wehub-resource-sync
77 lines
2.8 KiB
Markdown
77 lines
2.8 KiB
Markdown
# Differential Transformer V2 (DIFF V2)
|
||
|
||
[Read the blog post here](https://spiky-homegrown-4cb.notion.site/Differential-Transformer-V2-2e7baa052def80ecaa93d4d67d125417)
|
||
|
||
The implementation is provided in `multihead_flashdiffv2.py`.
|
||
|
||
## TL;DR
|
||
|
||
We introduce **Differential Transformer V2** (DIFF V2), an improved version of [Differential Transformer](https://arxiv.org/abs/2410.05258) (DIFF V1). This revision focuses on inference efficiency, training stability for production-level LLMs, and architectural elegance.
|
||
|
||
### Key Improvements
|
||
|
||
1. **Faster Inference & No Need of Custom Attention Kernels**
|
||
Instead of forcing the attention parameter count to match the baseline Transformer (as in DIFF V1), we introduce additional parameters for $Q_2$. This design allows DIFF V2 to match the baseline Transformer’s decoding speed and directly use [FlashAttention](https://github.com/Dao-AILab/flash-attention) without custom kernels.
|
||
|
||
2. **Improved Training Stability**
|
||
We remove the per-head RMSNorm after differential attention. We find the per-head RMSNorm can lead to instability in later stages of large-scale pretraining of LLM.
|
||
|
||
3. **Simpler Parameterization & Initialization**
|
||
We replace the globally shared $\lambda$ with a token-specific, head-wise projected $\lambda$. This eliminates the exponential re-parameterization and initialization complexity of $\lambda$ in V1.
|
||
|
||
## Implementation Details
|
||
|
||
### Pseudocode
|
||
|
||
In the script, `h` represents number of query heads, `h_kv` represents number of key-value heads, and `d` means head dimension. The $\lambda$ in DIFF V2 is projected from $X$ for each token each head.
|
||
|
||
(For simplicity, we omit the batch dimension and assume that both the input and output of the following `flash_attn_func` are three-dimensional tensors `(tokens, heads, head dimension)`. Heads belonging to the same GQA group are arranged contiguously in the output)
|
||
|
||
```python
|
||
def DiffAttnV2(
|
||
q, k, v, lam
|
||
):
|
||
"""
|
||
q: (N, 2h, d)
|
||
k: (N, h_kv, d)
|
||
v: (N, h_kv, d)
|
||
lam: (N, h, 1)
|
||
"""
|
||
|
||
attn = flash_attn_func(q, k, v)
|
||
attn1, attn2 = (attn[:, 0::2],
|
||
attn[:, 1::2])
|
||
|
||
lam_val = sigmoid(lam)
|
||
attn = attn1 - lam_val * attn2
|
||
return attn
|
||
```
|
||
|
||
### Note
|
||
|
||
DIFF V2 subtracts two heads that are **in the same GQA group, which means they share the same key and value**.
|
||
|
||
```python
|
||
# Subtraction of two heads that are **not** in the same GQA group
|
||
# ❌ Wrong Implementation of DIFF V2!
|
||
...
|
||
attn = flash_attn_func(q, k, v)
|
||
nh = attn.size(1)
|
||
attn1, attn2 = (attn[:, :nh//2],
|
||
attn[:, nh//2:])
|
||
# similarly, also wrong implementation:
|
||
# attn1, attn2 = attn.chunk(2, dim=1)
|
||
...
|
||
```
|
||
|
||
```python
|
||
# DIFF V2: Subtraction of two heads that are **in** the same GQA group
|
||
# ✅ Correct Implementation of DIFF V2
|
||
...
|
||
attn = flash_attn_func(q, k, v)
|
||
|
||
attn1, attn2 = (attn[:, 0::2],
|
||
attn[:, 1::2])
|
||
...
|
||
```
|