Add BF16 support for FlashQwen3

Author: alvarobartt

backends/candle/src/lib.rs

@@ -535,7 +535,7 @@ impl CandleBackend {
             }
             #[cfg(feature = "cuda")]
             (Config::Qwen3(config), Device::Cuda(_)) => {
-                if dtype != DType::F16
+                if (dtype != DType::F16 && dtype != DType::BF16)
                     || !cfg!(any(feature = "flash-attn", feature = "flash-attn-v1"))
                     || &std::env::var("USE_FLASH_ATTENTION")
                         .unwrap_or("True".to_string())

backends/candle/src/models/flash_qwen3.rs

@@ -300,8 +300,8 @@ impl FlashQwen3Model {
             _ => candle::bail!("FlashQwen3 requires Cuda"),
         }
 
-        if vb.dtype() != DType::F16 {
-            candle::bail!("FlashQwen3 requires DType::F16")
+        if vb.dtype() != DType::F16 && vb.dtype() != DType::BF16 {
+            candle::bail!("FlashQwen3 requires DType::F16 or DType::BF16")
         }
 
         let pool = match model_type {