LightDiffusion-Next / src /clip /KleinEncoder.py
Aatricks's picture
Deploy ZeroGPU Gradio Space snapshot
b701455
"""Klein text encoder for Flux2 models in LightDiffusion-Next.
This module provides the Klein (Qwen3-4B based) text encoder used by
Flux2 Klein models, including:
- KleinTokenizer: Qwen3-based tokenizer with special formatting
- Qwen3Model: Transformer-based language model for text encoding
Adapted from ComfyUI's Klein implementation.
"""
import logging
import math
from typing import Optional, List, Dict, Any
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange
from src.cond import cast as ops_module
from src.Device import Device
logger = logging.getLogger(__name__)
def get_ops():
"""Get the operations module for weight initialization."""
return ops_module.disable_weight_init
class QwenRMSNorm(nn.Module):
"""RMS Normalization for Qwen3."""
def __init__(self, dim: int, eps: float = 1e-6, dtype=None, device=None, operations=None):
super().__init__()
if operations is None:
operations = get_ops()
self.eps = eps
self.weight = nn.Parameter(torch.ones(dim, dtype=dtype, device=device))
def forward(self, x: torch.Tensor) -> torch.Tensor:
# RMS normalization - compute in float32 for precision, cast back to input dtype
input_dtype = x.dtype
x_float = x.float()
rms = torch.rsqrt(x_float.pow(2).mean(-1, keepdim=True) + self.eps)
return (x_float * rms * self.weight.float()).to(input_dtype)
class QwenRotaryEmbedding(nn.Module):
"""Rotary position embeddings for Qwen3."""
def __init__(self, dim: int, max_position_embeddings: int = 32768, base: float = 1000000.0):
super().__init__()
self.dim = dim
self.max_seq_len = max_position_embeddings
self.base = base
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
def forward(self, x: torch.Tensor, seq_len: int = None):
if seq_len is None:
seq_len = x.shape[1]
t = torch.arange(seq_len, device=x.device, dtype=self.inv_freq.dtype)
freqs = torch.outer(t, self.inv_freq)
emb = torch.cat((freqs, freqs), dim=-1)
return emb.cos(), emb.sin()
def rotate_half(x: torch.Tensor) -> torch.Tensor:
"""Rotate half for RoPE."""
x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
return torch.cat((-x2, x1), dim=-1)
def apply_rotary_pos_emb(q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor):
"""Apply rotary position embeddings to query and key."""
cos = cos.unsqueeze(0).unsqueeze(0) # [1, 1, seq, dim]
sin = sin.unsqueeze(0).unsqueeze(0)
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
return q_embed, k_embed
class QwenAttention(nn.Module):
"""Multi-head attention for Qwen3 with Grouped Query Attention."""
def __init__(
self,
hidden_size: int,
num_heads: int,
num_kv_heads: int = None,
head_dim: int = 128,
dtype=None,
device=None,
operations=None,
):
super().__init__()
if operations is None:
operations = get_ops()
self.hidden_size = hidden_size
self.num_heads = num_heads
self.num_kv_heads = num_kv_heads or num_heads
self.head_dim = head_dim
# Qwen3 uses separate projections with different output sizes
self.q_proj = operations.Linear(hidden_size, num_heads * head_dim, bias=False, dtype=dtype, device=device)
self.k_proj = operations.Linear(hidden_size, self.num_kv_heads * head_dim, bias=False, dtype=dtype, device=device)
self.v_proj = operations.Linear(hidden_size, self.num_kv_heads * head_dim, bias=False, dtype=dtype, device=device)
self.o_proj = operations.Linear(num_heads * head_dim, hidden_size, bias=False, dtype=dtype, device=device)
# Normalize Q and K
self.q_norm = QwenRMSNorm(head_dim, dtype=dtype, device=device, operations=operations)
self.k_norm = QwenRMSNorm(head_dim, dtype=dtype, device=device, operations=operations)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_embeddings: Optional[tuple] = None,
) -> torch.Tensor:
batch_size, seq_len, _ = hidden_states.shape
q = self.q_proj(hidden_states)
k = self.k_proj(hidden_states)
v = self.v_proj(hidden_states)
# Reshape for multi-head attention
q = q.view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
k = k.view(batch_size, seq_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
v = v.view(batch_size, seq_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
# Apply QK normalization
q = self.q_norm(q)
k = self.k_norm(k)
# Apply rotary embeddings
if position_embeddings is not None:
cos, sin = position_embeddings
q, k = apply_rotary_pos_emb(q, k, cos, sin)
# Grouped query attention - repeat K,V for each group
if self.num_kv_heads != self.num_heads:
n_rep = self.num_heads // self.num_kv_heads
k = k.repeat_interleave(n_rep, dim=1)
v = v.repeat_interleave(n_rep, dim=1)
# Ensure all tensors have same dtype for SDPA
attn_dtype = q.dtype
k = k.to(attn_dtype)
v = v.to(attn_dtype)
# Scaled dot-product attention with causal masking
# Use is_causal=True for efficiency, or attn_mask for custom masks
if attention_mask is None:
# Pure causal masking
attn_output = F.scaled_dot_product_attention(q, k, v, is_causal=True)
else:
# Custom mask (includes causal + padding) - ensure mask dtype matches
attention_mask = attention_mask.to(attn_dtype)
attn_output = F.scaled_dot_product_attention(q, k, v, attn_mask=attention_mask)
# Reshape back and ensure output dtype matches input for o_proj
attn_output = attn_output.transpose(1, 2).contiguous().view(batch_size, seq_len, -1)
attn_output = attn_output.to(hidden_states.dtype) # Match input dtype for o_proj
return self.o_proj(attn_output)
class QwenMLP(nn.Module):
"""MLP (Gate-Up-Down) for Qwen3."""
def __init__(
self,
hidden_size: int,
intermediate_size: int,
dtype=None,
device=None,
operations=None,
):
super().__init__()
if operations is None:
operations = get_ops()
self.gate_proj = operations.Linear(hidden_size, intermediate_size, bias=False, dtype=dtype, device=device)
self.up_proj = operations.Linear(hidden_size, intermediate_size, bias=False, dtype=dtype, device=device)
self.down_proj = operations.Linear(intermediate_size, hidden_size, bias=False, dtype=dtype, device=device)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x))
class QwenDecoderLayer(nn.Module):
"""Single transformer decoder layer for Qwen3."""
def __init__(
self,
hidden_size: int,
num_heads: int,
intermediate_size: int,
num_kv_heads: int = None,
head_dim: int = 128,
dtype=None,
device=None,
operations=None,
):
super().__init__()
if operations is None:
operations = get_ops()
self.self_attn = QwenAttention(
hidden_size=hidden_size,
num_heads=num_heads,
num_kv_heads=num_kv_heads,
head_dim=head_dim,
dtype=dtype,
device=device,
operations=operations,
)
self.mlp = QwenMLP(
hidden_size=hidden_size,
intermediate_size=intermediate_size,
dtype=dtype,
device=device,
operations=operations,
)
self.input_layernorm = QwenRMSNorm(hidden_size, dtype=dtype, device=device, operations=operations)
self.post_attention_layernorm = QwenRMSNorm(hidden_size, dtype=dtype, device=device, operations=operations)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_embeddings: Optional[tuple] = None,
) -> torch.Tensor:
# Self attention
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
hidden_states = self.self_attn(hidden_states, attention_mask, position_embeddings)
hidden_states = residual + hidden_states
# MLP
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = residual + hidden_states
return hidden_states
class Qwen3_4BModel(nn.Module):
"""Qwen3 4B model for Klein text encoding.
This is a decoder-only transformer used as a text encoder
for the Flux2 Klein model.
"""
def __init__(
self,
vocab_size: int = 151936,
hidden_size: int = 2560,
intermediate_size: int = 9728, # Matches checkpoint
num_hidden_layers: int = 36,
num_attention_heads: int = 32, # Matches checkpoint (4096/128)
num_key_value_heads: int = 8, # Matches checkpoint (1024/128)
head_dim: int = 128,
max_position_embeddings: int = 32768,
layer_indices: tuple = (9, 18, 27), # Layers to extract embeddings from
dtype=None,
device=None,
operations=None,
):
super().__init__()
if operations is None:
operations = get_ops()
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.layer_indices = layer_indices
# Token embeddings
self.embed_tokens = operations.Embedding(vocab_size, hidden_size, dtype=dtype, device=device)
# Rotary embeddings
self.rotary_emb = QwenRotaryEmbedding(
head_dim,
max_position_embeddings=max_position_embeddings,
)
# Transformer layers
self.layers = nn.ModuleList([
QwenDecoderLayer(
hidden_size=hidden_size,
num_heads=num_attention_heads,
intermediate_size=intermediate_size,
num_kv_heads=num_key_value_heads,
head_dim=head_dim,
dtype=dtype,
device=device,
operations=operations,
)
for _ in range(num_hidden_layers)
])
# Final norm
self.norm = QwenRMSNorm(hidden_size, dtype=dtype, device=device, operations=operations)
def forward(
self,
input_ids: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
) -> dict:
"""Forward pass returning hidden states from specified layers.
Args:
input_ids: Token IDs [batch, seq_len]
attention_mask: Optional attention mask
Returns:
Dict with 'hidden_states' from specified layers (concatenated)
"""
batch_size, seq_len = input_ids.shape
# Embed tokens - ensure output is at least bfloat16 for subsequent math
hidden_states = self.embed_tokens(input_ids).to(torch.bfloat16)
# Get rotary embeddings
cos, sin = self.rotary_emb(hidden_states, seq_len)
position_embeddings = (cos, sin)
# Prepare attention mask
# If we have a padding mask, create a combined causal + padding mask
# Otherwise, pass None and let the attention layer use is_causal=True
final_mask = None
if attention_mask is not None:
# Create mask matching ComfyUI's approach:
# 1. Convert padding mask from [B, L] to [B, 1, L, L] with expansion
# 2. Set padded positions (where mask=0) to -inf
# 3. Add causal mask
# Reshape and expand: [B, L] -> [B, 1, L, L]
mask = 1.0 - attention_mask.to(hidden_states.dtype) # 0 = valid, 1 = padding
mask = mask.reshape(mask.shape[0], 1, -1, mask.shape[-1]) # [B, 1, 1, L]
mask = mask.expand(mask.shape[0], 1, seq_len, seq_len) # [B, 1, L, L]
mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))
# Create causal mask [L, L]
causal_mask = torch.empty(seq_len, seq_len, dtype=hidden_states.dtype, device=input_ids.device).fill_(float("-inf")).triu_(1)
# Combine
final_mask = mask + causal_mask
# Collect outputs from specified layers
# NOTE: ComfyUI captures the INPUT to layers (before the layer runs),
# so we capture before applying each layer
layer_outputs = []
for i, layer in enumerate(self.layers):
hidden_states = layer(hidden_states, final_mask, position_embeddings)
# Capture AFTER the layer (output of layer i)
if i in self.layer_indices:
layer_outputs.append(hidden_states.clone())
# Apply final norm
hidden_states = self.norm(hidden_states)
# Concatenate layer outputs matching ComfyUI's interleaving pattern
# This is critical for Flux2/Klein cross-attention
if layer_outputs:
# layer_outputs is a list of [B, L, D] tensors
# stack: (B, 3, L, D)
stacked = torch.stack(layer_outputs, dim=1)
# permute: (B, L, 3, D) - interleave the 3 layers at each sequence position
permuted = stacked.permute(0, 2, 1, 3)
# reshape: (B, L, 3*D)
concatenated = permuted.reshape(batch_size, seq_len, -1)
else:
concatenated = hidden_states
return {
"last_hidden_state": hidden_states,
"hidden_states": concatenated,
"pooled_output": None, # Match ComfyUI: No pooling for Qwen -> Flux2 uses zeros
}
class KleinTokenizer:
"""Tokenizer for Klein (Qwen3-based) text encoder.
Uses Qwen2Tokenizer from Hugging Face transformers with
Klein-specific formatting template.
"""
# Klein template for prompt formatting
TEMPLATE = "<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
def __init__(
self,
tokenizer_path: str = None,
max_length: int = 99999999, # ComfyUI uses essentially unlimited
min_length: int = 512, # ComfyUI uses min_length=512 for Klein
padding: str = "do_not_pad", # ComfyUI uses pad_to_max_length=False
):
self.max_length = max_length
self.min_length = min_length
self.padding = padding
# Klein special tokens
self.pad_token_id = 151643 # <|endoftext|>
self.bos_token_id = 151644 # <|im_start|>
self.eos_token_id = 151645 # <|im_end|>
# Load the real tokenizer
if tokenizer_path is None:
# Default path relative to include folder
import os
# Try multiple locations
possible_paths = [
os.path.join(os.path.dirname(os.path.dirname(os.path.dirname(os.path.realpath(__file__)))), "include", "text_encoder", "qwen25_tokenizer"),
os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer"),
]
for path in possible_paths:
if os.path.exists(path):
tokenizer_path = path
break
else:
tokenizer_path = possible_paths[0] # Use first as default
try:
from transformers import Qwen2Tokenizer
self._tokenizer = Qwen2Tokenizer.from_pretrained(tokenizer_path)
# Use right padding for content-first alignment, matching ComfyUI default
self._tokenizer.padding_side = "right"
logger.info(f"Loaded Qwen2Tokenizer from {tokenizer_path}")
except Exception as e:
logger.error(f"Failed to load tokenizer: {e}")
raise RuntimeError(f"Could not load Klein tokenizer from {tokenizer_path}") from e
def apply_template(self, text: str) -> str:
"""Apply Klein's prompt template to input text."""
return self.TEMPLATE.format(text)
def tokenize_with_weights(self, text: str, return_word_ids: bool = False) -> dict:
"""Tokenize text with Klein template formatting.
Args:
text: Input text to tokenize
return_word_ids: Whether to return word IDs
Returns:
Dict with 'input_ids' and 'attention_mask'
"""
# Apply template
formatted_text = self.apply_template(text)
# Tokenize with the real tokenizer - pad to min_length (512)
# Matches ComfyUI Qwen3Tokenizer behavior
encoded = self._tokenizer(
formatted_text,
padding="max_length",
max_length=self.min_length,
truncation=True,
return_tensors="pt",
)
result = {
"input_ids": encoded["input_ids"],
"attention_mask": encoded["attention_mask"],
}
if return_word_ids:
# Word IDs from the tokenizer's encoding
word_ids = encoded.word_ids() if hasattr(encoded, 'word_ids') else list(range(encoded["input_ids"].shape[1]))
result["word_ids"] = word_ids
return result
def state_dict(self) -> dict:
"""Return tokenizer state for serialization."""
return {
"max_length": self.max_length,
"min_length": self.min_length,
"padding": self.padding,
}
class KleinCLIP:
"""Klein text encoder wrapper compatible with CLIP interface.
This provides the same interface as other CLIP models while
using the Qwen3-based Klein encoder internally.
VRAM Optimization: Model stays on CPU until encoding, then moves to GPU
and back to CPU. This follows ComfyUI's lazy loading approach.
"""
def __init__(
self,
tokenizer: KleinTokenizer = None,
model: Qwen3_4BModel = None,
dtype=None,
device=None,
offload_device=None,
):
self.tokenizer = tokenizer or KleinTokenizer()
self.dtype = dtype
self.device = device # Device to use for encoding (GPU)
self.offload_device = offload_device or torch.device("cpu") # Device when idle (CPU)
if model is None:
self.model = Qwen3_4BModel(dtype=dtype, device=self.offload_device)
else:
self.model = model
self.clip_options = {}
def reset_clip_options(self):
"""Reset clip options to defaults."""
self.clip_options = {}
def set_clip_options(self, options: dict):
"""Set clip options (for API compatibility)."""
self.clip_options.update(options)
def encode_token_weights(self, tokens: dict) -> tuple:
"""Encode token weights returning (embeddings, pooled, extra).
VRAM Optimization: Moves model to GPU only during encoding,
then offloads back to CPU to free VRAM for diffusion model.
Args:
tokens: Dict with 'input_ids' and 'attention_mask' tensors
Returns:
Tuple of (hidden_states, pooled_output, extra_dict)
where extra_dict contains 'attention_mask' for the diffusion model
"""
input_ids = tokens.get("input_ids")
if input_ids is None:
raise ValueError("tokens dict must contain 'input_ids'")
# Move model to GPU for encoding
logger.info(f"Moving text encoder to {self.device} for encoding...")
self.model = self.model.to(self.device)
input_ids = input_ids.to(self.device)
# Get attention mask if present - CRITICAL for proper masking of padding tokens
attention_mask = tokens.get("attention_mask")
if attention_mask is not None:
attention_mask = attention_mask.to(self.device)
try:
with torch.no_grad():
outputs = self.model(input_ids, attention_mask=attention_mask)
# Return concatenated hidden states, pooled output, and extra with attention_mask
hidden_states = outputs["hidden_states"].clone() # Clone to keep on GPU
pooled_out = outputs["pooled_output"]
pooled = pooled_out.clone() if pooled_out is not None else None # Clone if exists
finally:
# Offload model back to CPU to free VRAM for diffusion model
logger.info(f"Offloading text encoder to {self.offload_device}...")
self.model = self.model.to(self.offload_device)
if torch.cuda.is_available():
torch.cuda.empty_cache()
# Return attention mask in extra dict for the diffusion model to use
extra = {}
if attention_mask is not None:
extra["attention_mask"] = attention_mask
return hidden_states, pooled, extra
def tokenize(self, text, return_word_ids=False):
"""Tokenize text (CLIP interface compatibility for Adetailer).
Args:
text: Text to tokenize
return_word_ids: Whether to return word IDs
Returns:
Dict with 'input_ids' and 'attention_mask'
"""
return self.tokenizer.tokenize_with_weights(text, return_word_ids)
def encode_from_tokens(self, tokens, return_pooled=False, return_dict=False):
"""Encode from tokens (CLIP interface compatibility for Adetailer).
Args:
tokens: Dict with 'input_ids' and 'attention_mask'
return_pooled: Whether to return pooled output
return_dict: Whether to return as dict
Returns:
Embeddings, or (embeddings, pooled) if return_pooled, or dict if return_dict
"""
cond, pooled, extra = self.encode_token_weights(tokens)
if return_dict:
out = {"cond": cond, "pooled_output": pooled}
out.update(extra)
return out
return (cond, pooled) if return_pooled else cond
def load_model(self):
"""Load model to GPU (CLIP interface compatibility).
Returns:
Self for compatibility
"""
# Move model to device if not already there
if self.device is not None:
self.model = self.model.to(self.device)
return self
def load_sd(self, state_dict: dict) -> tuple:
"""Load state dictionary into model.
Args:
state_dict: Model weights
Returns:
Tuple of (missing_keys, unexpected_keys)
"""
# Filter and map state dict keys for Qwen3 model
model_sd = {}
for k, v in state_dict.items():
# Map state dict keys to model structure
if k.startswith("model."):
model_sd[k[6:]] = v # Remove "model." prefix
else:
model_sd[k] = v
return self.model.load_state_dict(model_sd, strict=False)
def klein_clip(dtype=None) -> dict:
"""Create Klein CLIP configuration.
Returns:
Dict with 'clip' and 'tokenizer' classes
"""
class Target:
clip = KleinCLIP
tokenizer = KleinTokenizer
params = {"dtype": dtype}
return Target
# Convenience function to detect Klein model from state dict
def detect_klein_model(state_dict: dict) -> bool:
"""Detect if state dict is from a Klein text encoder.
Args:
state_dict: Model state dictionary
Returns:
True if this appears to be a Klein model
"""
klein_indicators = [
"model.layers.0.self_attn.q_norm.weight",
"model.layers.0.self_attn.k_norm.weight",
"embed_tokens.weight",
]
keys = set(state_dict.keys())
for indicator in klein_indicators:
for key in keys:
if indicator in key:
return True
return False