Transformer + GFM
No estimate
No hardware requirements for this model
The weights for this model have not been published, so it cannot be downloaded or run on your own hardware at any size. It is reachable only through its provider, and no graphics card changes that.
On record
Full specification
Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.
Origin
Who built this model, where, and when it was published.
- Organisation
- Nanjing University
- Organisation type
- Academia
- Country
- China
- Published
- 1 December 2022
- Authors
- Hao Yu, Jianxin Wu
What it does
The problem areas the model was built for. A model can carry several of each.
- Domain
- Language
- Task
- Language modeling
- Base model
- FAIRSEQ Adaptive Inputs
Size
How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.
- Parameters
- 185.2M
- Training data
- 103,000,000 tokens
185.2M (Table 4) "We implemented our methods based on fairseq (Ott et al. 2019). The original transformer model follows the architectural choice described in Baevski and Auli (2018), which includes 16 decoder blocks and sinusoidal position embeddings in the input layer. Each MHSA module has 8 heads and adaptive input representations have three bands of size 20K, 40K and 200K. The embedding layer and FFN’s hidden-state have dimensions of 1024 and 4096, respectively. We sampled 4K sentences to fo…
Training compute
The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.
- Training compute
- 7.7 × 10¹⁸ FLOP
- How it was established
- Operation counting,Hardware
- Fine-tuning compute
- 4.4 × 10¹⁷ FLOP
7.30e18 FLOP [base transformer] + 4.3694162e+17 FLOP [GFM] = 7.7369416e+18 FLOP _______________ Estimations from the Algorithmic progress paper (upd - a100 gpu were assumed while the paper reports 3090 gpus): SOURCE: Compression of Baevski et al. transformer, impute During the GFM process, we removed layer dropout and trained on 8 GPUs. We limited the number of tokens per GPU to a maximum threshold 1536, which means each GPU processes 1536 tokens using the same model parameters. We accumulat…
6 FLOP/token/parameter * 185000000 parameters * 103000000 tokens = 1.1433e+17 FLOP or 0.6 hours - reported training time for another model in the paper, I assumed, this training was similar in time length 0.6 hours * 3600 sec / hour * 8 GPUs * 35580000000000 FLOP/s [assumed precision fp16] * 0.3 [assumed utilization] = 1.8444672e+17 FLOP or "During the GFM process, we removed layer dropout and trained on 8 GPUs. We limited the number of tokens per GPU to a maximum threshold 1536, which me…
The training run
What it physically took to train: which chips, how many, for how long, and what that drew from the wall.
- Training hardware
- NVIDIA GeForce RTX 3090
- Chips used
- 8
- Chip-hours
- 1
- Power draw
- 5.6 kW
Availability
Whether you can obtain the model and run it on your own hardware, which is what decides if any of the graphics-card figures on this page apply.
- Weights
- Closed — provider access only
- Model access
- Unreleased
- Training code
- Unreleased
How it is classified
Labels the source dataset applies when tracking notable models, and how confident it is in the entry.
- Record confidence
- Confident
- Benchmark data
- Transformer + GFM
Sources
Where this record came from and when it was last checked.
- Reference
- "Compressing Transformers: Features Are Low-Rank, but Weights Are Not"
- Last updated
- 28 November 2025
What the numbers mean
Where it came from
Transformer + GFM was published by Nanjing University, in the country recorded as China, during December 2022. The category the publisher falls under is academia.
It works in the domain of Language, and is recorded as performing the task of language modeling.
Rather than being trained from scratch, it is derived from FAIRSEQ Adaptive Inputs. That is why it shares the base model's general shape and size.
Because the weights are not available, none of the hardware figures elsewhere on this site apply to it.
Training and provenance
Producing it required arithmetic totalling around 7.7 × 10¹⁸ FLOP, on hardware recorded as NVIDIA GeForce RTX 3090. That figure measures what producing the model cost, and has no bearing on how fast it answers.
It was trained on a corpus of about 103,000,000 tokens of text.
Answers
Transformer + GFM — common questions
Transformer + GFM— what is it used for?
It works in the domain of Language, and is recorded as handling the task of language modeling. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.
Transformer + GFM— how much compute was used to train it?
Training consumed around 7.7 × 10¹⁸ FLOP, on hardware recorded as NVIDIA GeForce RTX 3090. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.
Transformer + GFM— what GPU do I need to run it?
None. This is a closed model — its weights were never published, so it cannot be downloaded or run on your own hardware at any price. It is reachable only through its provider.
Transformer + GFM— is it open source?
No. Its weights have not been published, so it exists only as a service controlled by its owner.
Transformer + GFM— how many parameters does it have?
It has a parameter count of 185.2M. 185.2M (Table 4) "We implemented our methods based on fairseq (Ott et al. 2019). The original transformer model follows the architectural choice described in Baevski and Auli (2018), which includes 16 decoder blocks and sinusoidal position embeddings in the input layer. Each MHSA module has 8 heads and adaptive input representations have three bands of size 20K, 40K and 200K. The embedding layer and FFN’s hidden-state have dimensions of 1024 and 4096, respectively. We sampled 4K sentences to form the proxy dataset D. Then, we reduced the parameters by 15%, 20% and 25%.". That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.
Transformer + GFM— who created it?
It was published by Nanjing University, based in China, an organisation categorised as academia.
Transformer + GFM— when was it released?
It was published in December 2022. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.
The other direction
Looking at it from the other side?
This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.