chore: import upstream snapshot with attribution

Diff-Transformer/Diff-Transformer-V2/README.md

@@ -0,0 +1,76 @@
+# 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])
+...
+```

Diff-Transformer/Diff-Transformer-V2/multihead_flashdiffv2.py

@@ -0,0 +1,76 @@
+import torch
+from torch import nn
+from typing import Optional, Tuple
+from ..kernel.rotary import apply_rotary_emb
+from flash_attn import flash_attn_func
+
+
+@torch.compile
+def diff_func(attn1: torch.Tensor, attn2: torch.Tensor, lambda_val: torch.Tensor) -> torch.Tensor:
+    return attn1 - torch.sigmoid(lambda_val).unsqueeze(-1) * attn2
+
+
+class MultiheadFlashDiffV2(nn.Module):
+    """
+    Differential Attention Version 2 (DiffAttnV2) implementation using Flash Attention.
+    """
+    def __init__(
+        self,
+        use_diff_v2: bool, # If False, acts as a baseline Transformer attention
+        d_model: int,      # Model dimension
+        num_heads: int,    # Number of output heads
+        num_kv_heads: Optional[int], # Number of KV heads for GQA
+        head_dim: int,     # Dimension per head
+    ):
+        super().__init__()
+        self.use_diff_v2 = use_diff_v2
+        self.d_model = d_model
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads if num_kv_heads is not None else num_heads
+        self.head_dim = head_dim
+        
+        self.num_q_heads = 2 * self.num_heads if self.use_diff_v2 else self.num_heads
+        self.q_proj = nn.Linear(self.d_model, self.num_q_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.d_model, self.num_kv_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.d_model, self.num_kv_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.d_model, bias=False)
+        self.lambda_proj = nn.Linear(self.d_model, self.num_heads, bias=False) if self.use_diff_v2 else None
+    
+    def forward(
+        self,
+        x: torch.Tensor,               # Input tensor [bsz, seq_len, d_model]
+        rel_pos: Tuple[torch.Tensor, torch.Tensor], # Rotary embedding (cos, sin)
+    ) -> torch.Tensor:
+        """
+        Forward pass for MultiheadFlashDiffV2.
+        
+        Args:
+            x: Input hidden states of shape [batch, length, d_model]
+            rel_pos: Tuple of (cos, sin) tensors for rotary positional embeddings
+            
+        Returns:
+            Output tensor of shape [batch, length, d_model]
+        """
+        bsz, tgt_len, _ = x.size()
+        src_len = tgt_len
+
+        q = self.q_proj(x)
+        k = self.k_proj(x)
+        v = self.v_proj(x)
+
+        q = q.view(bsz, tgt_len, self.num_q_heads, self.head_dim)
+        k = k.view(bsz, src_len, self.num_kv_heads, self.head_dim)
+        v = v.view(bsz, src_len, self.num_kv_heads, self.head_dim)
+
+        q = apply_rotary_emb(q, *rel_pos, interleaved=True)
+        k = apply_rotary_emb(k, *rel_pos, interleaved=True)
+
+        attn = flash_attn_func(q, k, v, causal=True)
+        if self.use_diff_v2:
+            lambda_val = self.lambda_proj(x)
+            attn1, attn2 = attn[:, :, 0::2], attn[:, :, 1::2]
+            attn = diff_func(attn1, attn2, lambda_val)
+        
+        attn = attn.reshape(bsz, tgt_len, self.num_heads * self.head_dim)
+        output = self.o_proj(attn)
+        return output