perf: avoid heap allocs during token detokenization

This commit is contained in:
Evandro Fonseca Junior committed 2026-03-26 17:37:01 -03:00
1 parent b38c77f875
commit 14ab598ae1
1 file changed
+24 -16
+24 -16
View File
@@ -297,22 +297,30 @@ fn decode_token_piece_with<F>(mut call: F) -> String
where where
F: FnMut(*mut std::ffi::c_char, i32) -> i32, F: FnMut(*mut std::ffi::c_char, i32) -> i32,
{ {
let mut capacity = 64usize; let mut stack_buffer = [0u8; 256];
let written = call(
stack_buffer.as_mut_ptr() as *mut std::ffi::c_char,
stack_buffer.len() as i32,
);
if written >= 0 {
return String::from_utf8_lossy(&stack_buffer[..written as usize]).into_owned();
}
let mut heap_buffer = vec![0u8; (-written) as usize];
loop { loop {
let mut buffer = vec![0u8; capacity];
let written = call( let written = call(
buffer.as_mut_ptr() as *mut std::ffi::c_char, heap_buffer.as_mut_ptr() as *mut std::ffi::c_char,
capacity as i32, heap_buffer.len() as i32,
); );
if written < 0 { if written < 0 {
capacity = (-written) as usize; heap_buffer.resize((-written) as usize, 0);
continue; continue;
} }
let written = written as usize; return String::from_utf8_lossy(&heap_buffer[..written as usize]).into_owned();
return String::from_utf8_lossy(&buffer[..written]).into_owned();
} }
} }
@@ -321,17 +329,14 @@ mod tests {
use super::decode_token_piece_with; use super::decode_token_piece_with;
#[test] #[test]
fn retries_when_piece_buffer_is_too_small() { fn uses_stack_buffer_when_piece_fits() {
let mut calls = 0; let mut calls = 0;
let decoded = decode_token_piece_with(|buf, len| { let decoded = decode_token_piece_with(|buf, len| {
calls += 1; calls += 1;
if len < 17 {
return -17;
}
let bytes = b" Quantum computer"; let bytes = b" Quantum computer";
assert!(len >= bytes.len() as i32);
unsafe { unsafe {
std::ptr::copy_nonoverlapping(bytes.as_ptr(), buf as *mut u8, bytes.len()); std::ptr::copy_nonoverlapping(bytes.as_ptr(), buf as *mut u8, bytes.len());
} }
@@ -350,18 +355,21 @@ mod tests {
calls += 1; calls += 1;
if calls == 1 { if calls == 1 {
return -18; return -300;
} }
assert!(len >= 18); let bytes = b" Quantum computer rendered from a resized heap buffer";
let bytes = b" Quantum computer"; assert!(len >= bytes.len() as i32);
unsafe { unsafe {
std::ptr::copy_nonoverlapping(bytes.as_ptr(), buf as *mut u8, bytes.len()); std::ptr::copy_nonoverlapping(bytes.as_ptr(), buf as *mut u8, bytes.len());
} }
bytes.len() as i32 bytes.len() as i32
}); });
assert_eq!(decoded, " Quantum computer"); assert_eq!(
decoded,
" Quantum computer rendered from a resized heap buffer"
);
assert_eq!(calls, 2); assert_eq!(calls, 2);
} }
} }