rtx-csm: WavLM-SV bug fixes — gelu_erf + gru_rel_pos_const loading

Two real bugs found via code inspection against HF source:

1. candle's .gelu() is the tanh approximation; PyTorch's default 'gelu'
   activation (used in WavLM via ACT2FN['gelu']) is the exact erf-based
   version. Switched all 3 sites (feature extractor convs, pos_conv,
   FFN) from .gelu() to .gelu_erf() to match the reference.

2. gru_rel_pos_const lookup used vb.pp("name").get(shape, "") which
   resolves to "<prefix>.name." (trailing dot) and fails to find the
   tensor. The .or_else(|_| zeros) silently swallowed the failure,
   leaving all 12 layers' gating constants at zero instead of the
   trained values. Fixed to attn.get(shape, "gru_rel_pos_const") which
   resolves correctly.

examples/wavlm_sv_inspect.rs: utility for sanity-checking specific
tensors inside converted safetensors (e.g. layer_weights).

Same-content same-speaker cosine: 0.9985 -> 0.9963 (≈unchanged).
Cross-content same-speaker cosine: 0.4882 -> 0.4118 (still drifting).
Phase 5d (Python reference comparison) remains the gate for
identifying the residual numerical drift.

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

crates/models/rtx-csm/Cargo.toml

@@ -149,6 +149,10 @@ path = "examples/wavlm_sv_convert.rs"
 name = "wavlm_sv_demo"
 path = "examples/wavlm_sv_demo.rs"
 
+[[example]]
+name = "wavlm_sv_inspect"
+path = "examples/wavlm_sv_inspect.rs"
+
 [[example]]
 name = "pipeline"
 path = "examples/pipeline.rs"

crates/models/rtx-csm/examples/wavlm_sv_inspect.rs

@@ -0,0 +1,52 @@
+//! Inspect specific tensors inside a converted WavLM-SV safetensors.
+//! Useful for sanity-checking weight loading and key naming.
+
+use anyhow::Result;
+use candle_core::Device;
+use clap::Parser;
+use std::path::PathBuf;
+
+#[derive(Debug, Parser)]
+struct Cli {
+    #[arg(long)]
+    weights: PathBuf,
+    /// Tensor name to dump (e.g. "layer_weights", "projector.weight").
+    #[arg(long)]
+    key: String,
+    /// Apply softmax along dim=0 before printing (for layer_weights).
+    #[arg(long)]
+    softmax: bool,
+    /// Number of leading elements to print (default: all).
+    #[arg(long)]
+    head: Option<usize>,
+}
+
+fn main() -> Result<()> {
+    let cli = Cli::parse();
+    let tensors = candle_core::safetensors::load(&cli.weights, &Device::Cpu)?;
+    let t = tensors
+        .get(&cli.key)
+        .ok_or_else(|| anyhow::anyhow!("key not found: {}", cli.key))?;
+    println!("{}: dtype={:?}, shape={:?}", cli.key, t.dtype(), t.dims());
+    let to_print = if cli.softmax {
+        candle_nn::ops::softmax(t, 0)?
+    } else {
+        t.clone()
+    };
+    let v: Vec<f32> = to_print
+        .flatten_all()?
+        .to_dtype(candle_core::DType::F32)?
+        .to_vec1()?;
+    let limit = cli.head.unwrap_or(v.len()).min(v.len());
+    let head: Vec<&f32> = v.iter().take(limit).collect();
+    println!("first {limit} values: {head:?}");
+    if v.len() > limit {
+        let tail: Vec<&f32> = v.iter().rev().take(8).collect();
+        println!("(last 8 values, reversed): {tail:?}");
+    }
+    let sum: f32 = v.iter().sum();
+    let max = v.iter().cloned().fold(f32::NEG_INFINITY, f32::max);
+    let min = v.iter().cloned().fold(f32::INFINITY, f32::min);
+    println!("sum={sum:.6}, min={min:.6}, max={max:.6}");
+    Ok(())
+}

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

@@ -100,7 +100,7 @@ impl ConvLayer {
             Some(gn) => h.apply(gn)?,
             None => h,
         };
-        h.gelu()
+        h.gelu_erf()
     }
 }
 
@@ -237,7 +237,7 @@ impl Module for PosConv {
         } else {
             h
         };
-        let h = h.gelu()?;
+        let h = h.gelu_erf()?;
         // Back to (B, T, C).
         h.transpose(1, 2)?.contiguous()
     }
@@ -298,12 +298,7 @@ impl WavLmEncoderLayer {
         let fc1 = linear(HIDDEN_DIM, FFN_DIM, ff.pp("intermediate_dense"))?;
         let fc2 = linear(FFN_DIM, HIDDEN_DIM, ff.pp("output_dense"))?;
         let final_norm = layer_norm(HIDDEN_DIM, 1e-5, vb.pp("final_layer_norm"))?;
-        let gru_rel_pos_const = attn
+        let gru_rel_pos_const = attn.get((1, NUM_HEADS, 1, 1), "gru_rel_pos_const")?;
-            .pp("gru_rel_pos_const")
-            .get((1, NUM_HEADS, 1, 1), "")
-            .or_else(|_| {
-                Tensor::zeros((1, NUM_HEADS, 1, 1), vb.dtype(), vb.device())
-            })?;
         let gru_rel_pos_linear = linear(HEAD_DIM, 8, attn.pp("gru_rel_pos_linear"))?;
         let rel_attn_embed = if layer_idx == 0 {
             Some(candle_nn::embedding(
@@ -440,7 +435,7 @@ impl WavLmEncoderLayer {
         let h = (xs + attn_out)?;
         let h = h.apply(&self.attn_norm)?;
         // FFN: 768 → 3072 → GELU → 768
-        let ffn = h.apply(&self.fc1)?.gelu()?.apply(&self.fc2)?;
+        let ffn = h.apply(&self.fc1)?.gelu_erf()?.apply(&self.fc2)?;
         let h = (h + ffn)?;
         let h = h.apply(&self.final_norm)?;
         Ok((h, bias.clone()))