Instructions to use praveenramesh/awq_finetuned_embedding_gemma with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use praveenramesh/awq_finetuned_embedding_gemma with sentence-transformers:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("praveenramesh/awq_finetuned_embedding_gemma")

sentences = [
    "What is the main advantage of using AWQ over existing work on language modeling and domain-specific benchmarks?",
    "Large language models (LLMs) have transformed numerous AI applications. On-device LLM is becoming increasingly important: running LLMs locally on edge devices can reduce the cloud computing cost and protect users' privacy. However, the astronomical model size and the limited hardware resource pose significant deployment challenges. We propose Activation-aware Weight Quantization (AWQ), a hardware-friendly approach for LLM low-bit weight-only quantization. AWQ finds that not all weights in an LLM are equally important. Protecting only 1% salient weights can greatly reduce quantization error. To identify salient weight channels, we should refer to the activation distribution, not weights. To avoid the hardware-inefficient mix-precision quantization, we mathematically derive that scaling up the salient channels can reduce the quantization error. AWQ employs an equivalent transformation to scale the salient weight channels to protect them. The scale is determined by collecting the activation statistics offline. AWQ does not rely on any backpropagation or reconstruction, so it generalizes to different domains and modalities without overfitting the calibration set. AWQ outperforms existing work on various language modeling and domain-specific benchmarks (coding and math). Thanks to better generalization, it achieves excellent quantization performance for instruction-tuned LMs and, for the first time, multi-modal LMs. Alongside AWQ, we implement TinyChat, an efficient and flexible inference framework tailored for 4-bit on-device LLM/VLMs. With kernel fusion and platform-aware weight packing, TinyChat offers more than 3 × speedup over the Huggingface FP16 implementation on both desktop and mobile GPUs. It also democratizes the deployment of the 70B Llama-2 model on mobile GPUs.",
    "We propose an alternative method to reduce the quantization error of the salient weight by per-channel scaling , which does not suffer from the hardware inefficiency issue.",
    "Quantization. We focus on weight-only grouped quantization in this work. As shown in previous work (Dettmers & Zettlemoyer, 2022; Frantar et al., 2022), grouped quantization is always helpful for improving performance/model size trade-off. We used a group size of 128 throughout the work, except otherwise specified. We focus on INT4/INT3 quantization since they are able to mostly preserve the LLMs' performance (Dettmers & Zettlemoyer, 2022). For AWQ, we used a small calibration set from the Pile (Gao et al., 2020) dataset in order not to overfit to a specific downstream domain. We used a grid size of 20 to search for the optimal α in Equation 5.. Models. We benchmarked our method on LLaMA (Touvron et al., 2023a) and OPT (Zhang et al., 2022) families. There are other open LLMs like BLOOM (Scao et al., 2022), but they are generally worse in quality, so we do not include them in our study. We further benchmark an instructiontuned model Vicuna (Chiang et al., 2023) and visual language models OpenFlamingo-9B (Awadalla et al., 2023) and LLaVA-13B (Liu et al., 2023a) to demonstrate the generability of our method.. Figure 5. Comparing INT3-g128 quantized Vicuna models with FP16 counterparts under GPT-4 evaluation protocol (Chiang et al., 2023). More winning cases (in blue) indicate better performance. AWQ consistently improves the quantized performance compared to RTN and GPTQ (Frantar et al., 2022), showing generalization to instruction-tuned models.. 0. 20. 40. 60. 80. 52. 75. 71. 5. 1. 3. 23. 4. 6. Quantized Win. Tie. Quantized Lost. 0. 20. 40. 60. 80. 47. 57. 57. 11. 6. 9. 22. 17. 14. I. N. T3/g128. RTN. GPTQ. AWQ. (a) Vicuna-7B. (b) Vicuna-13B. Evaluations. Following previous literature (Dettmers et al., 2022; Xiao et al., 2022; Frantar et al., 2022; Dettmers &Zettlemoyer, 2022; Yao et al., 2022), we mainly profiled the quantized models on language modeling tasks (perplexity evaluation on WikiText-2 (Merity et al., 2016)) since perplexity can stably reflect the LLM's performance (Dettmers &Zettlemoyer, 2022).. Baselines. Our primary baseline is vanilla round-tonearest quantization (RTN). It is actually quite strong when using a small group size like 128 (Frantar et al., 2022; Dettmers & Zettlemoyer, 2022). We also compare with a state-of-the-art method GPTQ (Frantar et al., 2022) for LLM weight quantization. For GPTQ, we also compare with an updated version that uses a 'reorder' trick (denoted as GPTQ-Reorder or GPTQ-R). Other techniques like ZeroQuant (Yao et al., 2022), AdaRound (Nagel et al., 2020), and BRECQ (Li et al., 2021) rely on backpropagation to update the quantized weights, which may not easily scale up to large model sizes; they also do not outperform GPTQ (Frantar et al., 2022), thus not included for study."
]
embeddings = model.encode(sentences)

similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [4, 4]

Notebooks
Google Colab
Kaggle

SentenceTransformer based on google/embeddinggemma-300m

This is a sentence-transformers model finetuned from google/embeddinggemma-300m. It maps sentences & paragraphs to a 768-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

Model Description

Model Type: Sentence Transformer
Base model: google/embeddinggemma-300m
Maximum Sequence Length: 2048 tokens
Output Dimensionality: 768 dimensions
Similarity Function: Cosine Similarity

Model Sources

Documentation: Sentence Transformers Documentation
Repository: Sentence Transformers on GitHub
Hugging Face: Sentence Transformers on Hugging Face

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 2048, 'do_lower_case': False, 'architecture': 'Gemma3TextModel'})
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
  (2): Dense({'in_features': 768, 'out_features': 3072, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity'})
  (3): Dense({'in_features': 3072, 'out_features': 768, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity'})
  (4): Normalize()
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("praveenramesh/awq_finetuned_embedding_gemma")
# Run inference
queries = [
    "How does the use of INT4/INT3 quantization compare to other quantization methods in terms of preserving the performance of large language models?",
]
documents = [
    "Quantization. We focus on weight-only grouped quantization in this work. As shown in previous work (Dettmers & Zettlemoyer, 2022; Frantar et al., 2022), grouped quantization is always helpful for improving performance/model size trade-off. We used a group size of 128 throughout the work, except otherwise specified. We focus on INT4/INT3 quantization since they are able to mostly preserve the LLMs' performance (Dettmers & Zettlemoyer, 2022). For AWQ, we used a small calibration set from the Pile (Gao et al., 2020) dataset in order not to overfit to a specific downstream domain. We used a grid size of 20 to search for the optimal α in Equation 5.. Models. We benchmarked our method on LLaMA (Touvron et al., 2023a) and OPT (Zhang et al., 2022) families. There are other open LLMs like BLOOM (Scao et al., 2022), but they are generally worse in quality, so we do not include them in our study. We further benchmark an instructiontuned model Vicuna (Chiang et al., 2023) and visual language models OpenFlamingo-9B (Awadalla et al., 2023) and LLaVA-13B (Liu et al., 2023a) to demonstrate the generability of our method.. Figure 5. Comparing INT3-g128 quantized Vicuna models with FP16 counterparts under GPT-4 evaluation protocol (Chiang et al., 2023). More winning cases (in blue) indicate better performance. AWQ consistently improves the quantized performance compared to RTN and GPTQ (Frantar et al., 2022), showing generalization to instruction-tuned models.. 0. 20. 40. 60. 80. 52. 75. 71. 5. 1. 3. 23. 4. 6. Quantized Win. Tie. Quantized Lost. 0. 20. 40. 60. 80. 47. 57. 57. 11. 6. 9. 22. 17. 14. I. N. T3/g128. RTN. GPTQ. AWQ. (a) Vicuna-7B. (b) Vicuna-13B. Evaluations. Following previous literature (Dettmers et al., 2022; Xiao et al., 2022; Frantar et al., 2022; Dettmers &Zettlemoyer, 2022; Yao et al., 2022), we mainly profiled the quantized models on language modeling tasks (perplexity evaluation on WikiText-2 (Merity et al., 2016)) since perplexity can stably reflect the LLM's performance (Dettmers &Zettlemoyer, 2022).. Baselines. Our primary baseline is vanilla round-tonearest quantization (RTN). It is actually quite strong when using a small group size like 128 (Frantar et al., 2022; Dettmers & Zettlemoyer, 2022). We also compare with a state-of-the-art method GPTQ (Frantar et al., 2022) for LLM weight quantization. For GPTQ, we also compare with an updated version that uses a 'reorder' trick (denoted as GPTQ-Reorder or GPTQ-R). Other techniques like ZeroQuant (Yao et al., 2022), AdaRound (Nagel et al., 2020), and BRECQ (Li et al., 2021) rely on backpropagation to update the quantized weights, which may not easily scale up to large model sizes; they also do not outperform GPTQ (Frantar et al., 2022), thus not included for study.",
    'We propose an alternative method to reduce the quantization error of the salient weight by per-channel scaling , which does not suffer from the hardware inefficiency issue.',
    'We thank MIT AI Hardware Program, National Science Foundation, MIT-IBM Watson AI Lab, Amazon and MIT Science Hub, Microsoft Turing Academic Program, and Samsung for supporting this research.',
]
query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings.shape)
# [1, 768] [3, 768]

# Get the similarity scores for the embeddings
similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
# tensor([[ 0.9190, -0.3841, -0.6966]])

Training Details

Training Dataset

Unnamed Dataset

Size: 330 training samples
Columns: anchor, positive, and negative

Approximate statistics based on the first 330 samples:

	anchor	positive	negative
type	string	string	string
details	min: 9 tokens mean: 25.87 tokens max: 48 tokens	min: 2 tokens mean: 863.23 tokens max: 1961 tokens	min: 32 tokens mean: 65.48 tokens max: 905 tokens

Samples:

anchor	positive	negative
`Who are the authors of the paper?`	`Ji Lin * 1 Jiaming Tang * 1 2 Haotian Tang † 1 Shang Yang † 1 Wei-Ming Chen 3 Wei-Chen Wang 1 Guangxuan Xiao 1 Xingyu Dang 1 4 Chuang Gan 5 6 Song Han 1 3. https://github.com/mit-han-lab/llm-awq`	`We propose an alternative method to reduce the quantization error of the salient weight by per-channel scaling , which does not suffer from the hardware inefficiency issue.`
`Who are the authors of the paper?`	`Ji Lin * 1 Jiaming Tang * 1 2 Haotian Tang † 1 Shang Yang † 1 Wei-Ming Chen 3 Wei-Chen Wang 1 Guangxuan Xiao 1 Xingyu Dang 1 4 Chuang Gan 5 6 Song Han 1 3. https://github.com/mit-han-lab/llm-awq`	`We propose an alternative method to reduce the quantization error of the salient weight by per-channel scaling , which does not suffer from the hardware inefficiency issue.`
`Who are the authors of the paper?`	`Ji Lin * 1 Jiaming Tang * 1 2 Haotian Tang † 1 Shang Yang † 1 Wei-Ming Chen 3 Wei-Chen Wang 1 Guangxuan Xiao 1 Xingyu Dang 1 4 Chuang Gan 5 6 Song Han 1 3. https://github.com/mit-han-lab/llm-awq`	`We thank MIT AI Hardware Program, National Science Foundation, MIT-IBM Watson AI Lab, Amazon and MIT Science Hub, Microsoft Turing Academic Program, and Samsung for supporting this research.`

Loss: MultipleNegativesRankingLoss with these parameters:

{
    "scale": 20.0,
    "similarity_fct": "cos_sim",
    "gather_across_devices": false
}

Training Hyperparameters

Non-Default Hyperparameters

per_device_train_batch_size: 1
learning_rate: 2e-05
num_train_epochs: 5
warmup_ratio: 0.1

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: no
prediction_loss_only: True
per_device_train_batch_size: 1
per_device_eval_batch_size: 8
per_gpu_train_batch_size: None
per_gpu_eval_batch_size: None
gradient_accumulation_steps: 1
eval_accumulation_steps: None
torch_empty_cache_steps: None
learning_rate: 2e-05
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
max_grad_norm: 1.0
num_train_epochs: 5
max_steps: -1
lr_scheduler_type: linear
lr_scheduler_kwargs: {}
warmup_ratio: 0.1
warmup_steps: 0
log_level: passive
log_level_replica: warning
log_on_each_node: True
logging_nan_inf_filter: True
save_safetensors: True
save_on_each_node: False
save_only_model: False
restore_callback_states_from_checkpoint: False
no_cuda: False
use_cpu: False
use_mps_device: False
seed: 42
data_seed: None
jit_mode_eval: False
use_ipex: False
bf16: False
fp16: False
fp16_opt_level: O1
half_precision_backend: auto
bf16_full_eval: False
fp16_full_eval: False
tf32: None
local_rank: 0
ddp_backend: None
tpu_num_cores: None
tpu_metrics_debug: False
debug: []
dataloader_drop_last: False
dataloader_num_workers: 0
dataloader_prefetch_factor: None
past_index: -1
disable_tqdm: False
remove_unused_columns: True
label_names: None
load_best_model_at_end: False
ignore_data_skip: False
fsdp: []
fsdp_min_num_params: 0
fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
fsdp_transformer_layer_cls_to_wrap: None
accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
parallelism_config: None
deepspeed: None
label_smoothing_factor: 0.0
optim: adamw_torch_fused
optim_args: None
adafactor: False
group_by_length: False
length_column_name: length
ddp_find_unused_parameters: None
ddp_bucket_cap_mb: None
ddp_broadcast_buffers: False
dataloader_pin_memory: True
dataloader_persistent_workers: False
skip_memory_metrics: True
use_legacy_prediction_loop: False
push_to_hub: False
resume_from_checkpoint: None
hub_model_id: None
hub_strategy: every_save
hub_private_repo: None
hub_always_push: False
hub_revision: None
gradient_checkpointing: False
gradient_checkpointing_kwargs: None
include_inputs_for_metrics: False
include_for_metrics: []
eval_do_concat_batches: True
fp16_backend: auto
push_to_hub_model_id: None
push_to_hub_organization: None
mp_parameters:
auto_find_batch_size: False
full_determinism: False
torchdynamo: None
ray_scope: last
ddp_timeout: 1800
torch_compile: False
torch_compile_backend: None
torch_compile_mode: None
include_tokens_per_second: False
include_num_input_tokens_seen: False
neftune_noise_alpha: None
optim_target_modules: None
batch_eval_metrics: False
eval_on_start: False
use_liger_kernel: False
liger_kernel_config: None
eval_use_gather_object: False
average_tokens_across_devices: False
prompts: None
batch_sampler: batch_sampler
multi_dataset_batch_sampler: proportional
router_mapping: {}
learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss
1.0	330	0.3303
2.0	660	0.4762
3.0	990	0.0672
4.0	1320	0.0185
5.0	1650	0.0051

Framework Versions

Python: 3.13.7
Sentence Transformers: 5.1.1
Transformers: 4.56.2
PyTorch: 2.8.0+cu128
Accelerate: 1.10.1
Datasets: 4.1.1
Tokenizers: 0.22.1

Citation

BibTeX

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

MultipleNegativesRankingLoss

@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}