cohere: support untied head weight

scripts/convert-cohere.py

@@ -554,9 +554,9 @@ def add(src_name: str, gguf_name: str, transform) -> None:
                 )
             # Read as torch bf16 tensor, convert to fp32 numpy.
             t = st.get_tensor(src_name)
-            if t.dtype != torch.bfloat16:
+            if t.dtype not in (torch.bfloat16, torch.float32):
                 raise ValueError(
-                    f"{src_name}: expected source dtype torch.bfloat16, got {t.dtype}"
+                    f"{src_name}: expected source dtype torch.bfloat16 or torch.float32, got {t.dtype}"
                 )
             arr = t.float().numpy()
             arr = transform(arr)
@@ -609,11 +609,31 @@ def add(src_name: str, gguf_name: str, transform) -> None:
         for src_name, gguf_name, transform in HEAD_TABLE:
             add(src_name, gguf_name, transform)
 
+        # Some fine-tunes (e.g. syvai/hviske-v5*) train the head separately
+        # from the embedding. Store it as head.weight so the runtime does
+        # not fall back to the tied embedding.
+        head_untied = not torch.equal(
+            st.get_tensor(tied_key),
+            st.get_tensor("transf_decoder._embedding.token_embedding.weight"),
+        )
+        if head_untied:
+            print("Head weight is untied from the embedding; writing head.weight")
+            add(tied_key, "head.weight", passthrough)
+
         # Mel frontend buffers (filterbank + window) — always stored as f32.
         # These are the exact values the model was trained with; using them
         # instead of recomputing from scratch eliminates mel-level divergence.
         fb_src = "preprocessor.featurizer.fb"
-        fb_tensor = st.get_tensor(fb_src).float().numpy()
+        if fb_src in st_keys:
+            fb_tensor = st.get_tensor(fb_src).float().numpy()
+        else:
+            # Checkpoints saved without the preprocessor buffer: rebuild it the
+            # way the remote-code CohereAsrFeatureExtractor does.
+            import librosa
+            fb_tensor = librosa.filters.mel(
+                sr=hp["fe_sample_rate"], n_fft=hp["fe_n_fft"], n_mels=hp["fe_num_mels"],
+                fmin=0.0, fmax=hp["fe_sample_rate"] / 2, norm="slaney",
+            ).astype(np.float32)
         if fb_tensor.ndim == 3:
             fb_tensor = fb_tensor.squeeze(0)  # [1, 128, 257] -> [128, 257]
         writer.add_tensor("frontend.mel_filterbank", fb_tensor)
@@ -645,6 +665,7 @@ def add(src_name: str, gguf_name: str, transform) -> None:
             + hp["dec_n_layers"] * len(DECODER_BLOCK_TABLE)    # 8 * 26 = 208
             + len(DEC_FINAL_NORM_TABLE)                        # 2
             + len(HEAD_TABLE)                                  # 1
+            + int(head_untied)                                 # optional head.weight
             + 2                                                # frontend fb + window
         )
         if n_added != expected:

src/arch/cohere/decoder.cpp

@@ -536,7 +536,7 @@ DecoderBuild build_decoder_graph(ggml_context *        ctx,
 
     // Head: tied weight (the [hidden, vocab_size] token embedding) + bias +
     // log-softmax. mul_mat(token_w, x[hidden,seq]) -> [vocab_size, seq_len].
-    ggml_tensor * logits = ggml_mul_mat(ctx, w.dec_embed.token_w, x);
+    ggml_tensor * logits = ggml_mul_mat(ctx, w.head.logits_w(w.dec_embed.token_w), x);
     if (w.head.bias != nullptr) {
         logits = ggml_add(ctx, logits, w.head.bias);
     }
@@ -745,7 +745,7 @@ DecoderBuild build_decoder_graph_kv(ggml_context *        ctx,
     db.dumps.out_before_head = x;
 
     // Head: tied weight + bias + optional log-softmax.
-    ggml_tensor * logits = ggml_mul_mat(ctx, w.dec_embed.token_w, x);
+    ggml_tensor * logits = ggml_mul_mat(ctx, w.head.logits_w(w.dec_embed.token_w), x);
     if (w.head.bias != nullptr) {
         logits = ggml_add(ctx, logits, w.head.bias);
     }
@@ -883,7 +883,7 @@ StepBuild build_step_graph(ggml_context *        ctx,
     x = layer_norm(ctx, x, w.dec_final.norm_w, w.dec_final.norm_b);
 
     // Tied LM head: token embedding transposed.
-    ggml_tensor * logits = ggml_mul_mat(ctx, w.dec_embed.token_w, x);
+    ggml_tensor * logits = ggml_mul_mat(ctx, w.head.logits_w(w.dec_embed.token_w), x);
     if (w.head.bias != nullptr) {
         logits = ggml_add(ctx, logits, w.head.bias);
     }
@@ -1156,7 +1156,7 @@ StepBuildBatched build_step_graph_batched(ggml_context *        ctx,
     }
 
     x                    = layer_norm(ctx, x, w.dec_final.norm_w, w.dec_final.norm_b);
-    ggml_tensor * logits = ggml_mul_mat(ctx, w.dec_embed.token_w, x);  // [vocab, B]
+    ggml_tensor * logits = ggml_mul_mat(ctx, w.head.logits_w(w.dec_embed.token_w), x);  // [vocab, B]
     if (w.head.bias != nullptr) {
         logits = ggml_add(ctx, logits, w.head.bias);
     }

src/arch/cohere/weights.cpp

@@ -437,6 +437,9 @@ transcribe_status build_cohere_weights(ggml_context * ctx_meta, const CohereHPar
 
     // ----- head -----
     GET_F32(weights.head.bias, "head.bias", vocab_size);
+    if (ggml_get_tensor(ctx_meta, "head.weight") != nullptr) {
+        GET_LIN(weights.head.weight, "head.weight", dec_h, vocab_size);
+    }
 
     return TRANSCRIBE_OK;
 }

src/arch/cohere/weights.h

@@ -212,9 +212,13 @@ struct CohereDecFinal {
     ggml_tensor * norm_b = nullptr;
 };
 
-// Head (bias only; weight is tied to dec_embed.token_w).
+// Head. The weight is tied to dec_embed.token_w unless the GGUF carries an
+// untied head.weight (fine-tunes such as syvai/hviske-v5*).
 struct CohereHead {
-    ggml_tensor * bias = nullptr;  // [vocab_size]
+    ggml_tensor * weight = nullptr;  // optional [dec_hidden, vocab_size]
+    ggml_tensor * bias   = nullptr;  // [vocab_size]
+
+    ggml_tensor * logits_w(ggml_tensor * token_w) const { return weight != nullptr ? weight : token_w; }
 };
 
 struct CohereWeights {