Support sd3.5 medium and MMDiT-X (#2587)

* extract attn out of joint_attn

* further adjust attn and joint_attn

* add mmdit-x support

* support sd3.5-medium in the example

* update README.md

candle-examples/examples/stable-diffusion-3/README.md

@@ -1,18 +1,26 @@
-# candle-stable-diffusion-3: Candle Implementation of Stable Diffusion 3 Medium
+# candle-stable-diffusion-3: Candle Implementation of Stable Diffusion 3/3.5
 
 ![](assets/stable-diffusion-3.jpg)
 
-*A cute rusty robot holding a candle torch in its hand, with glowing neon text \"LETS GO RUSTY\" displayed on its chest, bright background, high quality, 4k*
+*A cute rusty robot holding a candle torch in its hand, with glowing neon text \"LETS GO RUSTY\" displayed on its chest, bright background, high quality, 4k*, generated by Stable Diffusion 3 Medium
 
 Stable Diffusion 3 Medium is a text-to-image model based on Multimodal Diffusion Transformer (MMDiT) architecture.
 
 - [huggingface repo](https://huggingface.co/stabilityai/stable-diffusion-3-medium)
 - [research paper](https://arxiv.org/pdf/2403.03206)
 - [announcement blog post](https://stability.ai/news/stable-diffusion-3-medium)
 
+Stable Diffusion 3.5 is a family of text-to-image models with latest improvements:
+- [announcement blog post](https://stability.ai/news/introducing-stable-diffusion-3-5)
+
+It has three variants:
+- [Stable Diffusion 3.5 Large](https://huggingface.co/stabilityai/stable-diffusion-3.5-large) @ 8.1b params, with scaled and slightly modified MMDiT architecture.
+- [Stable Diffusion 3.5 Large Turbo](https://huggingface.co/stabilityai/stable-diffusion-3.5-large-turbo) distilled version that enables 4-step inference.
+- [Stable Diffusion 3.5 Medium](https://huggingface.co/stabilityai/stable-diffusion-3.5-medium) @ 2.5b params, with improved MMDiT-X architecture.
+
 ## Getting access to the weights
 
-The weights of Stable Diffusion 3 Medium is released by Stability AI under the Stability Community License. You will need to accept the conditions and acquire a license by visiting [the repo on HuggingFace Hub](https://huggingface.co/stabilityai/stable-diffusion-3-medium) to gain access to the weights for your HuggingFace account.
+The weights of Stable Diffusion 3/3.5 is released by Stability AI under the Stability Community License. You will need to accept the conditions and acquire a license by visiting the repos on HuggingFace Hub to gain access to the weights for your HuggingFace account.
 
 To allow your computer to gain access to the public-gated repos on HuggingFace, you might need to create a [HuggingFace User Access Tokens](https://huggingface.co/docs/hub/en/security-tokens) (recommended) and log in on your computer if you haven't done that before. A convenient way to do the login is to use [huggingface-cli](https://huggingface.co/docs/huggingface_hub/en/guides/cli):
 
@@ -27,10 +35,12 @@ On the first run, the weights will be automatically downloaded from the Huggingf
 
 ```shell
 cargo run --example stable-diffusion-3 --release --features=cuda -- \
-  --height 1024 --width 1024 \
+  --which 3-medium --height 1024 --width 1024 \
   --prompt 'A cute rusty robot holding a candle torch in its hand, with glowing neon text \"LETS GO RUSTY\" displayed on its chest, bright background, high quality, 4k'
 ```
 
+To use different models, changed the value of `--which` option. (Possible values: `3-medium`, `3.5-large`, `3.5-large-turbo` and `3.5-medium`).
+
 To display other options available,
 
 ```shell
@@ -45,7 +55,7 @@ cargo run --example stable-diffusion-3 --release --features=cuda,flash-attn -- -
 
 ## Performance Benchmark
 
-Below benchmark is done by generating 1024-by-1024 image from 28 steps of Euler sampling and measure the average speed (iteration per seconds).
+Below benchmark is done with Stable Diffusion 3 Medium by generating 1024-by-1024 image from 28 steps of Euler sampling and measure the average speed (iteration per seconds).
 
 [candle](https://github.com/huggingface/candle) and [candle-flash-attn](https://github.com/huggingface/candle/tree/main/candle-flash-attn) is based on the commit of [0d96ec3](https://github.com/huggingface/candle/commit/0d96ec31e8be03f844ed0aed636d6217dee9c7bc).
 

candle-examples/examples/stable-diffusion-3/main.rs

@@ -19,13 +19,15 @@ enum Which {
     V3_5Large,
     #[value(name = "3.5-large-turbo")]
     V3_5LargeTurbo,
+    #[value(name = "3.5-medium")]
+    V3_5Medium,
 }
 
 impl Which {
     fn is_3_5(&self) -> bool {
         match self {
             Self::V3Medium => false,
-            Self::V3_5Large | Self::V3_5LargeTurbo => true,
+            Self::V3_5Large | Self::V3_5LargeTurbo | Self::V3_5Medium => true,
         }
     }
 }
@@ -117,36 +119,59 @@ fn main() -> Result<()> {
     let default_inference_steps = match which {
         Which::V3_5Large => 28,
         Which::V3_5LargeTurbo => 4,
+        Which::V3_5Medium => 28,
         Which::V3Medium => 28,
     };
     let num_inference_steps = num_inference_steps.unwrap_or(default_inference_steps);
     let default_cfg_scale = match which {
         Which::V3_5Large => 4.0,
         Which::V3_5LargeTurbo => 1.0,
+        Which::V3_5Medium => 4.0,
         Which::V3Medium => 4.0,
     };
     let cfg_scale = cfg_scale.unwrap_or(default_cfg_scale);
 
     let api = hf_hub::api::sync::Api::new()?;
     let (mmdit_config, mut triple, vb) = if which.is_3_5() {
-        let sai_repo = {
+        let sai_repo_for_text_encoders = {
+            let name = match which {
+                Which::V3_5Large => "stabilityai/stable-diffusion-3.5-large",
+                Which::V3_5LargeTurbo => "stabilityai/stable-diffusion-3.5-large-turbo",
+
+                // Unfortunately, stabilityai/stable-diffusion-3.5-medium doesn't have the monolithic text encoders that's usually
+                // placed under the text_encoders directory, like the case in stabilityai/stable-diffusion-3.5-large and -large-turbo.
+                // To make things worse, it currently only has partitioned model.fp16-00001-of-00002.safetensors and model.fp16-00002-of-00002.safetensors
+                // under the text_encoder_3 directory, for the t5xxl_fp16.safetensors model. This means that we need to merge the two partitions
+                // to get the monolithic text encoders. This is not a trivial task.
+                // Since the situation can change, we do not want to spend efforts to handle the uniqueness of stabilityai/stable-diffusion-3.5-medium,
+                // which involves different paths and merging the two partitions files for t5xxl_fp16.safetensors.
+                // so for now, we'll use the text encoder models from the stabilityai/stable-diffusion-3.5-large repository.
+                // TODO: Change to "stabilityai/stable-diffusion-3.5-medium" once the maintainers of the repository add back the monolithic text encoders.
+                Which::V3_5Medium => "stabilityai/stable-diffusion-3.5-large",
+                Which::V3Medium => unreachable!(),
+            };
+            api.repo(hf_hub::Repo::model(name.to_string()))
+        };
+        let sai_repo_for_mmdit = {
             let name = match which {
                 Which::V3_5Large => "stabilityai/stable-diffusion-3.5-large",
                 Which::V3_5LargeTurbo => "stabilityai/stable-diffusion-3.5-large-turbo",
+                Which::V3_5Medium => "stabilityai/stable-diffusion-3.5-medium",
                 Which::V3Medium => unreachable!(),
             };
             api.repo(hf_hub::Repo::model(name.to_string()))
         };
-        let clip_g_file = sai_repo.get("text_encoders/clip_g.safetensors")?;
-        let clip_l_file = sai_repo.get("text_encoders/clip_l.safetensors")?;
-        let t5xxl_file = sai_repo.get("text_encoders/t5xxl_fp16.safetensors")?;
+        let clip_g_file = sai_repo_for_text_encoders.get("text_encoders/clip_g.safetensors")?;
+        let clip_l_file = sai_repo_for_text_encoders.get("text_encoders/clip_l.safetensors")?;
+        let t5xxl_file = sai_repo_for_text_encoders.get("text_encoders/t5xxl_fp16.safetensors")?;
         let model_file = {
             let model_file = match which {
                 Which::V3_5Large => "sd3.5_large.safetensors",
                 Which::V3_5LargeTurbo => "sd3.5_large_turbo.safetensors",
+                Which::V3_5Medium => "sd3.5_medium.safetensors",
                 Which::V3Medium => unreachable!(),
             };
-            sai_repo.get(model_file)?
+            sai_repo_for_mmdit.get(model_file)?
         };
         let triple = StableDiffusion3TripleClipWithTokenizer::new_split(
             &clip_g_file,
@@ -157,7 +182,12 @@ fn main() -> Result<()> {
         let vb = unsafe {
             candle_nn::VarBuilder::from_mmaped_safetensors(&[model_file], DType::F16, &device)?
         };
-        (MMDiTConfig::sd3_5_large(), triple, vb)
+        match which {
+            Which::V3_5Large => (MMDiTConfig::sd3_5_large(), triple, vb),
+            Which::V3_5LargeTurbo => (MMDiTConfig::sd3_5_large(), triple, vb),
+            Which::V3_5Medium => (MMDiTConfig::sd3_5_medium(), triple, vb),
+            Which::V3Medium => unreachable!(),
+        }
     } else {
         let sai_repo = {
             let name = "stabilityai/stable-diffusion-3-medium";

candle-transformers/src/models/mmdit/blocks.rs

@@ -36,7 +36,6 @@ impl Module for LayerNormNoAffine {
 
 impl DiTBlock {
     pub fn new(hidden_size: usize, num_heads: usize, vb: nn::VarBuilder) -> Result<Self> {
-        // {'hidden_size': 1536, 'num_heads': 24}
         let norm1 = LayerNormNoAffine::new(1e-6);
         let attn = AttnProjections::new(hidden_size, num_heads, vb.pp("attn"))?;
         let norm2 = LayerNormNoAffine::new(1e-6);
@@ -103,6 +102,117 @@ impl DiTBlock {
     }
 }
 
+pub struct SelfAttnModulateIntermediates {
+    gate_msa: Tensor,
+    shift_mlp: Tensor,
+    scale_mlp: Tensor,
+    gate_mlp: Tensor,
+    gate_msa2: Tensor,
+}
+
+pub struct SelfAttnDiTBlock {
+    norm1: LayerNormNoAffine,
+    attn: AttnProjections,
+    attn2: AttnProjections,
+    norm2: LayerNormNoAffine,
+    mlp: Mlp,
+    ada_ln_modulation: nn::Sequential,
+}
+
+impl SelfAttnDiTBlock {
+    pub fn new(hidden_size: usize, num_heads: usize, vb: nn::VarBuilder) -> Result<Self> {
+        let norm1 = LayerNormNoAffine::new(1e-6);
+        let attn = AttnProjections::new(hidden_size, num_heads, vb.pp("attn"))?;
+        let attn2 = AttnProjections::new(hidden_size, num_heads, vb.pp("attn2"))?;
+        let norm2 = LayerNormNoAffine::new(1e-6);
+        let mlp_ratio = 4;
+        let mlp = Mlp::new(hidden_size, hidden_size * mlp_ratio, vb.pp("mlp"))?;
+        let n_mods = 9;
+        let ada_ln_modulation = nn::seq().add(nn::Activation::Silu).add(nn::linear(
+            hidden_size,
+            n_mods * hidden_size,
+            vb.pp("adaLN_modulation.1"),
+        )?);
+
+        Ok(Self {
+            norm1,
+            attn,
+            attn2,
+            norm2,
+            mlp,
+            ada_ln_modulation,
+        })
+    }
+
+    pub fn pre_attention(
+        &self,
+        x: &Tensor,
+        c: &Tensor,
+    ) -> Result<(Qkv, Qkv, SelfAttnModulateIntermediates)> {
+        let modulation = self.ada_ln_modulation.forward(c)?;
+        let chunks = modulation.chunk(9, D::Minus1)?;
+        let (
+            shift_msa,
+            scale_msa,
+            gate_msa,
+            shift_mlp,
+            scale_mlp,
+            gate_mlp,
+            shift_msa2,
+            scale_msa2,
+            gate_msa2,
+        ) = (
+            chunks[0].clone(),
+            chunks[1].clone(),
+            chunks[2].clone(),
+            chunks[3].clone(),
+            chunks[4].clone(),
+            chunks[5].clone(),
+            chunks[6].clone(),
+            chunks[7].clone(),
+            chunks[8].clone(),
+        );
+
+        let norm_x = self.norm1.forward(x)?;
+        let modulated_x = modulate(&norm_x, &shift_msa, &scale_msa)?;
+        let qkv = self.attn.pre_attention(&modulated_x)?;
+
+        let modulated_x2 = modulate(&norm_x, &shift_msa2, &scale_msa2)?;
+        let qkv2 = self.attn2.pre_attention(&modulated_x2)?;
+
+        Ok((
+            qkv,
+            qkv2,
+            SelfAttnModulateIntermediates {
+                gate_msa,
+                shift_mlp,
+                scale_mlp,
+                gate_mlp,
+                gate_msa2,
+            },
+        ))
+    }
+
+    pub fn post_attention(
+        &self,
+        attn: &Tensor,
+        attn2: &Tensor,
+        x: &Tensor,
+        mod_interm: &SelfAttnModulateIntermediates,
+    ) -> Result<Tensor> {
+        let attn_out = self.attn.post_attention(attn)?;
+        let x = x.add(&attn_out.broadcast_mul(&mod_interm.gate_msa.unsqueeze(1)?)?)?;
+        let attn_out2 = self.attn2.post_attention(attn2)?;
+        let x = x.add(&attn_out2.broadcast_mul(&mod_interm.gate_msa2.unsqueeze(1)?)?)?;
+
+        let norm_x = self.norm2.forward(&x)?;
+        let modulated_x = modulate(&norm_x, &mod_interm.shift_mlp, &mod_interm.scale_mlp)?;
+        let mlp_out = self.mlp.forward(&modulated_x)?;
+        let x = x.add(&mlp_out.broadcast_mul(&mod_interm.gate_mlp.unsqueeze(1)?)?)?;
+        Ok(x)
+    }
+}
+
 pub struct QkvOnlyDiTBlock {
     norm1: LayerNormNoAffine,
     attn: QkvOnlyAttnProjections,
@@ -190,14 +300,18 @@ fn modulate(x: &Tensor, shift: &Tensor, scale: &Tensor) -> Result<Tensor> {
     shift.broadcast_add(&x.broadcast_mul(&scale_plus_one)?)
 }
 
-pub struct JointBlock {
+pub trait JointBlock {
+    fn forward(&self, context: &Tensor, x: &Tensor, c: &Tensor) -> Result<(Tensor, Tensor)>;
+}
+
+pub struct MMDiTJointBlock {
     x_block: DiTBlock,
     context_block: DiTBlock,
     num_heads: usize,
     use_flash_attn: bool,
 }
 
-impl JointBlock {
+impl MMDiTJointBlock {
     pub fn new(
         hidden_size: usize,
         num_heads: usize,
@@ -214,8 +328,10 @@ impl JointBlock {
             use_flash_attn,
         })
     }
+}
 
-    pub fn forward(&self, context: &Tensor, x: &Tensor, c: &Tensor) -> Result<(Tensor, Tensor)> {
+impl JointBlock for MMDiTJointBlock {
+    fn forward(&self, context: &Tensor, x: &Tensor, c: &Tensor) -> Result<(Tensor, Tensor)> {
         let (context_qkv, context_interm) = self.context_block.pre_attention(context, c)?;
         let (x_qkv, x_interm) = self.x_block.pre_attention(x, c)?;
         let (context_attn, x_attn) =
@@ -228,6 +344,49 @@ impl JointBlock {
     }
 }
 
+pub struct MMDiTXJointBlock {
+    x_block: SelfAttnDiTBlock,
+    context_block: DiTBlock,
+    num_heads: usize,
+    use_flash_attn: bool,
+}
+
+impl MMDiTXJointBlock {
+    pub fn new(
+        hidden_size: usize,
+        num_heads: usize,
+        use_flash_attn: bool,
+        vb: nn::VarBuilder,
+    ) -> Result<Self> {
+        let x_block = SelfAttnDiTBlock::new(hidden_size, num_heads, vb.pp("x_block"))?;
+        let context_block = DiTBlock::new(hidden_size, num_heads, vb.pp("context_block"))?;
+
+        Ok(Self {
+            x_block,
+            context_block,
+            num_heads,
+            use_flash_attn,
+        })
+    }
+}
+
+impl JointBlock for MMDiTXJointBlock {
+    fn forward(&self, context: &Tensor, x: &Tensor, c: &Tensor) -> Result<(Tensor, Tensor)> {
+        let (context_qkv, context_interm) = self.context_block.pre_attention(context, c)?;
+        let (x_qkv, x_qkv2, x_interm) = self.x_block.pre_attention(x, c)?;
+        let (context_attn, x_attn) =
+            joint_attn(&context_qkv, &x_qkv, self.num_heads, self.use_flash_attn)?;
+        let x_attn2 = attn(&x_qkv2, self.num_heads, self.use_flash_attn)?;
+        let context_out =
+            self.context_block
+                .post_attention(&context_attn, context, &context_interm)?;
+        let x_out = self
+            .x_block
+            .post_attention(&x_attn, &x_attn2, x, &x_interm)?;
+        Ok((context_out, x_out))
+    }
+}
+
 pub struct ContextQkvOnlyJointBlock {
     x_block: DiTBlock,
     context_block: QkvOnlyDiTBlock,
@@ -309,26 +468,30 @@ fn joint_attn(
         v: Tensor::cat(&[&context_qkv.v, &x_qkv.v], 1)?,
     };
 
-    let (batch_size, seqlen, _) = qkv.q.dims3()?;
+    let seqlen = qkv.q.dim(1)?;
+    let attn = attn(&qkv, num_heads, use_flash_attn)?;
+    let context_qkv_seqlen = context_qkv.q.dim(1)?;
+    let context_attn = attn.narrow(1, 0, context_qkv_seqlen)?;
+    let x_attn = attn.narrow(1, context_qkv_seqlen, seqlen - context_qkv_seqlen)?;
+
+    Ok((context_attn, x_attn))
+}
+
+fn attn(qkv: &Qkv, num_heads: usize, use_flash_attn: bool) -> Result<Tensor> {
+    let batch_size = qkv.q.dim(0)?;
+    let seqlen = qkv.q.dim(1)?;
     let qkv = Qkv {
         q: qkv.q.reshape((batch_size, seqlen, num_heads, ()))?,
         k: qkv.k.reshape((batch_size, seqlen, num_heads, ()))?,
-        v: qkv.v,
+        v: qkv.v.clone(),
     };
 
     let headdim = qkv.q.dim(D::Minus1)?;
     let softmax_scale = 1.0 / (headdim as f64).sqrt();
-
     let attn = if use_flash_attn {
         flash_attn(&qkv.q, &qkv.k, &qkv.v, softmax_scale as f32, false)?
     } else {
         flash_compatible_attention(&qkv.q, &qkv.k, &qkv.v, softmax_scale as f32)?
     };
-
-    let attn = attn.reshape((batch_size, seqlen, ()))?;
-    let context_qkv_seqlen = context_qkv.q.dim(1)?;
-    let context_attn = attn.narrow(1, 0, context_qkv_seqlen)?;
-    let x_attn = attn.narrow(1, context_qkv_seqlen, seqlen - context_qkv_seqlen)?;
-
-    Ok((context_attn, x_attn))
+    attn.reshape((batch_size, seqlen, ()))
 }