diff --git a/docs/supported_models/diffusion_models.md b/docs/supported_models/diffusion_models.md index ec5aa55c3..8ed55a944 100644 --- a/docs/supported_models/diffusion_models.md +++ b/docs/supported_models/diffusion_models.md @@ -4,7 +4,7 @@ SGLang Diffusion is an inference framework for accelerated image and video gener ## Key Features -- **Broad Model Support**: Wan series, FastWan series, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux, and more +- **Broad Model Support**: Wan series, FastWan series, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux, Z-Image, GLM-Image, and more - **Fast Inference**: Optimized kernels from sgl-kernel, efficient scheduler loop, and Cache-DiT acceleration - **Ease of Use**: OpenAI-compatible API, CLI, and Python SDK - **Multi-Platform**: NVIDIA GPUs (H100, H200, A100, B200, 4090) and AMD GPUs (MI300X, MI325X) @@ -110,12 +110,16 @@ default parameters when initializing and generating videos. ### Image Generation Models -| Model Name | HuggingFace Model ID | Resolutions | -|:----------------|:-------------------------------|:---------------| -| FLUX.1-dev | `black-forest-labs/FLUX.1-dev` | Any resolution | -| FLUX.2-dev | `black-forest-labs/FLUX.2-dev` | Any resolution | -| Qwen Image | `Qwen/Qwen-Image` | Any resolution | -| Qwen Image Edit | `Qwen/Qwen-Image-Edit` | Any resolution | +| Model Name | HuggingFace Model ID | Resolutions | +|:-----------------|:----------------------------------------|:---------------| +| FLUX.1-dev | `black-forest-labs/FLUX.1-dev` | Any resolution | +| FLUX.2-dev | `black-forest-labs/FLUX.2-dev` | Any resolution | +| FLUX.2-Klein | `black-forest-labs/FLUX.2-klein-4B` | Any resolution | +| Z-Image-Turbo | `Tongyi-MAI/Z-Image-Turbo` | Any resolution | +| GLM-Image | `zai-org/GLM-Image` | Any resolution | +| Qwen Image | `Qwen/Qwen-Image` | Any resolution | +| Qwen Image 2512 | `Qwen/Qwen-Image-2512` | Any resolution | +| Qwen Image Edit | `Qwen/Qwen-Image-Edit` | Any resolution | ## Verified LoRA Examples @@ -1004,13 +1008,14 @@ All Cache-DiT parameters can be set via the following environment variables: SGLang Diffusion x Cache-DiT supports almost all models originally supported in SGLang Diffusion: -| Model Family | Example Models | -|--------------|-----------------------------| -| Wan | Wan2.1, Wan2.2 | -| Flux | FLUX.1-dev, FLUX.2-dev | -| Z-Image | Z-Image-Turbo | -| Qwen | Qwen-Image, Qwen-Image-Edit | -| Hunyuan | HunyuanVideo | +| Model Family | Example Models | +|--------------|-------------------------------------------| +| Wan | Wan2.1, Wan2.2 | +| Flux | FLUX.1-dev, FLUX.2-dev, FLUX.2-Klein | +| Z-Image | Z-Image-Turbo | +| Qwen | Qwen-Image, Qwen-Image-Edit | +| GLM | GLM-Image | +| Hunyuan | HunyuanVideo | ## Performance Tips diff --git a/python/sglang/multimodal_gen/README.md b/python/sglang/multimodal_gen/README.md index 6acdcffa1..76d925543 100644 --- a/python/sglang/multimodal_gen/README.md +++ b/python/sglang/multimodal_gen/README.md @@ -9,7 +9,7 @@ SGLang diffusion features an end-to-end unified pipeline for accelerating diffus ## Key Features SGLang Diffusion has the following features: - - Broad model support: Wan series, FastWan series, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux + - Broad model support: Wan series, FastWan series, Hunyuan, Qwen-Image, Qwen-Image-Edit, Flux, Z-Image, GLM-Image - Fast inference speed: enpowered by highly optimized kernel from sgl-kernel and efficient scheduler loop - Ease of use: OpenAI-compatible api, CLI, and python sdk support - Multi-platform support: NVIDIA GPUs (H100, H200, A100, B200, 4090) and AMD GPUs (MI300X, MI325X) diff --git a/python/sglang/multimodal_gen/configs/models/encoders/__init__.py b/python/sglang/multimodal_gen/configs/models/encoders/__init__.py index 70851bfa5..b66389ed9 100644 --- a/python/sglang/multimodal_gen/configs/models/encoders/__init__.py +++ b/python/sglang/multimodal_gen/configs/models/encoders/__init__.py @@ -11,6 +11,7 @@ from sglang.multimodal_gen.configs.models.encoders.clip import ( CLIPVisionConfig, ) from sglang.multimodal_gen.configs.models.encoders.llama import LlamaConfig +from sglang.multimodal_gen.configs.models.encoders.qwen3 import Qwen3TextConfig from sglang.multimodal_gen.configs.models.encoders.t5 import T5Config __all__ = [ @@ -21,5 +22,6 @@ __all__ = [ "CLIPTextConfig", "CLIPVisionConfig", "LlamaConfig", + "Qwen3TextConfig", "T5Config", ] diff --git a/python/sglang/multimodal_gen/configs/models/encoders/qwen3.py b/python/sglang/multimodal_gen/configs/models/encoders/qwen3.py new file mode 100644 index 000000000..3b56e690d --- /dev/null +++ b/python/sglang/multimodal_gen/configs/models/encoders/qwen3.py @@ -0,0 +1,85 @@ +# SPDX-License-Identifier: Apache-2.0 +"""Qwen3 text encoder configuration for SGLang diffusion models.""" +from dataclasses import dataclass, field + +from sglang.multimodal_gen.configs.models.encoders.base import ( + TextEncoderArchConfig, + TextEncoderConfig, +) + + +def _is_transformer_layer(n: str, m) -> bool: + return "layers" in n and str.isdigit(n.split(".")[-1]) + + +def _is_embeddings(n: str, m) -> bool: + return n.endswith("embed_tokens") + + +def _is_final_norm(n: str, m) -> bool: + return n.endswith("norm") + + +@dataclass +class Qwen3TextArchConfig(TextEncoderArchConfig): + """Architecture config for Qwen3 text encoder. + + Qwen3 is similar to LLaMA but with QK-Norm (RMSNorm on Q and K before attention). + """ + + vocab_size: int = 151936 + hidden_size: int = 2560 + intermediate_size: int = 9728 + num_hidden_layers: int = 36 + num_attention_heads: int = 32 + num_key_value_heads: int = 8 + hidden_act: str = "silu" + max_position_embeddings: int = 40960 + initializer_range: float = 0.02 + rms_norm_eps: float = 1e-6 + use_cache: bool = True + pad_token_id: int = 151643 + bos_token_id: int = 151643 + eos_token_id: int = 151645 + tie_word_embeddings: bool = True + rope_theta: float = 1000000.0 + rope_scaling: dict | None = None + attention_bias: bool = False + attention_dropout: float = 0.0 + mlp_bias: bool = False + head_dim: int = 128 + text_len: int = 512 + output_hidden_states: bool = True # Klein needs hidden states from layers 9, 18, 27 + + # Stacked params for weight loading with tensor parallelism + stacked_params_mapping: list[tuple[str, str, str]] = field( + default_factory=lambda: [ + # (param_name, shard_name, shard_id) + (".qkv_proj", ".q_proj", "q"), + (".qkv_proj", ".k_proj", "k"), + (".qkv_proj", ".v_proj", "v"), + (".gate_up_proj", ".gate_proj", 0), + (".gate_up_proj", ".up_proj", 1), + ] + ) + + # FSDP sharding conditions for CPU offload + _fsdp_shard_conditions: list = field( + default_factory=lambda: [_is_transformer_layer, _is_embeddings, _is_final_norm] + ) + + def __post_init__(self) -> None: + self.tokenizer_kwargs = { + "padding": "max_length", + "truncation": True, + "max_length": self.text_len, + "return_tensors": "pt", + } + + +@dataclass +class Qwen3TextConfig(TextEncoderConfig): + """Top-level config for Qwen3 text encoder.""" + + arch_config: TextEncoderArchConfig = field(default_factory=Qwen3TextArchConfig) + prefix: str = "qwen3" diff --git a/python/sglang/multimodal_gen/configs/pipeline_configs/__init__.py b/python/sglang/multimodal_gen/configs/pipeline_configs/__init__.py index 65324c5d0..48e2ffe25 100644 --- a/python/sglang/multimodal_gen/configs/pipeline_configs/__init__.py +++ b/python/sglang/multimodal_gen/configs/pipeline_configs/__init__.py @@ -7,7 +7,11 @@ from sglang.multimodal_gen.configs.pipeline_configs.base import ( from sglang.multimodal_gen.configs.pipeline_configs.diffusers_generic import ( DiffusersGenericPipelineConfig, ) -from sglang.multimodal_gen.configs.pipeline_configs.flux import FluxPipelineConfig +from sglang.multimodal_gen.configs.pipeline_configs.flux import ( + Flux2KleinPipelineConfig, + Flux2PipelineConfig, + FluxPipelineConfig, +) from sglang.multimodal_gen.configs.pipeline_configs.flux_finetuned import ( Flux2FinetunedPipelineConfig, ) @@ -29,6 +33,8 @@ __all__ = [ "HunyuanConfig", "FastHunyuanConfig", "FluxPipelineConfig", + "Flux2PipelineConfig", + "Flux2KleinPipelineConfig", "Flux2FinetunedPipelineConfig", "PipelineConfig", "SlidingTileAttnConfig", diff --git a/python/sglang/multimodal_gen/configs/pipeline_configs/flux.py b/python/sglang/multimodal_gen/configs/pipeline_configs/flux.py index 22ebf4f58..a5c58e684 100644 --- a/python/sglang/multimodal_gen/configs/pipeline_configs/flux.py +++ b/python/sglang/multimodal_gen/configs/pipeline_configs/flux.py @@ -15,6 +15,7 @@ from sglang.multimodal_gen.configs.models.encoders import ( TextEncoderConfig, ) from sglang.multimodal_gen.configs.models.encoders.base import TextEncoderArchConfig +from sglang.multimodal_gen.configs.models.encoders.qwen3 import Qwen3TextConfig from sglang.multimodal_gen.configs.models.encoders.qwen_image import ( _is_transformer_layer, ) @@ -340,6 +341,20 @@ def flux2_postprocess_text(outputs: BaseEncoderOutput, _text_inputs) -> torch.Te return prompt_embeds +def flux2_klein_postprocess_text( + outputs: BaseEncoderOutput, _text_inputs +) -> torch.Tensor: + hidden_states_layers: list[int] = [9, 18, 27] + + out = torch.stack([outputs.hidden_states[k] for k in hidden_states_layers], dim=1) + batch_size, num_channels, seq_len, hidden_dim = out.shape + prompt_embeds = out.permute(0, 2, 1, 3).reshape( + batch_size, seq_len, num_channels * hidden_dim + ) + + return prompt_embeds + + @dataclass class Flux2MistralTextArchConfig(TextEncoderArchConfig): stacked_params_mapping: list[tuple[str, str, str]] = field( @@ -607,3 +622,61 @@ class Flux2PipelineConfig(FluxPipelineConfig): # remove noise over input image noise = noise[:, : latents.size(1) :] return noise + + +@dataclass +class Flux2KleinPipelineConfig(Flux2PipelineConfig): + # Klein is distilled, so no guidance embeddings + should_use_guidance: bool = False + task_type: ModelTaskType = ModelTaskType.T2I + + text_encoder_precisions: tuple[str, ...] = field(default_factory=lambda: ("bf16",)) + + text_encoder_configs: tuple[EncoderConfig, ...] = field( + default_factory=lambda: (Qwen3TextConfig(),) + ) + + preprocess_text_funcs: tuple[Callable[[str], str], ...] = field( + default_factory=lambda: (preprocess_text,), + ) + + postprocess_text_funcs: tuple[Callable[[str], str], ...] = field( + default_factory=lambda: (flux2_klein_postprocess_text,) + ) + + def tokenize_prompt(self, prompts: list[str], tokenizer, tok_kwargs) -> dict: + if prompts and isinstance(prompts[0], list): + prompts = [p for prompt in prompts for p in prompt] + + def _apply_chat_template(prompt: str) -> str: + messages = [{"role": "user", "content": prompt}] + try: + return tokenizer.apply_chat_template( + messages, + tokenize=False, + add_generation_prompt=True, + enable_thinking=False, + ) + except TypeError: + return tokenizer.apply_chat_template( + messages, + tokenize=False, + add_generation_prompt=True, + ) + + texts = [_apply_chat_template(prompt) for prompt in prompts] + + tok_kwargs = dict(tok_kwargs or {}) + max_length = tok_kwargs.pop("max_length", 512) + padding = tok_kwargs.pop("padding", "max_length") + truncation = tok_kwargs.pop("truncation", True) + return_tensors = tok_kwargs.pop("return_tensors", "pt") + + return tokenizer( + texts, + padding=padding, + truncation=truncation, + max_length=max_length, + return_tensors=return_tensors, + **tok_kwargs, + ) diff --git a/python/sglang/multimodal_gen/docs/support_matrix.md b/python/sglang/multimodal_gen/docs/support_matrix.md index d7af4756f..2f9baec13 100644 --- a/python/sglang/multimodal_gen/docs/support_matrix.md +++ b/python/sglang/multimodal_gen/docs/support_matrix.md @@ -38,12 +38,16 @@ default parameters when initializing and generating videos. ### Image Generation Models -| Model Name | HuggingFace Model ID | Resolutions | -|:----------------|:-------------------------------|:---------------| -| FLUX.1-dev | `black-forest-labs/FLUX.1-dev` | Any resolution | -| FLUX.2-dev | `black-forest-labs/FLUX.2-dev` | Any resolution | -| Qwen Image | `Qwen/Qwen-Image` | Any resolution | -| Qwen Image Edit | `Qwen/Qwen-Image-Edit` | Any resolution | +| Model Name | HuggingFace Model ID | Resolutions | +|:-----------------|:----------------------------------------|:---------------| +| FLUX.1-dev | `black-forest-labs/FLUX.1-dev` | Any resolution | +| FLUX.2-dev | `black-forest-labs/FLUX.2-dev` | Any resolution | +| FLUX.2-Klein | `black-forest-labs/FLUX.2-klein-4B` | Any resolution | +| Z-Image-Turbo | `Tongyi-MAI/Z-Image-Turbo` | Any resolution | +| GLM-Image | `zai-org/GLM-Image` | Any resolution | +| Qwen Image | `Qwen/Qwen-Image` | Any resolution | +| Qwen Image 2512 | `Qwen/Qwen-Image-2512` | Any resolution | +| Qwen Image Edit | `Qwen/Qwen-Image-Edit` | Any resolution | ## Verified LoRA Examples diff --git a/python/sglang/multimodal_gen/registry.py b/python/sglang/multimodal_gen/registry.py index c244bc8b5..94110ec47 100644 --- a/python/sglang/multimodal_gen/registry.py +++ b/python/sglang/multimodal_gen/registry.py @@ -38,7 +38,10 @@ from sglang.multimodal_gen.configs.pipeline_configs import ( ZImagePipelineConfig, ) from sglang.multimodal_gen.configs.pipeline_configs.base import PipelineConfig -from sglang.multimodal_gen.configs.pipeline_configs.flux import Flux2PipelineConfig +from sglang.multimodal_gen.configs.pipeline_configs.flux import ( + Flux2KleinPipelineConfig, + Flux2PipelineConfig, +) from sglang.multimodal_gen.configs.pipeline_configs.glm_image import ( GlmImagePipelineConfig, ) @@ -534,13 +537,27 @@ def _register_configs(): ], model_detectors=[lambda hf_id: "flux.1" in hf_id.lower()], ) + register_configs( + sampling_param_cls=FluxSamplingParams, + pipeline_config_cls=Flux2KleinPipelineConfig, + hf_model_paths=[ + "black-forest-labs/FLUX.2-klein-4B", + "black-forest-labs/FLUX.2-klein-9B", + ], + model_detectors=[ + lambda hf_id: "flux.2-klein" in hf_id.lower() + or "flux2-klein" in hf_id.lower() + ], + ) register_configs( sampling_param_cls=FluxSamplingParams, pipeline_config_cls=Flux2PipelineConfig, hf_model_paths=[ "black-forest-labs/FLUX.2-dev", ], - model_detectors=[lambda hf_id: "flux.2" in hf_id.lower()], + model_detectors=[ + lambda hf_id: "flux.2" in hf_id.lower() and "klein" not in hf_id.lower() + ], ) register_configs( sampling_param_cls=ZImageSamplingParams, diff --git a/python/sglang/multimodal_gen/runtime/models/dits/flux_2.py b/python/sglang/multimodal_gen/runtime/models/dits/flux_2.py index 411f97ada..533102690 100644 --- a/python/sglang/multimodal_gen/runtime/models/dits/flux_2.py +++ b/python/sglang/multimodal_gen/runtime/models/dits/flux_2.py @@ -550,7 +550,11 @@ class Flux2TransformerBlock(nn.Module): class Flux2TimestepGuidanceEmbeddings(nn.Module): def __init__( - self, in_channels: int = 256, embedding_dim: int = 6144, bias: bool = False + self, + in_channels: int = 256, + embedding_dim: int = 6144, + bias: bool = False, + guidance_embeds: bool = True, ): super().__init__() @@ -561,24 +565,32 @@ class Flux2TimestepGuidanceEmbeddings(nn.Module): in_channels=in_channels, time_embed_dim=embedding_dim, sample_proj_bias=bias ) - self.guidance_embedder = TimestepEmbedding( - in_channels=in_channels, time_embed_dim=embedding_dim, sample_proj_bias=bias - ) + if guidance_embeds: + self.guidance_embedder = TimestepEmbedding( + in_channels=in_channels, + time_embed_dim=embedding_dim, + sample_proj_bias=bias, + ) + else: + self.guidance_embedder = None - def forward(self, timestep: torch.Tensor, guidance: torch.Tensor) -> torch.Tensor: + def forward( + self, timestep: torch.Tensor, guidance: Optional[torch.Tensor] = None + ) -> torch.Tensor: timesteps_proj = self.time_proj(timestep) timesteps_emb = self.timestep_embedder( timesteps_proj.to(timestep.dtype) ) # (N, D) - guidance_proj = self.time_proj(guidance) - guidance_emb = self.guidance_embedder( - guidance_proj.to(guidance.dtype) - ) # (N, D) - - time_guidance_emb = timesteps_emb + guidance_emb - - return time_guidance_emb + if guidance is not None and self.guidance_embedder is not None: + guidance_proj = self.time_proj(guidance) + guidance_emb = self.guidance_embedder( + guidance_proj.to(guidance.dtype) + ) # (N, D) + time_guidance_emb = timesteps_emb + guidance_emb + return time_guidance_emb + else: + return timesteps_emb class Flux2Modulation(nn.Module): @@ -650,8 +662,10 @@ class Flux2Transformer2DModel(CachableDiT, OffloadableDiTMixin): axes_dims_rope: Tuple[int, ...] = config.axes_dims_rope rope_theta: int = config.rope_theta eps: float = config.eps + guidance_embeds: bool = getattr(config, "guidance_embeds", True) self.out_channels = out_channels or in_channels self.inner_dim = num_attention_heads * attention_head_dim + self.guidance_embeds = guidance_embeds # 1. Sinusoidal positional embedding for RoPE on image and text tokens self.rotary_emb = Flux2PosEmbed(theta=rope_theta, axes_dim=axes_dims_rope) @@ -661,6 +675,7 @@ class Flux2Transformer2DModel(CachableDiT, OffloadableDiTMixin): in_channels=timestep_guidance_channels, embedding_dim=self.inner_dim, bias=False, + guidance_embeds=guidance_embeds, ) # 3. Modulation (double stream and single stream blocks share modulation parameters, resp.) @@ -767,7 +782,8 @@ class Flux2Transformer2DModel(CachableDiT, OffloadableDiTMixin): # 1. Calculate timestep embedding and modulation parameters timestep = timestep.to(hidden_states.dtype) - guidance = guidance.to(hidden_states.dtype) + if guidance is not None: + guidance = guidance.to(hidden_states.dtype) temb = self.time_guidance_embed(timestep, guidance) diff --git a/python/sglang/multimodal_gen/runtime/models/encoders/qwen3.py b/python/sglang/multimodal_gen/runtime/models/encoders/qwen3.py new file mode 100644 index 000000000..b8132e404 --- /dev/null +++ b/python/sglang/multimodal_gen/runtime/models/encoders/qwen3.py @@ -0,0 +1,422 @@ +from collections.abc import Iterable +from typing import Any + +import torch +from torch import nn + +from sglang.multimodal_gen.configs.models.encoders import BaseEncoderOutput +from sglang.multimodal_gen.configs.models.encoders.qwen3 import Qwen3TextConfig +from sglang.multimodal_gen.runtime.distributed import get_tp_world_size +from sglang.multimodal_gen.runtime.layers.activation import SiluAndMul +from sglang.multimodal_gen.runtime.layers.attention import LocalAttention +from sglang.multimodal_gen.runtime.layers.layernorm import RMSNorm +from sglang.multimodal_gen.runtime.layers.linear import ( + MergedColumnParallelLinear, + QKVParallelLinear, + RowParallelLinear, +) +from sglang.multimodal_gen.runtime.layers.quantization import QuantizationConfig +from sglang.multimodal_gen.runtime.layers.rotary_embedding import get_rope +from sglang.multimodal_gen.runtime.layers.vocab_parallel_embedding import ( + VocabParallelEmbedding, +) +from sglang.multimodal_gen.runtime.loader.weight_utils import ( + default_weight_loader, + maybe_remap_kv_scale_name, +) +from sglang.multimodal_gen.runtime.models.encoders.base import TextEncoder + + +class Qwen3MLP(nn.Module): + """Qwen3 MLP with SwiGLU activation and tensor parallelism.""" + + def __init__( + self, + hidden_size: int, + intermediate_size: int, + hidden_act: str, + quant_config: QuantizationConfig | None = None, + bias: bool = False, + prefix: str = "", + ) -> None: + super().__init__() + self.gate_up_proj = MergedColumnParallelLinear( + input_size=hidden_size, + output_sizes=[intermediate_size] * 2, + bias=bias, + quant_config=quant_config, + prefix=f"{prefix}.gate_up_proj", + ) + self.down_proj = RowParallelLinear( + input_size=intermediate_size, + output_size=hidden_size, + bias=bias, + quant_config=quant_config, + prefix=f"{prefix}.down_proj", + ) + if hidden_act != "silu": + raise ValueError( + f"Unsupported activation: {hidden_act}. Only silu is supported." + ) + self.act_fn = SiluAndMul() + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x, _ = self.gate_up_proj(x) + x = self.act_fn(x) + x, _ = self.down_proj(x) + return x + + +class Qwen3Attention(nn.Module): + """Qwen3 attention with QK-Norm and tensor parallelism. + + Key difference from LLaMA: RMSNorm is applied to Q and K before attention. + """ + + def __init__( + self, + config: Qwen3TextConfig, + hidden_size: int, + num_heads: int, + num_kv_heads: int, + rope_theta: float = 1000000.0, + rope_scaling: dict[str, Any] | None = None, + max_position_embeddings: int = 40960, + quant_config: QuantizationConfig | None = None, + bias: bool = False, + prefix: str = "", + ) -> None: + super().__init__() + self.hidden_size = hidden_size + tp_size = get_tp_world_size() + self.total_num_heads = num_heads + assert self.total_num_heads % tp_size == 0 + self.num_heads = self.total_num_heads // tp_size + self.total_num_kv_heads = num_kv_heads + if self.total_num_kv_heads >= tp_size: + assert self.total_num_kv_heads % tp_size == 0 + else: + assert tp_size % self.total_num_kv_heads == 0 + self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) + + self.head_dim = getattr( + config, "head_dim", self.hidden_size // self.total_num_heads + ) + self.rotary_dim = self.head_dim + self.q_size = self.num_heads * self.head_dim + self.kv_size = self.num_kv_heads * self.head_dim + self.scaling = self.head_dim**-0.5 + self.rope_theta = rope_theta + self.max_position_embeddings = max_position_embeddings + + # QKV projection with tensor parallelism + self.qkv_proj = QKVParallelLinear( + hidden_size=hidden_size, + head_size=self.head_dim, + total_num_heads=self.total_num_heads, + total_num_kv_heads=self.total_num_kv_heads, + bias=bias, + quant_config=quant_config, + prefix=f"{prefix}.qkv_proj", + ) + + # Output projection + self.o_proj = RowParallelLinear( + input_size=self.total_num_heads * self.head_dim, + output_size=hidden_size, + bias=bias, + quant_config=quant_config, + prefix=f"{prefix}.o_proj", + ) + + # QK-Norm: Key difference from LLaMA + rms_norm_eps = getattr(config, "rms_norm_eps", 1e-6) + self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps) + self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps) + + # Rotary embeddings + self.rotary_emb = get_rope( + self.head_dim, + rotary_dim=self.rotary_dim, + max_position=max_position_embeddings, + base=int(rope_theta), + rope_scaling=rope_scaling, + is_neox_style=True, + ) + + # Attention with FlashAttention/SageAttn support + self.attn = LocalAttention( + self.num_heads, + self.head_dim, + self.num_kv_heads, + softmax_scale=self.scaling, + causal=True, + supported_attention_backends=config._supported_attention_backends, + ) + + def forward( + self, + positions: torch.Tensor, + hidden_states: torch.Tensor, + ) -> torch.Tensor: + # QKV projection + qkv, _ = self.qkv_proj(hidden_states) + q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) + + # Reshape for QK-norm + batch_size, seq_len = q.shape[0], q.shape[1] + q = q.reshape(batch_size, seq_len, self.num_heads, self.head_dim) + k = k.reshape(batch_size, seq_len, self.num_kv_heads, self.head_dim) + v = v.reshape(batch_size, seq_len, self.num_kv_heads, self.head_dim) + + # Apply QK-Norm (key difference from LLaMA) + q = self.q_norm(q) + k = self.k_norm(k) + + # Reshape back for rotary embeddings + q = q.reshape(batch_size, seq_len, -1) + k = k.reshape(batch_size, seq_len, -1) + + # Apply rotary embeddings + q, k = self.rotary_emb(positions, q, k) + + # Reshape for attention + q = q.reshape(batch_size, seq_len, self.num_heads, self.head_dim) + k = k.reshape(batch_size, seq_len, self.num_kv_heads, self.head_dim) + + # Attention + attn_output = self.attn(q, k, v) + attn_output = attn_output.reshape(batch_size, seq_len, -1) + + # Output projection + output, _ = self.o_proj(attn_output) + return output + + +class Qwen3DecoderLayer(nn.Module): + """Qwen3 transformer decoder layer.""" + + def __init__( + self, + config: Qwen3TextConfig, + quant_config: QuantizationConfig | None = None, + prefix: str = "", + ) -> None: + super().__init__() + self.hidden_size = config.hidden_size + rope_theta = getattr(config, "rope_theta", 1000000.0) + rope_scaling = getattr(config, "rope_scaling", None) + max_position_embeddings = getattr(config, "max_position_embeddings", 40960) + attention_bias = getattr(config, "attention_bias", False) + + self.self_attn = Qwen3Attention( + config=config, + hidden_size=self.hidden_size, + num_heads=config.num_attention_heads, + num_kv_heads=getattr( + config, "num_key_value_heads", config.num_attention_heads + ), + rope_theta=rope_theta, + rope_scaling=rope_scaling, + max_position_embeddings=max_position_embeddings, + quant_config=quant_config, + bias=attention_bias, + prefix=f"{prefix}.self_attn", + ) + self.mlp = Qwen3MLP( + hidden_size=self.hidden_size, + intermediate_size=config.intermediate_size, + hidden_act=config.hidden_act, + quant_config=quant_config, + bias=getattr(config, "mlp_bias", False), + prefix=f"{prefix}.mlp", + ) + self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = RMSNorm( + config.hidden_size, eps=config.rms_norm_eps + ) + + def forward( + self, + positions: torch.Tensor, + hidden_states: torch.Tensor, + residual: torch.Tensor | None, + ) -> tuple[torch.Tensor, torch.Tensor]: + # Self Attention + if residual is None: + residual = hidden_states + hidden_states = self.input_layernorm(hidden_states) + else: + hidden_states, residual = self.input_layernorm(hidden_states, residual) + + hidden_states = self.self_attn(positions=positions, hidden_states=hidden_states) + + # MLP + hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) + hidden_states = self.mlp(hidden_states) + return hidden_states, residual + + +class Qwen3ForCausalLM(TextEncoder): + """Qwen3 causal language model for text encoding in diffusion models. + + Features: + - Tensor parallelism support + - FlashAttention/SageAttn/SDPA support via LocalAttention + - QK-Norm for better training stability + - FSDP sharding for CPU offload + """ + + def __init__(self, config: Qwen3TextConfig) -> None: + super().__init__(config) + + self.config = config + self.quant_config = config.quant_config + + # Embedding layer with tensor parallelism + if config.lora_config is not None: + max_loras = getattr(config.lora_config, "max_loras", 1) + lora_vocab_size = getattr(config.lora_config, "lora_extra_vocab_size", 1) + lora_vocab = lora_vocab_size * max_loras + else: + lora_vocab = 0 + self.vocab_size = config.vocab_size + lora_vocab + self.org_vocab_size = config.vocab_size + + self.embed_tokens = VocabParallelEmbedding( + self.vocab_size, + config.hidden_size, + org_num_embeddings=config.vocab_size, + quant_config=config.quant_config, + ) + + # Transformer layers + self.layers = nn.ModuleList( + [ + Qwen3DecoderLayer( + config=config, + quant_config=config.quant_config, + prefix=f"{config.prefix}.layers.{i}", + ) + for i in range(config.num_hidden_layers) + ] + ) + + # Final layer norm + self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor: + return self.embed_tokens(input_ids) + + def forward( + self, + input_ids: torch.Tensor | None = None, + position_ids: torch.Tensor | None = None, + attention_mask: torch.Tensor | None = None, + inputs_embeds: torch.Tensor | None = None, + output_hidden_states: bool | None = None, + **kwargs, + ) -> BaseEncoderOutput: + output_hidden_states = ( + output_hidden_states + if output_hidden_states is not None + else self.config.output_hidden_states + ) + + if inputs_embeds is not None: + hidden_states = inputs_embeds + else: + hidden_states = self.get_input_embeddings(input_ids) + + residual = None + + if position_ids is None: + position_ids = torch.arange( + 0, hidden_states.shape[1], device=hidden_states.device + ).unsqueeze(0) + + all_hidden_states: tuple[Any, ...] | None = () if output_hidden_states else None + + for layer in self.layers: + if all_hidden_states is not None: + all_hidden_states += ( + (hidden_states,) + if residual is None + else (hidden_states + residual,) + ) + hidden_states, residual = layer(position_ids, hidden_states, residual) + + hidden_states, _ = self.norm(hidden_states, residual) + + # Add hidden states from the last decoder layer + if all_hidden_states is not None: + all_hidden_states += (hidden_states,) + + return BaseEncoderOutput( + last_hidden_state=hidden_states, + hidden_states=all_hidden_states, + ) + + def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: + """Load weights with support for tensor parallelism and weight remapping.""" + params_dict = dict(self.named_parameters()) + loaded_params: set[str] = set() + + for name, loaded_weight in weights: + # Strip 'model.' prefix from HuggingFace Qwen3 weights + if name.startswith("model."): + name = name[6:] # len("model.") == 6 + + # Skip rotary embedding weights + if "rotary_emb.inv_freq" in name: + continue + if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name: + continue + + # Handle KV scale remapping + if "scale" in name: + kv_scale_name: str | None = maybe_remap_kv_scale_name(name, params_dict) + if kv_scale_name is None: + continue + else: + name = kv_scale_name + + # Handle stacked params mapping (qkv_proj, gate_up_proj) + for ( + param_name, + weight_name, + shard_id, + ) in self.config.arch_config.stacked_params_mapping: + if weight_name not in name: + continue + name = name.replace(weight_name, param_name) + + # Skip loading extra bias for GPTQ models + if name.endswith(".bias") and name not in params_dict: + continue + + if name not in params_dict: + continue + + param = params_dict[name] + weight_loader = param.weight_loader + weight_loader(param, loaded_weight, shard_id) + break + else: + # Skip loading extra bias for GPTQ models + if name.endswith(".bias") and name not in params_dict: + continue + + if name not in params_dict: + continue + + param = params_dict[name] + weight_loader = getattr(param, "weight_loader", default_weight_loader) + weight_loader(param, loaded_weight) + + loaded_params.add(name) + + return loaded_params + + +EntryClass = Qwen3ForCausalLM diff --git a/python/sglang/multimodal_gen/runtime/pipelines/flux_2_klein.py b/python/sglang/multimodal_gen/runtime/pipelines/flux_2_klein.py new file mode 100644 index 000000000..ff2d6d7d9 --- /dev/null +++ b/python/sglang/multimodal_gen/runtime/pipelines/flux_2_klein.py @@ -0,0 +1,8 @@ +from sglang.multimodal_gen.runtime.pipelines.flux_2 import Flux2Pipeline + + +class Flux2KleinPipeline(Flux2Pipeline): + pipeline_name = "Flux2KleinPipeline" + + +EntryClass = Flux2KleinPipeline