fix(cuda+csm): P0 rtx-backend-cuda compile fix + P2 rtx-csm clippy cleanup (#9)

Co-authored-by: Omar Sobh <[email protected]>
Co-committed-by: Omar Sobh <[email protected]>

crates/core/rtx-backend-cuda/Cargo.toml

@@ -49,3 +49,7 @@ path = "src/lib.rs"
 
 [lints]
 workspace = true
+
+[[test]]
+name = "backend_parity_tests"
+required-features = ["cuda"]

crates/core/rtx-backend-cuda/src/device.rs

@@ -6,6 +6,18 @@
 use cudarc::cublas::CudaBlas;
 #[cfg(feature = "cuda")]
 use cudarc::driver::{CudaContext, CudaStream};
+#[cfg(feature = "cuda")]
+use crate::CudaBackend;
+#[cfg(feature = "cuda")]
+use rtx_backend::{DeviceId, DeviceOps};
+#[cfg(feature = "cuda")]
+use once_cell::sync::{Lazy, OnceCell};
+#[cfg(feature = "cuda")]
+use parking_lot::RwLock;
+#[cfg(feature = "cuda")]
+use std::collections::HashMap;
+#[cfg(feature = "cuda")]
+use std::sync::Arc;
 
 use crate::{CudaError, CudaResult};
 

crates/core/rtx-backend-cuda/src/lib.rs

@@ -104,6 +104,8 @@ pub use device::CudaDevice;
 pub use error::{CudaError, CudaResult};
 pub use tensor::CudaTensorPrimitive;
 
+#[cfg(feature = "cuda")]
+use rtx_backend::{Backend, BoolU8, DeviceId, DeviceOps};
 use std::fmt::Debug;
 
 /// CUDA backend for RustyTorch++.

crates/core/rtx-backend-cuda/src/tensor.rs

@@ -5,6 +5,8 @@
 use crate::CudaDevice;
 #[cfg(feature = "cuda")]
 use cudarc::driver::CudaSlice;
+#[cfg(feature = "cuda")]
+use std::sync::Arc;
 
 #[cfg(feature = "cuda")]
 /// CUDA tensor primitive - GPU memory storage for tensors.

crates/models/rtx-csm/src/audioseal.rs

@@ -448,7 +448,7 @@ impl MsgProcessor {
         debug_assert_eq!(c, self.hidden);
         let mut indices = Vec::with_capacity(self.nbits);
         for k in 0..self.nbits {
-            let bit = ((message >> k) & 1) as u32;
+            let bit = (message >> k) & 1;
             indices.push(2 * k as u32 + bit);
         }
         let idx = Tensor::from_vec(indices, (self.nbits,), device)?;

crates/models/rtx-csm/src/csm_fork.rs

@@ -923,6 +923,7 @@ impl Model {
 
         self.decoder.clear_kv_cache();
         let mut decoder_pos = 0;
+        #[allow(clippy::needless_range_loop)]
         for i in 1..self.config.audio_num_codebooks {
             let proj_h = curr_h.apply(&self.projection)?;
             let decoder_h = self.decoder.forward(&proj_h, decoder_pos)?;
@@ -1019,6 +1020,7 @@ impl Model {
         let mut curr_h = Tensor::cat(&[h, c0_embed], 1)?;
         self.decoder.clear_kv_cache();
         let mut decoder_pos = 0usize;
+        #[allow(clippy::needless_range_loop)]
         for i in 1..self.config.audio_num_codebooks {
             let proj_h = curr_h.apply(&self.projection)?;
             let decoder_h = self.decoder.forward(&proj_h, decoder_pos)?;
@@ -1054,6 +1056,7 @@ impl Model {
     /// respectively).
     ///
     /// Reference: Koel-TTS (NVIDIA, arXiv 2502.05236), `cfg_scale ∈ [1.5, 3.0]`.
+    #[allow(clippy::too_many_arguments)]
     pub fn generate_frame_cfg(
         &mut self,
         cond_tokens: &Tensor,

crates/models/rtx-csm/src/csm_quantized.rs

@@ -406,9 +406,9 @@ impl Layer {
         let xs = xs.apply(&self.mlp_norm)?.apply(&self.mlp)?;
         let out = (residual + xs)?;
         // Optional layer-by-layer dump for qmatmul bisection.
-        if std::env::var("CSM_DUMP_LAYERS").is_ok() {
+        if std::env::var("CSM_DUMP_LAYERS").is_ok()
-            if let Some(idx) = dbg_idx {
+            && let Some(idx) = dbg_idx
-                if seqlen_offset == 0 {
+                && seqlen_offset == 0 {
                     let f = out.flatten_all()?.to_vec1::<f32>().unwrap_or_default();
                     let n: f32 = f.iter().map(|x| x * x).sum::<f32>().sqrt();
                     eprintln!(
@@ -419,8 +419,6 @@ impl Layer {
                         &f[..8.min(f.len())]
                     );
                 }
-            }
-        }
         Ok(out)
     }
 
@@ -761,6 +759,7 @@ impl Model {
         self.run_decoder(h, c0_sample, lp)
     }
 
+    #[allow(clippy::too_many_arguments)]
     pub fn generate_frame_cfg(
         &mut self,
         cond_tokens: &Tensor,

crates/models/rtx-csm/src/diarize.rs

@@ -231,12 +231,9 @@ fn vad_intervals(probs: &[f32], cfg: &DiarizationConfig) -> Vec<(usize, usize)>
     let mut last_speech: Option<usize> = None;
     for i in 0..active.len() {
         if active[i] {
-            if let Some(prev) = last_speech {
+            if let Some(prev) = last_speech
-                if i - prev <= max_silence_chunks + 1 {
+                && i - prev <= max_silence_chunks + 1 {
-                    for j in (prev + 1)..i {
+                    active[(prev + 1)..i].fill(true);
-                        active[j] = true;
-                    }
-                }
                 }
             last_speech = Some(i);
         }
@@ -308,11 +305,10 @@ fn agglomerative_cluster(
         if clusters.len() == 1 {
             break;
         }
-        if let Some(k) = n_speakers {
+        if let Some(k) = n_speakers
-            if clusters.len() <= k {
+            && clusters.len() <= k {
                 break;
             }
-        }
         // Find the closest pair using average linkage.
         let mut best: Option<(usize, usize, f32)> = None;
         for i in 0..clusters.len() {

crates/models/rtx-csm/src/emotion2vec.rs

@@ -504,7 +504,7 @@ impl RelativePositionalEncoder {
         groups: usize,
         vb: VarBuilder,
     ) -> CsmResult<Self> {
-        if kernel % 2 == 0 {
+        if kernel.is_multiple_of(2) {
             return Err(crate::CsmError::Config(format!(
                 "relative_positional_encoder kernel must be odd for same-padding, got {kernel}"
             )));

crates/models/rtx-csm/src/generator.rs

@@ -392,14 +392,13 @@ impl Generator {
                 break;
             }
 
-            if let Some(g) = rep_guard.as_mut() {
+            if let Some(g) = rep_guard.as_mut()
-                if g.observe(&sampled) {
+                && g.observe(&sampled) {
                     tracing::warn!(
                         "loop-escape: repetition guard tripped at frame {frame_idx}; ending generation"
                     );
                     break;
                 }
-            }
 
             all_frames.push(sampled.clone());
 
@@ -502,14 +501,13 @@ impl Generator {
                 tracing::info!("EOT detected at frame {frame_idx}");
                 break;
             }
-            if let Some(g) = rep_guard.as_mut() {
+            if let Some(g) = rep_guard.as_mut()
-                if g.observe(&sampled) {
+                && g.observe(&sampled) {
                     tracing::warn!(
                         "loop-escape: repetition guard tripped at frame {frame_idx}; ending"
                     );
                     break;
                 }
-            }
 
             pending.push(sampled.clone());
 
@@ -541,12 +539,11 @@ impl Generator {
             )?;
         }
         // Drain any internal Mimi buffering by feeding a `None` step.
-        if let Some(tail) = self.mimi.decode_step(None)? {
+        if let Some(tail) = self.mimi.decode_step(None)?
-            if !tail.is_empty() {
+            && !tail.is_empty() {
                 on_chunk(&tail)?;
                 full_pcm.extend_from_slice(&tail);
             }
-        }
 
         tracing::info!(
             "streaming generation finished: {} samples (~{:.2}s)",
@@ -630,12 +627,11 @@ where
     }
     let codes = Tensor::from_vec(flat, (1, num_codebooks, k), device)?.to_dtype(DType::U32)?;
     pending.clear();
-    if let Some(samples) = mimi.decode_step(Some(&codes))? {
+    if let Some(samples) = mimi.decode_step(Some(&codes))?
-        if !samples.is_empty() {
+        && !samples.is_empty() {
             on_chunk(&samples)?;
             full_pcm.extend_from_slice(&samples);
         }
-    }
     Ok(())
 }
 

crates/models/rtx-csm/src/llm_client.rs

@@ -239,13 +239,12 @@ impl LlmClient for OpenAiCompatibleClient {
         // default case so this is a no-op.
         let mut body = serde_json::to_value(&req)
             .map_err(|e| CsmError::Config(format!("LLM serialize: {e}")))?;
-        if !config.extra_body.is_empty() {
+        if !config.extra_body.is_empty()
-            if let Some(obj) = body.as_object_mut() {
+            && let Some(obj) = body.as_object_mut() {
                 for (k, v) in config.extra_body.iter() {
                     obj.insert(k.clone(), v.clone());
                 }
             }
-        }
         let response = self
             .http
             .post(&url)

crates/models/rtx-csm/src/longform.rs

@@ -106,9 +106,9 @@ fn split_on_sentence_boundaries(text: &str) -> Vec<String> {
 
 impl Generator {
     /// Generate audio for arbitrarily long text by chunking on sentence
-    /// boundaries with a rolling context. The previous generated chunk's audio
+    /// boundaries with a rolling context. The previous generated chunk's
-    /// + transcript becomes context for the next call. Returns the full
+    /// audio + transcript becomes context for the next call. Returns the
-    /// concatenated PCM.
+    /// full concatenated PCM.
     pub fn generate_long(
         &mut self,
         text: &str,
@@ -185,6 +185,7 @@ impl Generator {
     ///
     /// Order matches `generate_to_wav` exactly so installing a watermarker
     /// applies uniformly to short-form and long-form output.
+    #[allow(clippy::too_many_arguments)]
     pub fn generate_long_to_wav(
         &mut self,
         text: &str,

crates/models/rtx-csm/src/lora.rs

@@ -78,6 +78,7 @@ pub struct LoraDelta {
 }
 
 impl LoraDelta {
+    #[allow(clippy::too_many_arguments)]
     pub fn new(
         rank: usize,
         alpha: f64,
@@ -169,6 +170,7 @@ impl LoraLinear {
     /// Wrap an existing frozen `Linear` with a trainable rank-r LoRA adapter.
     /// The A/B params get registered into `vm` under `<prefix>.lora_a` /
     /// `<prefix>.lora_b` so AdamW (or any optimizer) can find and update them.
+    #[allow(clippy::too_many_arguments)]
     pub fn wrap(
         base: Linear,
         rank: usize,

crates/models/rtx-csm/src/model.rs

@@ -68,6 +68,7 @@ impl ModelBackend {
             Self::Quantized(m) => m.generate_frame(tokens, mask, input_pos, lp),
         }
     }
+    #[allow(clippy::too_many_arguments)]
     pub fn generate_frame_cfg(
         &mut self,
         cond_tokens: &Tensor,

crates/models/rtx-csm/src/moonshine.rs

@@ -77,7 +77,7 @@ impl MoonshineConfig {
     pub fn padded_head_dim(&self, num_heads: usize) -> usize {
         let raw = self.hidden_size / num_heads;
         let pad = self.pad_head_dim_to_multiple_of;
-        ((raw + pad - 1) / pad) * pad
+        raw.div_ceil(pad) * pad
     }
 }
 
@@ -191,7 +191,7 @@ struct RotaryCache {
 
 impl RotaryCache {
     fn new(rotary_dim: usize, max_seq: usize, theta: f64, dtype: DType, dev: &Device) -> Result<Self> {
-        assert!(rotary_dim % 2 == 0, "rotary_dim must be even");
+        assert!(rotary_dim.is_multiple_of(2), "rotary_dim must be even");
         let inv_freq: Vec<f32> = (0..rotary_dim)
             .step_by(2)
             .map(|i| (1.0 / theta.powf(i as f64 / rotary_dim as f64)) as f32)

crates/models/rtx-csm/src/post.rs

@@ -42,7 +42,7 @@ impl PostProcess {
         }
     }
 
-    pub fn apply(&self, samples: &mut Vec<f32>, sample_rate: u32) -> Result<()> {
+    pub fn apply(&self, samples: &mut [f32], sample_rate: u32) -> Result<()> {
         if samples.is_empty() {
             return Ok(());
         }

crates/models/rtx-csm/src/prompt.rs

@@ -2,8 +2,10 @@
 //! that CSM's dual-transformer expects.
 //!
 //! Slot layout per row of the (S, cb+1) tensor:
+//!
 //!   - slots [0..cb):  per-codebook audio token id (or 0 if unused)
 //!   - slot   cb:      Llama text token id (or 0 if unused)
+//!
 //! The mask (S, cb+1) u8 says which slots actually carry data.
 //!
 //! We delegate per-step tensor construction to candle's

crates/models/rtx-csm/src/ser.rs

@@ -178,11 +178,10 @@ impl ProsodyDetector {
                 min_lag,
                 max_lag,
                 self.voiced_threshold,
-            ) {
+            )
-                if voiced && f0 >= self.min_f0_hz && f0 <= self.max_f0_hz {
+                && voiced && f0 >= self.min_f0_hz && f0 <= self.max_f0_hz {
                     voiced_f0s.push(f0);
                 }
-            }
             total_frames += 1;
             t += self.hop_samples;
         }

crates/models/rtx-csm/src/silentcipher.rs

@@ -516,7 +516,7 @@ pub fn vb_from_ckpt(
 
 /// VCTK average power baseline used by the original to pre-condition
 /// audio energy before watermarking. Source: `server.py:60` constant.
-const AVERAGE_ENERGY_VCTK: f32 = 0.002837200844477648;
+const AVERAGE_ENERGY_VCTK: f32 = 0.002_837_200_9;
 
 /// SilentCipher watermarker holding all three networks + an STFT
 /// helper. Loads from the three released `.ckpt` files (`enc_c`,

crates/models/rtx-csm/src/silero_vad.rs

@@ -290,7 +290,7 @@ impl SileroVad {
         padded.extend_from_slice(samples);
         // Tail-pad to whole-chunk multiple.
         let extra = (CHUNK_SAMPLES - (samples.len() % CHUNK_SAMPLES)) % CHUNK_SAMPLES;
-        padded.extend(std::iter::repeat(0.0).take(extra));
+        padded.extend(std::iter::repeat_n(0.0, extra));
 
         let mut probs = Vec::new();
         let mut i = 0;

crates/models/rtx-csm/src/training.rs

@@ -385,7 +385,7 @@ impl<'a> Trainer<'a> {
                     self.generator.model.inner.refresh_lora(self.vm)?;
 
                     step += 1;
-                    if step % 10 == 0 || step == total_steps - 1 {
+                    if step.is_multiple_of(10) || step == total_steps - 1 {
                         tracing::info!(
                             "  step {step:>4}/{total_steps} epoch {epoch} ex {idx} frame {frame_idx} \
                              prompt_len={prompt_len} lr={lr:.2e} loss={loss_val:.4}"
@@ -407,9 +407,11 @@ impl<'a> Trainer<'a> {
 ///   - any other string matches `ex.stage == Some(label)` exactly
 ///
 /// Per `personal_voice_training_guide.md` §4 the canonical 3-stage recipe is:
+///
 ///   - audiobook: 3 epochs, peak_lr 1e-4
 ///   - podcast:   1 epoch,  peak_lr 3e-5
 ///   - va:        1 epoch,  peak_lr 1e-5
+///
 /// each over a growing pool. The `CurriculumTrainer` enforces the pool growth
 /// implicitly by virtue of how you label your manifest (a podcast example
 /// labeled `stage: "podcast"` is only matched by the podcast stage; if you

crates/models/rtx-csm/src/wav2vec2.rs

@@ -291,7 +291,7 @@ impl ConvPosEmbedding {
 /// Even kernels need asymmetric handling: pad symmetrically by k/2 and
 /// trim the trailing extra frame.
 fn pad_for(kernel: usize) -> (usize, bool) {
-    if kernel % 2 == 0 {
+    if kernel.is_multiple_of(2) {
         (kernel / 2, true)
     } else {
         ((kernel - 1) / 2, false)
@@ -595,15 +595,14 @@ pub fn ctc_greedy_decode(logits: &Tensor, vocab: &[&str]) -> CsmResult<String> {
         if id_i == prev {
             continue;
         }
-        if (id as usize) != CTC_BLANK_ID {
+        if (id as usize) != CTC_BLANK_ID
-            if let Some(tok) = vocab.get(id as usize) {
+            && let Some(tok) = vocab.get(id as usize) {
                 if *tok == "|" {
                     out.push(' ');
                 } else if !tok.starts_with('<') {
                     out.push_str(tok);
                 }
             }
-        }
         prev = id_i;
     }
     Ok(out)

crates/models/rtx-csm/src/wavlm_sv.rs

@@ -232,7 +232,7 @@ impl Module for PosConv {
         let h = xs_bct.apply(&self.conv)?;
         // SamePad strips 1 trailing frame because kernel=128 is even.
         let out_len = h.dim(D::Minus1)?;
-        let h = if self.kernel % 2 == 0 && out_len > in_len {
+        let h = if self.kernel.is_multiple_of(2) && out_len > in_len {
             h.narrow(D::Minus1, 0, in_len)?
         } else {
             h

crates/models/rtx-csm/src/wer.rs

@@ -79,6 +79,7 @@ fn lev_align(r: &[&str], h: &[&str]) -> (usize, usize, usize) {
     }
     let mut prev = vec![Cell::new(); m + 1];
     let mut curr = vec![Cell::new(); m + 1];
+    #[allow(clippy::needless_range_loop)]
     for j in 0..=m {
         prev[j] = Cell { cost: j, s: 0, d: 0, i: j };
     }