[Bugfix][Rust Frontend] Fix LogprobsTensors wire schema mismatch (#53939)

Signed-off-by: jungjiyu <libraryofjiyu@gmail.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: OpenAI Codex <codex@openai.com>

rust/src/engine-core-client/src/protocol/logprobs.rs

@@ -222,6 +222,7 @@ impl WireLogprobs {
                 data: WireArrayData::RawView(token_ranks.into()),
             },
             cu_num_generated_tokens: None,
+            cu_num_generated_tokens_tensor: None,
         })
     }
 
@@ -239,6 +240,15 @@ impl WireLogprobs {
             );
         }
 
+        // Unlike the sibling check above, don't Debug-print the payload:
+        // an opaque non-None value here may embed a full tensor blob.
+        if self.cu_num_generated_tokens_tensor.is_some() {
+            bail_ext_value_decode!(
+                "{field_prefix}.cu_num_generated_tokens_tensor: \
+                 expected None for per-request engine-core logprobs payload"
+            );
+        }
+
         let token_ids = array::decode_array2_u32(
             self.logprob_token_ids,
             &format!("{field_prefix}.logprob_token_ids"),

rust/src/engine-core-client/src/protocol/logprobs/tests.rs

@@ -87,6 +87,7 @@ fn inline_prompt_logprobs_value() -> Value {
         ndarray_value("float32", &[2, 3], probs),
         ndarray_value("int64", &[2], ranks),
         Value::Nil,
+        Value::Nil,
     ])
 }
 
@@ -240,6 +241,29 @@ fn decodes_inline_prompt_logprobs() {
     assert_eq!(logprobs, expected_prompt_logprobs());
 }
 
+#[test]
+fn rejects_non_none_cu_num_generated_tokens_tensor() {
+    let Value::Array(mut fields) = inline_prompt_logprobs_value() else {
+        panic!("inline_prompt_logprobs_value must be an array");
+    };
+    fields[4] = Value::from(42);
+
+    let frames = vec![Bytes::from(encode_value(&output_wire_with_custom_fields(
+        None,
+        Some(Value::Array(fields)),
+    )))];
+
+    let error = decode_engine_core_outputs(&frames).unwrap_err();
+    let crate::error::Error::ExtValueDecode { message } = &error else {
+        panic!("expected ValueDecodeExt");
+    };
+    assert_eq!(
+        message,
+        "new_prompt_logprobs_tensors.cu_num_generated_tokens_tensor: \
+         expected None for per-request engine-core logprobs payload"
+    );
+}
+
 #[test]
 fn decodes_big_endian_payloads() {
     let frames = vec![Bytes::from(encode_value(&output_wire_with_custom_fields(

rust/src/engine-core-client/src/protocol/logprobs/wire.rs

@@ -3,6 +3,7 @@
 
 use serde_tuple::{Deserialize_tuple, Serialize_tuple};
 
+use crate::protocol::OpaqueValue;
 use crate::protocol::tensor::WireNdArray;
 
 /// Python wire representation of `LogprobsLists` / `LogprobsTensors` before
@@ -13,7 +14,15 @@ use crate::protocol::tensor::WireNdArray;
 /// into semantic per-position logprobs records.
 ///
 /// Original Python definition:
-/// <https://github.com/vllm-project/vllm/blob/f22d6e026798a74e6542a52ef776c054f2de572a/vllm/v1/outputs.py#L23-L56>
+/// <https://github.com/vllm-project/vllm/blob/d5cadcee8641a9fcec15facb5a9157d157daa207/vllm/v1/outputs.py#L30-L106>
+///
+/// # Wire compatibility
+///
+/// Python's `LogprobsLists` / `LogprobsTensors` are `NamedTuple`s: every
+/// field is serialized positionally, and `Deserialize_tuple` here hard-fails
+/// on payloads longer than this struct. To preserve positional compatibility,
+/// append new Python fields and add matching trailing `#[serde(default)]`
+/// fields here in the same change.
 #[derive(Debug, Clone, PartialEq, Serialize_tuple, Deserialize_tuple)]
 pub struct WireLogprobs {
     /// Wire array with shape `[num_positions, max_num_logprobs + 1]`.
@@ -31,4 +40,12 @@ pub struct WireLogprobs {
     /// the semantic Rust decoder rejects any other value.
     #[serde(default)]
     pub cu_num_generated_tokens: Option<Vec<usize>>,
+    /// Device-only generation boundaries added to `LogprobsTensors` in
+    /// vllm-project/vllm#52242. Preserved only for wire compatibility:
+    /// NamedTuple encoding always emits the slot even when it is `None`,
+    /// so this field must exist for the 5-element payload to decode.
+    /// Scheduler-sliced per-request outputs should emit `None` here, and
+    /// the semantic Rust decoder rejects any other value.
+    #[serde(default)]
+    pub cu_num_generated_tokens_tensor: Option<OpaqueValue>,
 }

rust/src/engine-core-client/src/tests/python_compat.py

@@ -353,7 +353,8 @@ inline_prompt_logprobs = engine_outputs_wire(
                 ).tobytes(),
             ),
             ("int64", [2], np.array([3, 4], dtype=np.int64).tobytes()),
-            None,
+            None,  # cu_num_generated_tokens
+            None,  # cu_num_generated_tokens_tensor
         ),
     )
 )
@@ -376,7 +377,8 @@ multipart_prompt_logprobs = engine_outputs_wire(
                 ).tobytes(),
             ),
             ("int64", [2], np.array([3, 4], dtype=np.int64).tobytes()),
-            None,
+            None,  # cu_num_generated_tokens
+            None,  # cu_num_generated_tokens_tensor
         ),
     )
 )