rtx-nn / rtx-multimodal: cargo fmt reformatting

Pure formatting changes across rtx-nn (conv_transpose1d, conv/mod, rnn/lstm)
and rtx-multimodal (audio/generation, audio/source_separation): multi-line
braces, trailing commas, import ordering. No logic changes.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

crates/core/rtx-nn/src/layers/conv/conv_transpose1d.rs

@@ -74,7 +74,9 @@ impl ConvTranspose1dConfig {
             return Err(NNError::InvalidParameter("channels must be > 0".into()));
         }
         if self.kernel_size == 0 || self.stride == 0 || self.dilation == 0 {
-            return Err(NNError::InvalidParameter("kernel/stride/dilation must be > 0".into()));
+            return Err(NNError::InvalidParameter(
+                "kernel/stride/dilation must be > 0".into(),
+            ));
         }
         if self.groups == 0 {
             return Err(NNError::InvalidParameter("groups must be > 0".into()));
@@ -101,8 +103,7 @@ impl ConvTranspose1dConfig {
 
     /// Calculate output length for a given input length.
     pub fn output_length(&self, input_length: usize) -> usize {
-        (input_length - 1) * self.stride
+        (input_length - 1) * self.stride - 2 * self.padding
-            - 2 * self.padding
             + self.dilation * (self.kernel_size - 1)
             + self.output_padding
             + 1
@@ -147,7 +148,11 @@ impl ConvTranspose1d {
         let out_channels_per_group = config.out_channels / config.groups;
 
         // Weight shape: [in_channels, out_channels/groups, kernel_size]
-        let weight_shape = vec![config.in_channels, out_channels_per_group, config.kernel_size];
+        let weight_shape = vec![
+            config.in_channels,
+            out_channels_per_group,
+            config.kernel_size,
+        ];
         let mut weight = Tensor::zeros(weight_shape, device)?;
 
         let fan_in = config.in_channels * config.kernel_size / config.groups;
@@ -171,12 +176,24 @@ impl ConvTranspose1d {
         })
     }
 
-    pub fn in_channels(&self) -> usize { self.config.in_channels }
+    pub fn in_channels(&self) -> usize {
-    pub fn out_channels(&self) -> usize { self.config.out_channels }
+        self.config.in_channels
-    pub fn kernel_size(&self) -> usize { self.config.kernel_size }
+    }
-    pub fn stride(&self) -> usize { self.config.stride }
+    pub fn out_channels(&self) -> usize {
-    pub fn weight(&self) -> &Tensor { &self.weight }
+        self.config.out_channels
-    pub fn bias(&self) -> Option<&Tensor> { self.bias.as_ref() }
+    }
+    pub fn kernel_size(&self) -> usize {
+        self.config.kernel_size
+    }
+    pub fn stride(&self) -> usize {
+        self.config.stride
+    }
+    pub fn weight(&self) -> &Tensor {
+        &self.weight
+    }
+    pub fn bias(&self) -> Option<&Tensor> {
+        self.bias.as_ref()
+    }
 
     /// Perform the transposed convolution.
     ///
@@ -256,7 +273,8 @@ impl crate::layers::Module for ConvTranspose1d {
         if shape.dims()[1] != self.config.in_channels {
             return Err(NNError::InvalidParameter(format!(
                 "Expected {} input channels, got {}",
-                self.config.in_channels, shape.dims()[1]
+                self.config.in_channels,
+                shape.dims()[1]
             )));
         }
         self.conv_transpose1d_forward(input)
@@ -287,7 +305,10 @@ impl crate::layers::Module for ConvTranspose1d {
     }
 
     fn to_device(&mut self, device: &Device) -> Result<()> {
-        self.weight = self.weight.to_device(device).map_err(|e| NNError::Tensor(e))?;
+        self.weight = self
+            .weight
+            .to_device(device)
+            .map_err(|e| NNError::Tensor(e))?;
         if let Some(ref b) = self.bias {
             self.bias = Some(b.to_device(device).map_err(|e| NNError::Tensor(e))?);
         }
@@ -321,7 +342,9 @@ mod tests {
 
     #[test]
     fn config_output_length_with_padding() {
-        let cfg = ConvTranspose1dConfig::new(1, 1, 3).with_stride(2).with_padding(1);
+        let cfg = ConvTranspose1dConfig::new(1, 1, 3)
+            .with_stride(2)
+            .with_padding(1);
         // (10-1)*2 - 2 + 1*(3-1) + 0 + 1 = 19
         assert_eq!(cfg.output_length(10), 19);
     }
@@ -341,7 +364,9 @@ mod tests {
     #[test]
     fn forward_shape_stride2() {
         let device = Device::Cpu;
-        let cfg = ConvTranspose1dConfig::new(4, 8, 4).with_stride(2).with_padding(1);
+        let cfg = ConvTranspose1dConfig::new(4, 8, 4)
+            .with_stride(2)
+            .with_padding(1);
         let layer = ConvTranspose1d::from_config(cfg, &device).unwrap();
         let input = Tensor::randn(&[1, 4, 16], &device).unwrap();
         let output = layer.forward(&input).unwrap();
@@ -369,14 +394,13 @@ mod tests {
     #[test]
     fn no_bias_has_fewer_params() {
         let device = Device::Cpu;
-        let with_bias = ConvTranspose1d::from_config(
+        let with_bias =
-            ConvTranspose1dConfig::new(4, 8, 3),
+            ConvTranspose1d::from_config(ConvTranspose1dConfig::new(4, 8, 3), &device).unwrap();
-            &device,
-        ).unwrap();
         let no_bias = ConvTranspose1d::from_config(
             ConvTranspose1dConfig::new(4, 8, 3).with_bias(false),
             &device,
-        ).unwrap();
+        )
+        .unwrap();
         assert_eq!(with_bias.parameters().len(), 2);
         assert_eq!(no_bias.parameters().len(), 1);
     }
@@ -397,7 +421,11 @@ mod tests {
         let output_data = output.to_cpu().unwrap();
 
         for (i, &v) in input_data.iter().enumerate() {
-            assert!((output_data[i] - v).abs() < 1e-6, "mismatch at {i}: {v} vs {}", output_data[i]);
+            assert!(
+                (output_data[i] - v).abs() < 1e-6,
+                "mismatch at {i}: {v} vs {}",
+                output_data[i]
+            );
         }
     }
 }

crates/core/rtx-nn/src/layers/conv/mod.rs

@@ -13,9 +13,9 @@ pub mod conv_transpose1d;
 
 // Re-export main types
 pub use common::{PaddingMode, im2col, padding};
+pub use conv_transpose1d::{ConvTranspose1d, ConvTranspose1dConfig};
 pub use conv1d::{Conv1d, Conv1dConfig, Conv1dPadding};
 pub use conv2d::{Conv2d, Conv2dConfig};
-pub use conv_transpose1d::{ConvTranspose1d, ConvTranspose1dConfig};
 
 // Re-export for backward compatibility
 pub use common::PaddingMode as ConvPaddingMode;

crates/core/rtx-nn/src/layers/rnn/lstm.rs

@@ -162,12 +162,7 @@ impl LSTM {
     ///
     /// Input: `[batch, seq_len, input_size]` as flat f32 vec.
     /// Returns `LSTMOutput` with output and final states.
-    pub fn forward_cpu(
+    pub fn forward_cpu(&self, input: &[f32], batch: usize, seq_len: usize) -> Result<LSTMOutput> {
-        &self,
-        input: &[f32],
-        batch: usize,
-        seq_len: usize,
-    ) -> Result<LSTMOutput> {
         let h = self.config.hidden_size;
         let dirs = self.config.num_directions();
         let num_layers = self.config.num_layers;
@@ -264,15 +259,17 @@ impl LSTM {
     pub fn forward_tensor(&self, input: &Tensor) -> Result<(Tensor, Tensor, Tensor)> {
         let dims = input.shape().dims();
         if dims.len() != 3 {
-            return Err(NNError::InvalidParameter(
+            return Err(NNError::InvalidParameter(format!(
-                format!("LSTM expects 3D input [batch, seq, features], got {}D", dims.len())
+                "LSTM expects 3D input [batch, seq, features], got {}D",
-            ));
+                dims.len()
+            )));
         }
         let (batch, seq_len, feat) = (dims[0], dims[1], dims[2]);
         if feat != self.config.input_size {
-            return Err(NNError::InvalidParameter(
+            return Err(NNError::InvalidParameter(format!(
-                format!("Expected input_size {}, got {}", self.config.input_size, feat)
+                "Expected input_size {}, got {}",
-            ));
+                self.config.input_size, feat
+            )));
         }
 
         let input_data = input.to_cpu().map_err(|e| NNError::Tensor(e))?;
@@ -288,7 +285,9 @@ impl LSTM {
         Ok((output, h_n, c_n))
     }
 
-    pub fn config(&self) -> &LSTMConfig { &self.config }
+    pub fn config(&self) -> &LSTMConfig {
+        &self.config
+    }
 }
 
 #[inline]
@@ -362,7 +361,9 @@ mod tests {
 
     #[test]
     fn forward_multilayer_bidirectional_shape() {
-        let cfg = LSTMConfig::new(8, 16).with_num_layers(2).with_bidirectional(true);
+        let cfg = LSTMConfig::new(8, 16)
+            .with_num_layers(2)
+            .with_bidirectional(true);
         let lstm = LSTM::new(cfg, &Device::Cpu).unwrap();
 
         let batch = 3;
@@ -392,11 +393,8 @@ mod tests {
         let cfg = LSTMConfig::new(4, 8);
         let lstm = LSTM::new(cfg, &Device::Cpu).unwrap();
 
-        let input = Tensor::from_data(
+        let input =
-            vec![0.1f32; 2 * 5 * 4],
+            Tensor::from_data(vec![0.1f32; 2 * 5 * 4], vec![2, 5, 4], &Device::Cpu).unwrap();
-            vec![2, 5, 4],
-            &Device::Cpu,
-        ).unwrap();
 
         let (output, h_n, c_n) = lstm.forward_tensor(&input).unwrap();
         assert_eq!(output.shape().dims(), &[2, 5, 8]);
@@ -420,7 +418,8 @@ mod tests {
             vec![0.0f32; 2 * 5 * 3], // 3 != 4
             vec![2, 5, 3],
             &Device::Cpu,
-        ).unwrap();
+        )
+        .unwrap();
         assert!(lstm.forward_tensor(&input).is_err());
     }
 }

crates/models/rtx-multimodal/src/audio/generation/mod.rs

@@ -7,7 +7,7 @@
 pub mod stable_audio;
 
 #[cfg(feature = "generation")]
-pub use stable_audio::{StableAudioConfig, StableAudioModel, StableAudioError};
+pub use stable_audio::{StableAudioConfig, StableAudioError, StableAudioModel};
 
 use serde::{Deserialize, Serialize};
 

crates/models/rtx-multimodal/src/audio/generation/stable_audio.rs

@@ -67,7 +67,8 @@ impl StableAudioModel {
             "Loading Stable Audio Open model"
         );
 
-        let text_encoder = OnnxSession::from_file(&config.text_encoder_path, onnx_config.clone())
+        let text_encoder =
+            OnnxSession::from_file(&config.text_encoder_path, onnx_config.clone())
                 .map_err(|e| StableAudioError::ModelLoad(format!("text encoder: {e}")))?;
 
         let diffusion = OnnxSession::from_file(&config.diffusion_path, onnx_config.clone())
@@ -78,11 +79,19 @@ impl StableAudioModel {
 
         info!("Stable Audio Open loaded successfully");
 
-        Ok(Self { text_encoder, diffusion, vae_decoder, config })
+        Ok(Self {
+            text_encoder,
+            diffusion,
+            vae_decoder,
+            config,
+        })
     }
 
     /// Generate audio from a text prompt.
-    pub fn generate(&mut self, params: &GenerationParams) -> Result<GenerationOutput, StableAudioError> {
+    pub fn generate(
+        &mut self,
+        params: &GenerationParams,
+    ) -> Result<GenerationOutput, StableAudioError> {
         info!(prompt = %params.prompt, duration = params.duration_secs, steps = params.steps, "Generating audio");
 
         // Step 1: Encode text prompt
@@ -90,7 +99,8 @@ impl StableAudioModel {
         debug!("Text encoded");
 
         // Step 2: Initialize latent noise
-        let latent_frames = (params.duration_secs * self.config.sample_rate as f64 / 512.0) as usize;
+        let latent_frames =
+            (params.duration_secs * self.config.sample_rate as f64 / 512.0) as usize;
         let latent = self.initialize_latent(latent_frames, params.seed)?;
         debug!(latent_frames, "Latent initialized");
 
@@ -119,7 +129,11 @@ impl StableAudioModel {
         Ok(dummy)
     }
 
-    fn initialize_latent(&self, frames: usize, seed: Option<u64>) -> Result<Tensor, StableAudioError> {
+    fn initialize_latent(
+        &self,
+        frames: usize,
+        seed: Option<u64>,
+    ) -> Result<Tensor, StableAudioError> {
         // Initialize with random noise (or seeded noise for reproducibility)
         let latent = Tensor::randn(&[1, self.config.latent_dim, frames], &Device::Cpu)
             .map_err(|e| StableAudioError::Inference(e.to_string()))?;

crates/models/rtx-multimodal/src/audio/source_separation/demucs.rs

@@ -113,7 +113,11 @@ impl DemucsModel {
     /// For audio longer than `segment_length`, the input is split into overlapping
     /// segments, each processed independently, then recombined via overlap-add with
     /// a triangular cross-fade window.
-    pub fn separate(&mut self, waveform: &[f32], channels: usize) -> Result<Vec<StemOutput>, DemucsError> {
+    pub fn separate(
+        &mut self,
+        waveform: &[f32],
+        channels: usize,
+    ) -> Result<Vec<StemOutput>, DemucsError> {
         if waveform.is_empty() {
             return Err(DemucsError::EmptyInput);
         }
@@ -138,7 +142,8 @@ impl DemucsModel {
 
         // 4. Run inference on each segment
         let stem_types = StemType::stems_for(self.config.num_stems);
-        let mut stem_accumulators: Vec<Vec<f32>> = vec![vec![0.0; total_frames * channels]; self.config.num_stems];
+        let mut stem_accumulators: Vec<Vec<f32>> =
+            vec![vec![0.0; total_frames * channels]; self.config.num_stems];
         let mut weight_accumulator: Vec<f32> = vec![0.0; total_frames];
 
         for (start_frame, chunk) in &segments {
@@ -208,7 +213,12 @@ impl DemucsModel {
     }
 
     /// Segment audio into overlapping chunks.
-    fn segment(&self, channel_first: &[f32], channels: usize, total_frames: usize) -> Vec<(usize, Vec<f32>)> {
+    fn segment(
+        &self,
+        channel_first: &[f32],
+        channels: usize,
+        total_frames: usize,
+    ) -> Vec<(usize, Vec<f32>)> {
         let seg_len = self.config.segment_length;
         let hop = ((1.0 - self.config.overlap) * seg_len as f32) as usize;
         let hop = hop.max(1);
@@ -272,14 +282,19 @@ impl DemucsModel {
         let mut inputs = HashMap::new();
         inputs.insert("mix".to_string(), &input_tensor);
 
-        let outputs = self.session.run(inputs)
+        let outputs = self
+            .session
+            .run(inputs)
             .map_err(|e| DemucsError::Inference(e.to_string()))?;
 
         // Extract the output tensor (first output, whatever its name)
-        let output_tensor = outputs.into_values().next()
+        let output_tensor = outputs
+            .into_values()
+            .next()
             .ok_or_else(|| DemucsError::Inference("no output tensor from ONNX model".into()))?;
 
-        let output_data = output_tensor.to_vec_f32()
+        let output_data = output_tensor
+            .to_vec_f32()
             .map_err(|e| DemucsError::Inference(e.to_string()))?;
 
         Ok(output_data)
@@ -292,7 +307,11 @@ fn normalize(samples: &[f32]) -> (Vec<f32>, f32) {
         return (vec![], 1.0);
     }
 
-    let mean_sq: f64 = samples.iter().map(|&s| (s as f64) * (s as f64)).sum::<f64>() / samples.len() as f64;
+    let mean_sq: f64 = samples
+        .iter()
+        .map(|&s| (s as f64) * (s as f64))
+        .sum::<f64>()
+        / samples.len() as f64;
     let rms = mean_sq.sqrt() as f32;
     let scale = rms.max(1e-8);
 

crates/models/rtx-multimodal/src/audio/source_separation/htdemucs.rs

@@ -19,8 +19,8 @@
 
 use rtx_nn::layers::{
     Module,
-    conv::conv1d::{Conv1d, Conv1dConfig, Conv1dPadding},
     conv::conv_transpose1d::{ConvTranspose1d, ConvTranspose1dConfig},
+    conv::conv1d::{Conv1d, Conv1dConfig, Conv1dPadding},
     rnn::lstm::{LSTM, LSTMConfig},
 };
 use rtx_tensor::{Device, Tensor};
@@ -152,8 +152,7 @@ impl HtDemucsNative {
         let lstm_cfg = LSTMConfig::new(bottleneck_ch, bottleneck_ch)
             .with_num_layers(config.lstm_layers)
             .with_bidirectional(true);
-        let lstm = LSTM::new(lstm_cfg, device)
+        let lstm = LSTM::new(lstm_cfg, device).map_err(|e| HtDemucsError::Build(e.to_string()))?;
-            .map_err(|e| HtDemucsError::Build(e.to_string()))?;
 
         // Linear projection after BiLSTM (2*hidden → hidden)
         let proj_cfg = Conv1dConfig::new(bottleneck_ch * 2, bottleneck_ch, 1).with_bias(true);
@@ -195,10 +194,13 @@ impl HtDemucsNative {
 
         for (d, enc) in self.encoder_convs.iter().enumerate() {
             skips.push(x.clone());
-            x = enc.forward(&x)
+            x = enc
+                .forward(&x)
                 .map_err(|e| HtDemucsError::Forward(format!("encoder[{d}]: {e}")))?;
             // ReLU activation
-            let data = x.to_cpu().map_err(|e| HtDemucsError::Forward(e.to_string()))?;
+            let data = x
+                .to_cpu()
+                .map_err(|e| HtDemucsError::Forward(e.to_string()))?;
             let relu_data: Vec<f32> = data.iter().map(|&v| v.max(0.0)).collect();
             x = Tensor::from_data(relu_data, x.shape().dims().to_vec(), &self.device)
                 .map_err(|e| HtDemucsError::Forward(e.to_string()))?;
@@ -209,7 +211,9 @@ impl HtDemucsNative {
         let enc_dims = x.shape().dims();
         let (b, c, t) = (enc_dims[0], enc_dims[1], enc_dims[2]);
 
-        let x_data = x.to_cpu().map_err(|e| HtDemucsError::Forward(e.to_string()))?;
+        let x_data = x
+            .to_cpu()
+            .map_err(|e| HtDemucsError::Forward(e.to_string()))?;
         let mut transposed = vec![0.0f32; b * t * c];
         for bi in 0..b {
             for ci in 0..c {
@@ -222,19 +226,24 @@ impl HtDemucsNative {
         let lstm_input = Tensor::from_data(transposed, vec![b, t, c], &self.device)
             .map_err(|e| HtDemucsError::Forward(e.to_string()))?;
 
-        let (lstm_out, _, _) = self.lstm.forward_tensor(&lstm_input)
+        let (lstm_out, _, _) = self
+            .lstm
+            .forward_tensor(&lstm_input)
             .map_err(|e| HtDemucsError::Forward(format!("lstm: {e}")))?;
 
         // Project BiLSTM output (2*hidden → hidden) and transpose back to [batch, channels, time]
         let lo_dims = lstm_out.shape().dims();
-        let lo_data = lstm_out.to_cpu().map_err(|e| HtDemucsError::Forward(e.to_string()))?;
+        let lo_data = lstm_out
+            .to_cpu()
+            .map_err(|e| HtDemucsError::Forward(e.to_string()))?;
         let lstm_ch = lo_dims[2]; // 2 * hidden_size
 
         let mut back = vec![0.0f32; b * lstm_ch * t];
         for bi in 0..b {
             for ti in 0..t {
                 for ci in 0..lstm_ch {
-                    back[bi * lstm_ch * t + ci * t + ti] = lo_data[bi * t * lstm_ch + ti * lstm_ch + ci];
+                    back[bi * lstm_ch * t + ci * t + ti] =
+                        lo_data[bi * t * lstm_ch + ti * lstm_ch + ci];
                 }
             }
         }
@@ -243,7 +252,9 @@ impl HtDemucsNative {
             .map_err(|e| HtDemucsError::Forward(e.to_string()))?;
 
         // 1x1 projection: [batch, 2*hidden, time] → [batch, hidden, time]
-        x = self.output_conv.forward(&x)
+        x = self
+            .output_conv
+            .forward(&x)
             .map_err(|e| HtDemucsError::Forward(format!("projection: {e}")))?;
 
         // Decoder pass with skip connections
@@ -252,8 +263,12 @@ impl HtDemucsNative {
             let skip = &skips[skip_idx];
 
             // Concatenate skip connection along channel dimension
-            let x_data = x.to_cpu().map_err(|e| HtDemucsError::Forward(e.to_string()))?;
+            let x_data = x
-            let s_data = skip.to_cpu().map_err(|e| HtDemucsError::Forward(e.to_string()))?;
+                .to_cpu()
+                .map_err(|e| HtDemucsError::Forward(e.to_string()))?;
+            let s_data = skip
+                .to_cpu()
+                .map_err(|e| HtDemucsError::Forward(e.to_string()))?;
 
             let x_dims = x.shape().dims();
             let s_dims = skip.shape().dims();
@@ -285,13 +300,16 @@ impl HtDemucsNative {
             x = Tensor::from_data(cat_data, vec![xb, cat_c, min_t], &self.device)
                 .map_err(|e| HtDemucsError::Forward(e.to_string()))?;
 
-            x = dec.forward(&x)
+            x = dec
+                .forward(&x)
                 .map_err(|e| HtDemucsError::Forward(format!("decoder[{d}]: {e}")))?;
         }
 
         // Reshape output: [batch, num_sources * audio_channels, samples]
         // → split into [batch, audio_channels, samples] per source
-        let out_data = x.to_cpu().map_err(|e| HtDemucsError::Forward(e.to_string()))?;
+        let out_data = x
+            .to_cpu()
+            .map_err(|e| HtDemucsError::Forward(e.to_string()))?;
         let out_dims = x.shape().dims();
         let out_len = out_dims[2].min(samples); // Trim to original length
         let num_sources = self.config.num_sources;
@@ -382,11 +400,8 @@ mod tests {
         let model = HtDemucsNative::new(cfg, &Device::Cpu).unwrap();
 
         // Input: batch=1, stereo, 4096 samples
-        let input = Tensor::from_data(
+        let input =
-            vec![0.01f32; 1 * 2 * 4096],
+            Tensor::from_data(vec![0.01f32; 1 * 2 * 4096], vec![1, 2, 4096], &Device::Cpu).unwrap();
-            vec![1, 2, 4096],
-            &Device::Cpu,
-        ).unwrap();
 
         let stems = model.forward(&input).unwrap();
         assert_eq!(stems.len(), 4);

crates/models/rtx-multimodal/src/audio/source_separation/mod.rs

@@ -31,11 +31,18 @@ impl StemType {
     pub fn stems_for(num_stems: usize) -> Vec<StemType> {
         match num_stems {
             6 => vec![
-                StemType::Drums, StemType::Bass, StemType::Other,
+                StemType::Drums,
-                StemType::Vocals, StemType::Guitar, StemType::Piano,
+                StemType::Bass,
+                StemType::Other,
+                StemType::Vocals,
+                StemType::Guitar,
+                StemType::Piano,
             ],
             _ => vec![
-                StemType::Drums, StemType::Bass, StemType::Other, StemType::Vocals,
+                StemType::Drums,
+                StemType::Bass,
+                StemType::Other,
+                StemType::Vocals,
             ],
         }
     }