| """
|
| Paired with a good language model. Thanks!
|
| """
|
|
|
| import torch
|
| import torch.nn.functional as F
|
| from torch.nn.attention import sdpa_kernel, SDPBackend
|
| from typing import Optional, Tuple
|
| from diffusers.models.transformers.transformer_qwenimage import apply_rotary_emb_qwen
|
|
|
|
|
| class QwenDoubleStreamAttnProcessorFA3:
|
| """
|
| FA3-grade attention processor for Qwen double-stream architecture.
|
| Routes through PyTorch's cuDNN attention backend, which on Hopper (SM 9.0+)
|
| dispatches to the same FlashAttention-3 family of kernels as vLLM's FA3 —
|
| bundled with the base image so there's no external-kernel ABI risk.
|
|
|
| Notes / limitations:
|
| - General attention masks are not supported here. `is_causal=False` and no arbitrary mask.
|
| - Expects an available `apply_rotary_emb_qwen` in scope (same as your non-FA3 processor).
|
| """
|
|
|
| _attention_backend = "fa3"
|
|
|
| @torch.no_grad()
|
| def __call__(
|
| self,
|
| attn,
|
| hidden_states: torch.FloatTensor,
|
| encoder_hidden_states: torch.FloatTensor = None,
|
| encoder_hidden_states_mask: torch.FloatTensor = None,
|
| attention_mask: Optional[torch.FloatTensor] = None,
|
| image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
| ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
|
| if encoder_hidden_states is None:
|
| raise ValueError("QwenDoubleStreamAttnProcessorFA3 requires encoder_hidden_states (text stream).")
|
| if attention_mask is not None:
|
| raise NotImplementedError("attention_mask is not supported in this FA3 implementation.")
|
|
|
| B, S_img, _ = hidden_states.shape
|
| S_txt = encoder_hidden_states.shape[1]
|
|
|
|
|
| img_q = attn.to_q(hidden_states)
|
| img_k = attn.to_k(hidden_states)
|
| img_v = attn.to_v(hidden_states)
|
|
|
|
|
| txt_q = attn.add_q_proj(encoder_hidden_states)
|
| txt_k = attn.add_k_proj(encoder_hidden_states)
|
| txt_v = attn.add_v_proj(encoder_hidden_states)
|
|
|
|
|
| H = attn.heads
|
| img_q = img_q.unflatten(-1, (H, -1))
|
| img_k = img_k.unflatten(-1, (H, -1))
|
| img_v = img_v.unflatten(-1, (H, -1))
|
|
|
| txt_q = txt_q.unflatten(-1, (H, -1))
|
| txt_k = txt_k.unflatten(-1, (H, -1))
|
| txt_v = txt_v.unflatten(-1, (H, -1))
|
|
|
|
|
| if getattr(attn, "norm_q", None) is not None:
|
| img_q = attn.norm_q(img_q)
|
| if getattr(attn, "norm_k", None) is not None:
|
| img_k = attn.norm_k(img_k)
|
| if getattr(attn, "norm_added_q", None) is not None:
|
| txt_q = attn.norm_added_q(txt_q)
|
| if getattr(attn, "norm_added_k", None) is not None:
|
| txt_k = attn.norm_added_k(txt_k)
|
|
|
|
|
| if image_rotary_emb is not None:
|
| img_freqs, txt_freqs = image_rotary_emb
|
| img_q = apply_rotary_emb_qwen(img_q, img_freqs, use_real=False)
|
| img_k = apply_rotary_emb_qwen(img_k, img_freqs, use_real=False)
|
| txt_q = apply_rotary_emb_qwen(txt_q, txt_freqs, use_real=False)
|
| txt_k = apply_rotary_emb_qwen(txt_k, txt_freqs, use_real=False)
|
|
|
|
|
| q = torch.cat([txt_q, img_q], dim=1)
|
| k = torch.cat([txt_k, img_k], dim=1)
|
| v = torch.cat([txt_v, img_v], dim=1)
|
|
|
|
|
| with sdpa_kernel(SDPBackend.CUDNN_ATTENTION):
|
| out = F.scaled_dot_product_attention(
|
| q.transpose(1, 2),
|
| k.transpose(1, 2),
|
| v.transpose(1, 2),
|
| is_causal=False,
|
| ).transpose(1, 2)
|
|
|
|
|
| out = out.flatten(2, 3).to(q.dtype)
|
|
|
|
|
| txt_attn_out = out[:, :S_txt, :]
|
| img_attn_out = out[:, S_txt:, :]
|
|
|
|
|
| img_attn_out = attn.to_out[0](img_attn_out)
|
| if len(attn.to_out) > 1:
|
| img_attn_out = attn.to_out[1](img_attn_out)
|
|
|
| txt_attn_out = attn.to_add_out(txt_attn_out)
|
|
|
| return img_attn_out, txt_attn_out
|
|
|