fix mask bug

src/models/common/mod.rs

@@ -146,7 +146,7 @@ impl AttentionNobias {
                 let attn_weights = (attn_weights * scale)?;
                 let attn_weights = match attention_mask {
                     None => attn_weights,
-                    Some(mask) => attn_weights.broadcast_add(mask)?,
+                    Some(mask) => attn_weights.broadcast_add(&mask.to_dtype(attn_weights.dtype())?)?,
                 };
                 let attn_weights = candle_nn::ops::softmax_last_dim(&attn_weights)?;
                 let attn_weights = attn_weights.matmul(&value_states)?;

src/models/minicpm4/model.rs

@@ -233,7 +233,7 @@ impl MiniCPMModel {
                 Some(&prepare_causal_attention_mask(
                     bs,
                     seq_len,
-                    position_id,
+                    0,
                     input_ids.device(),
                 )?)
             }
@@ -267,7 +267,7 @@ impl MiniCPMModel {
                 Some(&prepare_causal_attention_mask(
                     bs,
                     seq_len,
-                    position_id,
+                    0,
                     input_ids.device(),
                 )?)
             }

src/models/voxcpm/generate.rs

@@ -130,7 +130,7 @@ impl VoxCPMGenerate {
         Ok(audio)
     }
     pub fn generate_simple(&mut self, target_text: String) -> Result<Tensor> {
-        let audio = self.generate(target_text, None, None, 2, 1000, 10, 2.0, false, 6.0)?;
+        let audio = self.generate(target_text, None, None, 2, 100, 10, 2.0, false, 6.0)?;
         Ok(audio)
     }
     pub fn generate(

src/models/voxcpm/minicpm4.rs

@@ -266,7 +266,7 @@ impl MiniCPMModel {
                 Some(&prepare_causal_attention_mask(
                     bs,
                     seq_len,
-                    position_id,
+                    0,
                     input_embeds.device(),
                 )?)
             }
@@ -294,7 +294,7 @@ impl MiniCPMModel {
                 Some(&prepare_causal_attention_mask(
                     bs,
                     seq_len,
-                    position_id,
+                    0,
                     input_embeds.device(),
                 )?)
             }

src/utils/tensor_utils.rs

@@ -21,14 +21,14 @@ pub fn prepare_causal_attention_mask(
         .collect();
     let mask = Tensor::from_slice(&mask, (tgt_len, tgt_len), device)?;
     let mask = if seqlen_offset > 0 {
-        let mask0 = Tensor::zeros((tgt_len, seqlen_offset), DType::U32, device)?;
+        let mask0 = Tensor::zeros((tgt_len, seqlen_offset), DType::F32, device)?;
         Tensor::cat(&[&mask0, &mask], D::Minus1)?
     } else {
         mask
     };
     let mask = mask
         .expand((b_size, 1, tgt_len, tgt_len + seqlen_offset))?
-        .to_dtype(DType::U32)?;
+        .to_dtype(DType::F32)?;
     Ok(mask)
 }
 

tests/test_minicpm4.rs

@@ -21,7 +21,7 @@ fn minicpm_generate() -> Result<()> {
         "messages": [
             {
                 "role": "user",
-                "content": "你好啊，你是谁"
+                "content": "你吃饭了没"
             }
         ]
     }