Lazy upcasting for t5.

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

@@ -118,7 +118,7 @@ impl T5WithTokenizer {
             .to_vec();
         tokens.resize(self.max_position_embeddings, 0);
         let input_token_ids = Tensor::new(&tokens[..], device)?.unsqueeze(0)?;
-        let embeddings = self.t5.forward(&input_token_ids)?;
+        let embeddings = self.t5.forward_dt(&input_token_ids, Some(DType::F32))?;
         Ok(embeddings)
     }
 }
@@ -144,7 +144,7 @@ impl StableDiffusion3TripleClipWithTokenizer {
             candle_nn::VarBuilder::from_mmaped_safetensors(&[clip_l_file], DType::F16, device)?
         };
         let vb_t5 = unsafe {
-            candle_nn::VarBuilder::from_mmaped_safetensors(&[t5xxl_file], DType::F32, device)?
+            candle_nn::VarBuilder::from_mmaped_safetensors(&[t5xxl_file], DType::F16, device)?
         };
         let max_position_embeddings = 77usize;
         let clip_l = ClipWithTokenizer::new(
@@ -164,11 +164,6 @@ impl StableDiffusion3TripleClipWithTokenizer {
             max_position_embeddings,
         )?;
 
-        // Current T5 implementation does not support fp16, so we use fp32 VarBuilder for T5.
-        // This is a temporary workaround until the T5 implementation is updated to support fp16.
-        // Also see:
-        // https://github.com/huggingface/candle/issues/2480
-        // https://github.com/huggingface/candle/pull/2481
         let t5 = T5WithTokenizer::new(vb_t5, max_position_embeddings)?;
         Ok(Self {
             clip_l,
@@ -178,34 +173,26 @@ impl StableDiffusion3TripleClipWithTokenizer {
         })
     }
 
-    pub fn new(vb_fp16: candle_nn::VarBuilder, vb_fp32: candle_nn::VarBuilder) -> Result<Self> {
+    pub fn new(vb: candle_nn::VarBuilder) -> Result<Self> {
         let max_position_embeddings = 77usize;
         let clip_l = ClipWithTokenizer::new(
-            vb_fp16.pp("clip_l.transformer"),
+            vb.pp("clip_l.transformer"),
             stable_diffusion::clip::Config::sdxl(),
             "openai/clip-vit-large-patch14",
             max_position_embeddings,
         )?;
 
         let clip_g = ClipWithTokenizer::new(
-            vb_fp16.pp("clip_g.transformer"),
+            vb.pp("clip_g.transformer"),
             stable_diffusion::clip::Config::sdxl2(),
             "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k",
             max_position_embeddings,
         )?;
 
-        let text_projection = candle_nn::linear_no_bias(
-            1280,
-            1280,
-            vb_fp16.pp("clip_g.transformer.text_projection"),
-        )?;
+        let text_projection =
+            candle_nn::linear_no_bias(1280, 1280, vb.pp("clip_g.transformer.text_projection"))?;
 
-        // Current T5 implementation does not support fp16, so we use fp32 VarBuilder for T5.
-        // This is a temporary workaround until the T5 implementation is updated to support fp16.
-        // Also see:
-        // https://github.com/huggingface/candle/issues/2480
-        // https://github.com/huggingface/candle/pull/2481
-        let t5 = T5WithTokenizer::new(vb_fp32.pp("t5xxl.transformer"), max_position_embeddings)?;
+        let t5 = T5WithTokenizer::new(vb.pp("t5xxl.transformer"), max_position_embeddings)?;
         Ok(Self {
             clip_l,
             clip_g,

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

@@ -194,18 +194,11 @@ fn main() -> Result<()> {
             api.repo(hf_hub::Repo::model(name.to_string()))
         };
         let model_file = sai_repo.get("sd3_medium_incl_clips_t5xxlfp16.safetensors")?;
-        let vb_fp16 = unsafe {
+        let vb = unsafe {
             candle_nn::VarBuilder::from_mmaped_safetensors(&[&model_file], DType::F16, &device)?
         };
-
-        let vb_fp32 = unsafe {
-            candle_nn::VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)?
-        };
-        let triple = StableDiffusion3TripleClipWithTokenizer::new(
-            vb_fp16.pp("text_encoders"),
-            vb_fp32.pp("text_encoders"),
-        )?;
-        (MMDiTConfig::sd3_medium(), triple, vb_fp16)
+        let triple = StableDiffusion3TripleClipWithTokenizer::new(vb.pp("text_encoders"))?;
+        (MMDiTConfig::sd3_medium(), triple, vb)
     };
     let (context, y) = triple.encode_text_to_embedding(prompt.as_str(), &device)?;
     let (context_uncond, y_uncond) =

candle-transformers/src/models/t5.rs

@@ -1,12 +1,38 @@
 // T5 Text Model
 // https://github.com/huggingface/transformers/blob/main/src/transformers/models/t5/modeling_t5.py
 
-use crate::models::with_tracing::{linear_no_bias, Embedding, Linear};
+use crate::models::with_tracing::Embedding;
 use candle::{DType, Device, Module, Result, Tensor, D};
 use candle_nn::{Activation, VarBuilder};
 use serde::Deserialize;
 use std::sync::Arc;
 
+#[derive(Debug, Clone)]
+pub struct Linear {
+    weight: Tensor,
+    span: tracing::Span,
+}
+
+pub fn linear_no_bias(d1: usize, d2: usize, vb: VarBuilder) -> Result<Linear> {
+    let init_ws = candle_nn::init::DEFAULT_KAIMING_NORMAL;
+    let weight = vb.get_with_hints((d2, d1), "weight", init_ws)?;
+    let span = tracing::span!(tracing::Level::TRACE, "linear");
+    Ok(Linear { weight, span })
+}
+
+impl Module for Linear {
+    fn forward(&self, xs: &Tensor) -> Result<Tensor> {
+        let _enter = self.span.enter();
+        let weight = self.weight.to_dtype(xs.dtype())?;
+        let w = match *xs.dims() {
+            [b1, b2, _, _] => weight.broadcast_left((b1, b2))?.t()?,
+            [bsize, _, _] => weight.broadcast_left(bsize)?.t()?,
+            _ => weight.t()?,
+        };
+        xs.matmul(&w)
+    }
+}
+
 fn default_relative_attention_max_distance() -> usize {
     128
 }
@@ -185,7 +211,7 @@ impl Module for T5LayerNorm {
         let variance = xs_f32.sqr()?.mean_keepdim(D::Minus1)?;
         let xs = xs_f32.broadcast_div(&(variance + self.variance_epsilon)?.sqrt()?)?;
         let xs = xs.to_dtype(dtype)?;
-        let xs = xs.broadcast_mul(&self.weight)?;
+        let xs = xs.broadcast_mul(&self.weight.to_dtype(dtype)?)?;
         Ok(xs)
     }
 }
@@ -472,7 +498,8 @@ impl T5Attention {
                     let position_bias = relative_attention_bias
                         .forward(&relative_buckets)?
                         .permute((2, 0, 1))?
-                        .unsqueeze(0)?;
+                        .unsqueeze(0)?
+                        .to_dtype(scores.dtype())?;
                     (scores.broadcast_add(&position_bias)?, Some(position_bias))
                     // TODO: position_bias_masked?
                 }
@@ -678,9 +705,22 @@ impl T5Stack {
         &mut self,
         input_ids: &Tensor,
         encoder_hidden_states: Option<&Tensor>,
+    ) -> Result<Tensor> {
+        self.forward_dt(input_ids, encoder_hidden_states, None)
+    }
+
+    fn forward_dt(
+        &mut self,
+        input_ids: &Tensor,
+        encoder_hidden_states: Option<&Tensor>,
+        dtype: Option<DType>,
     ) -> Result<Tensor> {
         let _enter = self.span.enter();
         let input_embeds = self.shared.as_ref().forward(input_ids)?;
+        let input_embeds = match dtype {
+            None => input_embeds,
+            Some(dtype) => input_embeds.to_dtype(dtype)?,
+        };
         let mut hidden_states = input_embeds;
         let mut position_bias = None;
         for block in self.block.iter_mut() {
@@ -729,6 +769,11 @@ impl T5EncoderModel {
         self.encoder.forward(input_ids, None)
     }
 
+    pub fn forward_dt(&mut self, input_ids: &Tensor, dtype: Option<DType>) -> Result<Tensor> {
+        let _enter = self.span.enter();
+        self.encoder.forward_dt(input_ids, None, dtype)
+    }
+
     pub fn device(&self) -> &Device {
         &self.device
     }