Implement top_p / nucleus sampling (#819)

* Implement top_p / nucleus sampling

* Update changelog

* rustfmt

* Add tests

* Fix clippy warning

* Fix another clippy error

1 files changed, 29 insertions, 0 deletions

diff --git a/candle-transformers/tests/generation_tests.rs b/candle-transformers/tests/generation_tests.rs
new file mode 100644
index 00000000..76f994d0
--- /dev/null
+++ b/candle-transformers/tests/generation_tests.rs
@@ -0,0 +1,29 @@
+use candle::{Device, Result, Tensor};
+use candle_transformers::generation::LogitsProcessor;
+
+#[test]
+fn sample_with_zero_temperature() -> Result<()> {
+    let mut logits_process = LogitsProcessor::new(1337, None, None);
+    let logits = Tensor::new(&[0.1, 0.2, 0.3, 0.4], &Device::Cpu)?;
+    let token = logits_process.sample(&logits)?;
+    assert_eq!(token, 3);
+    Ok(())
+}
+
+#[test]
+fn sample_with_temperature() -> Result<()> {
+    let mut logits_process = LogitsProcessor::new(42, Some(0.9), None);
+    let logits = Tensor::new(&[0.1, 0.2, 0.3, 0.4], &Device::Cpu)?;
+    let token = logits_process.sample(&logits)?;
+    assert_eq!(token, 0);
+    Ok(())
+}
+
+#[test]
+fn sample_with_top_p() -> Result<()> {
+    let mut logits_process = LogitsProcessor::new(42, Some(1.0), Some(0.5));
+    let logits = Tensor::new(&[0.1, 0.2, 0.3, 0.4], &Device::Cpu)?;
+    let token = logits_process.sample(&logits)?;
+    assert_eq!(token, 2);
+    Ok(())
+}