=SAVE_DIR,
)

# Save to disk compressed.
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)
Creationpythontransformers
import requests
import torch
from PIL import Image
from transformers import AutoProcessor, Gemma3nForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.utils import dispatch_for_generation

# Load model.
model_id = "google/gemma-3n-E2B-it"
model = Gemma3nForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "flickr30k"
DATASET_SPLIT = {"calibration": "test[:512]"}
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048


# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {key: torch.tensor(value) for key, value in batch[0].items()}

dampening_frac=0.01

# Recipe
recipe = [
    GPTQModifier(
        targets="Linear",
        scheme="W8A8",
        ignore=[
            "re:.*embed_audio.*",
            "re:.*embed_vision.*",
            "re:.*audio_tower.*",
            "re:.*vision_tower.*",
            "re:.*altup.*",
            "re:.*lm_head.*",
            "re:.*laurel.*",
            "re:model\.language_model\.layers\.\d+\.per_layer_input_gate",
            "re:model\.language_model\.layers\.\d+\.per_layer_projection",
            "model.language_model.per_layer_model_projection",
        ],
        dampening_frac=dampening_frac
    ),
]

SAVE_DIR = f"{model_id.split('/')[1]}-quantized.{recipe[0].scheme}"

# Perform oneshot
oneshot(
    model=model,
    tokenizer=model_id,
    dataset=DATASET_ID,
    splits=DATASET_SPLIT,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
    data_collator=data_collator,
    # gemma3n has broken weight offloading which is required by the sequential pipeline
    pipeline="basic",
    # gemma3n does not support untying word embeddings
    tie_word_embeddings=True,
    output_dir=SAVE_DIR,
)

# Save to disk compressed.
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)
Creationpythontransformers
import requests
import torch
from PIL import Image
from transformers import AutoProcessor, Gemma3nForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.utils import dispatch_for_generation

# Load model.
model_id = "google/gemma-3n-E2B-it"
model = Gemma3nForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "flickr30k"
DATASET_SPLIT = {"calibration": "test[:512]"}
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048


# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {key: torch.tensor(value) for key, value in batch[0].items()}

dampening_frac=0.01

# Recipe
recipe = [
    GPTQModifier(
        targets="Linear",
        scheme="W8A8",
        ignore=[
            "re:.*embed_audio.*",
            "re:.*embed_vision.*",
            "re:.*audio_tower.*",
            "re:.*vision_tower.*",
            "re:.*altup.*",
            "re:.*lm_head.*",
            "re:.*laurel.*",
            "re:model\.language_model\.layers\.\d+\.per_layer_input_gate",
            "re:model\.language_model\.layers\.\d+\.per_layer_projection",
            "model.language_model.per_layer_model_projection",
        ],
        dampening_frac=dampening_frac
    ),
]

SAVE_DIR = f"{model_id.split('/')[1]}-quantized.{recipe[0].scheme}"

# Perform oneshot
oneshot(
    model=model,
    tokenizer=model_id,
    dataset=DATASET_ID,
    splits=DATASET_SPLIT,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
    data_collator=data_collator,
    # gemma3n has broken weight offloading which is required by the sequential pipeline
    pipeline="basic",
    # gemma3n does not support untying word embeddings
    tie_word_embeddings=True,
    output_dir=SAVE_DIR,
)

# Save to disk compressed.
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)
Creationpythontransformers
import requests
import torch
from PIL import Image
from transformers import AutoProcessor, Gemma3nForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.utils import dispatch_for_generation

# Load model.
model_id = "google/gemma-3n-E2B-it"
model = Gemma3nForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "flickr30k"
DATASET_SPLIT = {"calibration": "test[:512]"}
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048


# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {key: torch.tensor(value) for key, value in batch[0].items()}

dampening_frac=0.01

# Recipe
recipe = [
    GPTQModifier(
        targets="Linear",
        scheme="W8A8",
        ignore=[
            "re:.*embed_audio.*",
            "re:.*embed_vision.*",
            "re:.*audio_tower.*",
            "re:.*vision_tower.*",
            "re:.*altup.*",
            "re:.*lm_head.*",
            "re:.*laurel.*",
            "re:model\.language_model\.layers\.\d+\.per_layer_input_gate",
            "re:model\.language_model\.layers\.\d+\.per_layer_projection",
            "model.language_model.per_layer_model_projection",
        ],
        dampening_frac=dampening_frac
    ),
]

SAVE_DIR = f"{model_id.split('/')[1]}-quantized.{recipe[0].scheme}"

# Perform oneshot
oneshot(
    model=model,
    tokenizer=model_id,
    dataset=DATASET_ID,
    splits=DATASET_SPLIT,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
    data_collator=data_collator,
    # gemma3n has broken weight offloading which is required by the sequential pipeline
    pipeline="basic",
    # gemma3n does not support untying word embeddings
    tie_word_embeddings=True,
    output_dir=SAVE_DIR,
)

# Save to disk compressed.
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)
Creationpythontransformers
import requests
import torch
from PIL import Image
from transformers import AutoProcessor, Gemma3nForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.utils import dispatch_for_generation

# Load model.
model_id = "google/gemma-3n-E2B-it"
model = Gemma3nForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "flickr30k"
DATASET_SPLIT = {"calibration": "test[:512]"}
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048


# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {key: torch.tensor(value) for key, value in batch[0].items()}

dampening_frac=0.01

# Recipe
recipe = [
    GPTQModifier(
        targets="Linear",
        scheme="W8A8",
        ignore=[
            "re:.*embed_audio.*",
            "re:.*embed_vision.*",
            "re:.*audio_tower.*",
            "re:.*vision_tower.*",
            "re:.*altup.*",
            "re:.*lm_head.*",
            "re:.*laurel.*",
            "re:model\.language_model\.layers\.\d+\.per_layer_input_gate",
            "re:model\.language_model\.layers\.\d+\.per_layer_projection",
            "model.language_model.per_layer_model_projection",
        ],
        dampening_frac=dampening_frac
    ),
]

SAVE_DIR = f"{model_id.split('/')[1]}-quantized.{recipe[0].scheme}"

# Perform oneshot
oneshot(
    model=model,
    tokenizer=model_id,
    dataset=DATASET_ID,
    splits=DATASET_SPLIT,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
    data_collator=data_collator,
    # gemma3n has broken weight offloading which is required by the sequential pipeline
    pipeline="basic",
    # gemma3n does not support untying word embeddings
    tie_word_embeddings=True,
    output_dir=SAVE_DIR,
)

# Save to disk compressed.
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)
Creationpythontransformers
import requests
import torch
from PIL import Image
from transformers import AutoProcessor, Gemma3nForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.utils import dispatch_for_generation

# Load model.
model_id = "google/gemma-3n-E2B-it"
model = Gemma3nForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "flickr30k"
DATASET_SPLIT = {"calibration": "test[:512]"}
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048


# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {key: torch.tensor(value) for key, value in batch[0].items()}

dampening_frac=0.01

# Recipe
recipe = [
    GPTQModifier(
        targets="Linear",
        scheme="W8A8",
        ignore=[
            "re:.*embed_audio.*",
            "re:.*embed_vision.*",
            "re:.*audio_tower.*",
            "re:.*vision_tower.*",
            "re:.*altup.*",
            "re:.*lm_head.*",
            "re:.*laurel.*",
            "re:model\.language_model\.layers\.\d+\.per_layer_input_gate",
            "re:model\.language_model\.layers\.\d+\.per_layer_projection",
            "model.language_model.per_layer_model_projection",
        ],
        dampening_frac=dampening_frac
    ),
]

SAVE_DIR = f"{model_id.split('/')[1]}-quantized.{recipe[0].scheme}"

# Perform oneshot
oneshot(
    model=model,
    tokenizer=model_id,
    dataset=DATASET_ID,
    splits=DATASET_SPLIT,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
    data_collator=data_collator,
    # gemma3n has broken weight offloading which is required by the sequential pipeline
    pipeline="basic",
    # gemma3n does not support untying word embeddings
    tie_word_embeddings=True,
    output_dir=SAVE_DIR,
)

# Save to disk compressed.
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)
Creationpythontransformers
import requests
import torch
from PIL import Image
from transformers import AutoProcessor, Gemma3nForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.utils import dispatch_for_generation

# Load model.
model_id = "google/gemma-3n-E2B-it"
model = Gemma3nForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "flickr30k"
DATASET_SPLIT = {"calibration": "test[:512]"}
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048


# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {key: torch.tensor(value) for key, value in batch[0].items()}

dampening_frac=0.01

# Recipe
recipe = [
    GPTQModifier(
        targets="Linear",
        scheme="W8A8",
        ignore=[
            "re:.*embed_audio.*",
            "re:.*embed_vision.*",
            "re:.*audio_tower.*",
            "re:.*vision_tower.*",
            "re:.*altup.*",
            "re:.*lm_head.*",
            "re:.*laurel.*",
            "re:model\.language_model\.layers\.\d+\.per_layer_input_gate",
            "re:model\.language_model\.layers\.\d+\.per_layer_projection",
            "model.language_model.per_layer_model_projection",
        ],
        dampening_frac=dampening_frac
    ),
]

SAVE_DIR = f"{model_id.split('/')[1]}-quantized.{recipe[0].scheme}"

# Perform oneshot
oneshot(
    model=model,
    tokenizer=model_id,
    dataset=DATASET_ID,
    splits=DATASET_SPLIT,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
    data_collator=data_collator,
    # gemma3n has broken weight offloading which is required by the sequential pipeline
    pipeline="basic",
    # gemma3n does not support untying word embeddings
    tie_word_embeddings=True,
    output_dir=SAVE_DIR,
)

# Save to disk compressed.
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)
Creationpythontransformers
import requests
import torch
from PIL import Image
from transformers import AutoProcessor, Gemma3nForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.utils import dispatch_for_generation

# Load model.
model_id = "google/gemma-3n-E2B-it"
model = Gemma3nForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "flickr30k"
DATASET_SPLIT = {"calibration": "test[:512]"}
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048


# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {key: torch.tensor(value) for key, value in batch[0].items()}

dampening_frac=0.01

# Recipe
recipe = [
    GPTQModifier(
        targets="Linear",
        scheme="W8A8",
        ignore=[
            "re:.*embed_audio.*",
            "re:.*embed_vision.*",
            "re:.*audio_tower.*",
            "re:.*vision_tower.*",
            "re:.*altup.*",
            "re:.*lm_head.*",
            "re:.*laurel.*",
            "re:model\.language_model\.layers\.\d+\.per_layer_input_gate",
            "re:model\.language_model\.layers\.\d+\.per_layer_projection",
            "model.language_model.per_layer_model_projection",
        ],
        dampening_frac=dampening_frac
    ),
]

SAVE_DIR = f"{model_id.split('/')[1]}-quantized.{recipe[0].scheme}"

# Perform oneshot
oneshot(
    model=model,
    tokenizer=model_id,
    dataset=DATASET_ID,
    splits=DATASET_SPLIT,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
    data_collator=data_collator,
    # gemma3n has broken weight offloading which is required by the sequential pipeline
    pipeline="basic",
    # gemma3n does not support untying word embeddings
    tie_word_embeddings=True,
    output_dir=SAVE_DIR,
)

# Save to disk compressed.
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)
Creationpythontransformers
import requests
import torch
from PIL import Image
from transformers import AutoProcessor, Gemma3nForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.utils import dispatch_for_generation

# Load model.
model_id = "google/gemma-3n-E2B-it"
model = Gemma3nForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "flickr30k"
DATASET_SPLIT = {"calibration": "test[:512]"}
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048


# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {key: torch.tensor(value) for key, value in batch[0].items()}

dampening_frac=0.01

# Recipe
recipe = [
    GPTQModifier(
        targets="Linear",
        scheme="W8A8",
        ignore=[
            "re:.*embed_audio.*",
            "re:.*embed_vision.*",
            "re:.*audio_tower.*",
            "re:.*vision_tower.*",
            "re:.*altup.*",
            "re:.*lm_head.*",
            "re:.*laurel.*",
            "re:model\.language_model\.layers\.\d+\.per_layer_input_gate",
            "re:model\.language_model\.layers\.\d+\.per_layer_projection",
            "model.language_model.per_layer_model_projection",
        ],
        dampening_frac=dampening_frac
    ),
]

SAVE_DIR = f"{model_id.split('/')[1]}-quantized.{recipe[0].scheme}"

# Perform oneshot
oneshot(
    model=model,
    tokenizer=model_id,
    dataset=DATASET_ID,
    splits=DATASET_SPLIT,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
    data_collator=data_collator,
    # gemma3n has broken weight offloading which is required by the sequential pipeline
    pipeline="basic",
    # gemma3n does not support untying word embeddings
    tie_word_embeddings=True,
    output_dir=SAVE_DIR,
)

# Save to disk compressed.
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)
Creationpythontransformers
import requests
import torch
from PIL import Image
from transformers import AutoProcessor, Gemma3nForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.utils import dispatch_for_generation

# Load model.
model_id = "google/gemma-3n-E2B-it"
model = Gemma3nForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "flickr30k"
DATASET_SPLIT = {"calibration": "test[:512]"}
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048


# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {key: torch.tensor(value) for key, value in batch[0].items()}

dampening_frac=0.01

# Recipe
recipe = [
    GPTQModifier(
        targets="Linear",
        scheme="W8A8",
        ignore=[
            "re:.*embed_audio.*",
            "re:.*embed_vision.*",
            "re:.*audio_tower.*",
            "re:.*vision_tower.*",
            "re:.*altup.*",
            "re:.*lm_head.*",
            "re:.*laurel.*",
            "re:model\.language_model\.layers\.\d+\.per_layer_input_gate",
            "re:model\.language_model\.layers\.\d+\.per_layer_projection",
            "model.language_model.per_layer_model_projection",
        ],
        dampening_frac=dampening_frac
    ),
]

SAVE_DIR = f"{model_id.split('/')[1]}-quantized.{recipe[0].scheme}"

# Perform oneshot
oneshot(
    model=model,
    tokenizer=model_id,
    dataset=DATASET_ID,
    splits=DATASET_SPLIT,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
    data_collator=data_collator,
    # gemma3n has broken weight offloading which is required by the sequential pipeline
    pipeline="basic",
    # gemma3n does not support untying word embeddings
    tie_word_embeddings=True,
    output_dir=SAVE_DIR,
)

# Save to disk compressed.
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)
Creationpythontransformers
import requests
import torch
from PIL import Image
from transformers import AutoProcessor, Gemma3nForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.utils import dispatch_for_generation

# Load model.
model_id = "google/gemma-3n-E2B-it"
model = Gemma3nForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Oneshot arguments
DATASET_ID = "flickr30k"
DATASET_SPLIT = {"calibration": "test[:512]"}
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048


# Define a oneshot data collator for multimodal inputs.
def data_collator(batch):
    assert len(batch) == 1
    return {key: torch.tensor(value) for key, value in batch[0].items()}

dampening_frac=0.01

# Recipe
recipe = [
    GPTQModifier(
        targets="Linear",
        scheme="W8A8",
        ignore=[
            "re:.*embed_audio.*",
            "re:.*embed_vision.*",
            "re:.*audio_tower.*",
            "re:.*vision_tower.*",
            "re:.*altup.*",
            "re:.*lm_head.*",
            "re:.*laurel.*",
            "re:model\.language_model\.layers\.\d+\.per_layer_input_gate",
            "re:model\.language_model\.layers\.\d+\.per_layer_projection",
            "model.language_model.per_layer_model_projection",
        ],
        dampening_frac=dampening_frac
    ),
]

SAVE_DIR = f"{model_id.split('/')[1]}-quantized.{recipe[0].scheme}"

# Perform oneshot
oneshot(
    model=model,
    tokenizer=model_id,
    dataset=DATASET_ID,
    splits=DATASET_SPLIT,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    trust_remote_code_model=True,
    data_collator=data_collator,
    # gemma3n has broken weight offloading which is required by the sequential pipeline
    pipeline="basic",
    # gemma3n does not support untying word embeddings
    tie_word_embeddings=True,
    output_dir=SAVE_DIR,
)

# Save to disk compressed.
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)

Deploy This Model

Production-ready deployment in minutes

Together.ai

Instant API access to this model

Fastest API

Production-ready inference API. Start free, scale to millions.

Try Free API

Replicate

One-click model deployment

Easiest Setup

Run models in the cloud with simple API. No DevOps required.

Deploy Now

Disclosure: We may earn a commission from these partners. This helps keep LLMYourWay free.