Add LP/v1 logprobs deserialization for tracing gateway payloads (FIR-21499).

SunnySoldier357 · cursoragent · SunnySoldier357 · commit ac3db7b2b243 · 2026-05-21T17:59:52.000-07:00
Decode logprobs payloads into completion_logprobs and Message.logprobs on
EvaluationRow. Pop base_url from completion_params before OpenAI SDK calls
so dev inference API can be encoded in gateway tracing URLs.

Co-authored-by: Cursor &lt;cursoragent@cursor.com&gt;
diff --git a/eval_protocol/adapters/fireworks_tracing.py b/eval_protocol/adapters/fireworks_tracing.py
@@ -116,6 +116,34 @@ def convert_trace_dict_to_evaluation_row(
                 except Exception as e:
                     logger.warning("Failed to decompress R3 payload for trace %s: %s", trace.get("id"), e)
 
+            logprobs_payload = payloads.get("logprobs")
+            if isinstance(logprobs_payload, dict) and logprobs_payload.get("data"):
+                try:
+                    from .lp_deserializer import decompress_and_parse_lp
+
+                    logprobs, token_ids, lp_meta = decompress_and_parse_lp(logprobs_payload["data"])
+                    if execution_metadata.extra is None:
+                        execution_metadata.extra = {}
+                    execution_metadata.extra["completion_logprobs"] = logprobs
+                    if token_ids is not None:
+                        execution_metadata.extra["completion_token_ids"] = token_ids
+                    execution_metadata.extra["logprobs_metadata"] = lp_meta
+
+                    for i in range(len(messages) - 1, -1, -1):
+                        if messages[i].role == "assistant":
+                            content_entries = [{"logprob": lp} for lp in logprobs]
+                            if token_ids is not None:
+                                for entry, tid in zip(content_entries, token_ids):
+                                    entry["token_id"] = tid
+                            messages[i].logprobs = {"content": content_entries}
+                            break
+                except Exception as e:
+                    logger.warning(
+                        "Failed to decompress logprobs payload for trace %s: %s",
+                        trace.get("id"),
+                        e,
+                    )
+
         return EvaluationRow(
             messages=messages,
             tools=tools,
diff --git a/eval_protocol/adapters/lp_deserializer.py b/eval_protocol/adapters/lp_deserializer.py
@@ -0,0 +1,113 @@
+"""LP/v1 binary deserializer for per-token logprobs payloads.
+
+Implements the inverse of the tracing gateway's ``logprobs_serializer.serialize_logprobs``.
+See that module for the full header specification.
+"""
+
+from __future__ import annotations
+
+import base64
+import struct
+from typing import Any, Dict, List, Optional, Tuple
+
+import zstandard as zstd
+
+MAGIC = b"LP01"
+HEADER_VERSION = 1
+MISSING_TOKEN_ID = -1
+ENTRY_FORMAT = "<if"
+ENTRY_SIZE = struct.calcsize(ENTRY_FORMAT)  # 8 bytes
+HEADER_FORMAT = "<4sBBHIIQ"
+HEADER_SIZE = struct.calcsize(HEADER_FORMAT)  # 24 bytes
+
+
+def _parse_header(raw: bytes) -> Dict[str, Any]:
+    if len(raw) < HEADER_SIZE:
+        raise ValueError(f"Payload too short for lp/v1 header: {len(raw)} < {HEADER_SIZE}")
+
+    (
+        magic,
+        version,
+        flags,
+        reserved_u16,
+        token_count,
+        body_byte_length,
+        reserved_u64,
+    ) = struct.unpack(HEADER_FORMAT, raw[:HEADER_SIZE])
+
+    if magic != MAGIC:
+        raise ValueError(f"Bad LP/v1 magic: {magic!r}")
+    if version != HEADER_VERSION:
+        raise ValueError(f"Unsupported lp/v1 header version: {version}")
+
+    return {
+        "flags": flags,
+        "reserved_u16": reserved_u16,
+        "token_count": token_count,
+        "body_byte_length": body_byte_length,
+        "reserved_u64": reserved_u64,
+    }
+
+
+def _parse_logprobs_raw(raw: bytes) -> Tuple[List[float], Optional[List[int]], Dict[str, Any]]:
+    header = _parse_header(raw)
+    token_count = header["token_count"]
+    body_byte_length = header["body_byte_length"]
+
+    if token_count == 0:
+        raise ValueError("LP/v1 token_count must be > 0")
+    if body_byte_length != token_count * ENTRY_SIZE:
+        raise ValueError(
+            f"body_byte_length ({body_byte_length}) != token_count * {ENTRY_SIZE} "
+            f"({token_count * ENTRY_SIZE})"
+        )
+
+    expected_len = HEADER_SIZE + body_byte_length
+    if len(raw) != expected_len:
+        raise ValueError(f"LP/v1 payload length mismatch: {len(raw)} != {expected_len}")
+
+    logprobs: List[float] = []
+    token_ids: List[int] = []
+    all_token_ids_valid = True
+    offset = HEADER_SIZE
+    for _ in range(token_count):
+        wire_id, logprob = struct.unpack(ENTRY_FORMAT, raw[offset : offset + ENTRY_SIZE])
+        offset += ENTRY_SIZE
+        logprobs.append(logprob)
+        if wire_id == MISSING_TOKEN_ID:
+            all_token_ids_valid = False
+            token_ids.append(wire_id)
+        else:
+            token_ids.append(wire_id)
+
+    metadata: Dict[str, Any] = {
+        "scope": "completion_only",
+        "completion_token_count": token_count,
+        "all_token_ids_valid": all_token_ids_valid,
+    }
+    header.update(metadata)
+    ids_out: Optional[List[int]] = token_ids if all_token_ids_valid else None
+    return logprobs, ids_out, header
+
+
+def parse_logprobs(raw: bytes) -> Tuple[List[float], Optional[List[int]], Dict[str, Any]]:
+    """Parse uncompressed LP/v1 bytes into logprobs, optional token ids, and metadata."""
+    return _parse_logprobs_raw(raw)
+
+
+def decompress_and_parse_lp(data_b64: str) -> Tuple[List[float], Optional[List[int]], Dict[str, Any]]:
+    """Decompress and unpack an LP/v1 payload into completion logprobs and token ids.
+
+    Args:
+        data_b64: Base64-encoded zstd-compressed LP binary blob from
+            ``payloads.logprobs.data``.
+
+    Returns:
+        ``(logprobs, token_ids, metadata)`` where ``logprobs`` is per-completion-token
+        scalars, ``token_ids`` is ``None`` if any wire id was ``MISSING_TOKEN_ID``,
+        and ``metadata`` includes ``all_token_ids_valid`` and ``completion_token_count``.
+    """
+    compressed = base64.b64decode(data_b64)
+    decompressor = zstd.ZstdDecompressor()
+    raw = decompressor.decompress(compressed)
+    return _parse_logprobs_raw(raw)
diff --git a/eval_protocol/pytest/tracing_utils.py b/eval_protocol/pytest/tracing_utils.py
@@ -103,8 +103,8 @@ def build_init_request(
     if not completion_params_dict.get("model"):
         raise ValueError("Model must be provided in completion_params")
 
-    # Extract base_url from completion_params
-    completion_params_base_url: Optional[str] = completion_params_dict.get("base_url")
+    # Extract base_url from completion_params (encoded into gateway path, not sent to OpenAI SDK)
+    completion_params_base_url: Optional[str] = completion_params_dict.pop("base_url", None)
 
     # Strip non-OpenAI fields from messages
     # Use dump_mdoel_for_chat_completion_request() to automatically exclude unsupported fields (weight, control_plane_step, reasoning_content)
diff --git a/tests/adapters/test_fireworks_tracing_logprobs.py b/tests/adapters/test_fireworks_tracing_logprobs.py
@@ -0,0 +1,93 @@
+"""Tests for logprobs payload handling in fireworks_tracing adapter."""
+
+from __future__ import annotations
+
+import base64
+import struct
+
+import pytest
+import zstandard as zstd
+
+pytest.importorskip("mcp")
+
+from eval_protocol.adapters.fireworks_tracing import convert_trace_dict_to_evaluation_row
+from eval_protocol.adapters.lp_deserializer import (
+    ENTRY_FORMAT,
+    ENTRY_SIZE,
+    HEADER_FORMAT,
+    MAGIC,
+    MISSING_TOKEN_ID,
+)
+
+
+def _lp_b64(tokens: list[tuple[int, float]]) -> str:
+    token_count = len(tokens)
+    body_byte_length = token_count * ENTRY_SIZE
+    header = struct.pack(
+        HEADER_FORMAT,
+        MAGIC,
+        1,
+        0,
+        0,
+        token_count,
+        body_byte_length,
+        0,
+    )
+    body = b"".join(struct.pack(ENTRY_FORMAT, tid, lp) for tid, lp in tokens)
+    raw = header + body
+    compressed = zstd.ZstdCompressor().compress(raw)
+    return base64.b64encode(compressed).decode("ascii")
+
+
+def _base_trace(*, with_token_ids: bool = True) -> dict:
+    tokens = [(10, -0.1), (11, -0.2)] if with_token_ids else [(MISSING_TOKEN_ID, -0.1), (12, -0.2)]
+    return {
+        "id": "trace-1",
+        "input": {
+            "messages": [
+                {"role": "user", "content": "hi"},
+                {"role": "assistant", "content": "hello"},
+            ],
+        },
+        "output": {"role": "assistant", "content": "hello"},
+        "payloads": {
+            "logprobs": {
+                "data": _lp_b64(tokens),
+                "manifest": {"PayloadVersion": "lp/v1"},
+            },
+        },
+    }
+
+
+class TestConvertTraceLogprobs:
+    def test_attaches_completion_logprobs_and_message_logprobs(self):
+        row = convert_trace_dict_to_evaluation_row(_base_trace())
+        assert row is not None
+
+        extra = row.execution_metadata.extra
+        assert extra is not None
+        assert extra["completion_logprobs"] == pytest.approx([-0.1, -0.2])
+        assert extra["completion_token_ids"] == [10, 11]
+
+        assistant = row.messages[-1]
+        assert assistant.role == "assistant"
+        content = assistant.logprobs["content"]
+        assert len(content) == len(extra["completion_logprobs"])
+        assert content[0]["token_id"] == 10
+        assert content[1]["token_id"] == 11
+        assert content[0]["logprob"] == pytest.approx(-0.1)
+        assert content[1]["logprob"] == pytest.approx(-0.2)
+
+    def test_omits_token_id_keys_when_any_missing(self):
+        row = convert_trace_dict_to_evaluation_row(_base_trace(with_token_ids=False))
+        assert row is not None
+
+        extra = row.execution_metadata.extra
+        assert "completion_logprobs" in extra
+        assert "completion_token_ids" not in extra
+
+        content = row.messages[-1].logprobs["content"]
+        assert len(content) == 2
+        assert all("token_id" not in entry for entry in content)
+        assert content[0]["logprob"] == pytest.approx(-0.1)
+        assert content[1]["logprob"] == pytest.approx(-0.2)
diff --git a/tests/adapters/test_lp_deserializer.py b/tests/adapters/test_lp_deserializer.py
@@ -0,0 +1,78 @@
+"""Tests for LP/v1 binary deserializer (gateway-compatible)."""
+
+from __future__ import annotations
+
+import base64
+import struct
+
+import pytest
+import zstandard as zstd
+
+from eval_protocol.adapters.lp_deserializer import (
+    ENTRY_FORMAT,
+    ENTRY_SIZE,
+    HEADER_FORMAT,
+    HEADER_SIZE,
+    MAGIC,
+    MISSING_TOKEN_ID,
+    decompress_and_parse_lp,
+    parse_logprobs,
+)
+
+# Golden raw bytes: two tokens (7, -0.25) and (8, -0.5) — must match gateway serializer.
+GOLDEN_RAW_HEX = (
+    "4c503031010000000200000010000000000000000000000007000000000080be"
+    "08000000000000bf"
+)
+
+
+def _build_raw(tokens: list[tuple[int, float]]) -> bytes:
+    token_count = len(tokens)
+    body_byte_length = token_count * ENTRY_SIZE
+    header = struct.pack(
+        HEADER_FORMAT,
+        MAGIC,
+        1,
+        0,
+        0,
+        token_count,
+        body_byte_length,
+        0,
+    )
+    body = b"".join(struct.pack(ENTRY_FORMAT, tid, lp) for tid, lp in tokens)
+    return header + body
+
+
+def _compress_b64(raw: bytes) -> str:
+    return base64.b64encode(zstd.ZstdCompressor().compress(raw)).decode("ascii")
+
+
+class TestParseLogprobs:
+    def test_golden_bytes_match_gateway(self):
+        raw = bytes.fromhex(GOLDEN_RAW_HEX)
+        logprobs, token_ids, meta = parse_logprobs(raw)
+        assert logprobs == [-0.25, -0.5]
+        assert token_ids == [7, 8]
+        assert meta["all_token_ids_valid"] is True
+        assert meta["token_count"] == 2
+
+    def test_missing_token_id_omits_token_ids_list(self):
+        raw = _build_raw([(MISSING_TOKEN_ID, -0.3), (42, -0.4)])
+        logprobs, token_ids, meta = parse_logprobs(raw)
+        assert logprobs == pytest.approx([-0.3, -0.4])
+        assert token_ids is None
+        assert meta["all_token_ids_valid"] is False
+
+    def test_decompress_and_parse_round_trip(self):
+        raw = bytes.fromhex(GOLDEN_RAW_HEX)
+        b64 = _compress_b64(raw)
+        logprobs, token_ids, meta = decompress_and_parse_lp(b64)
+        assert logprobs == [-0.25, -0.5]
+        assert token_ids == [7, 8]
+        assert meta["scope"] == "completion_only"
+
+    def test_rejects_bad_magic(self):
+        raw = _build_raw([(1, -0.1)])
+        bad = b"XXXX" + raw[4:]
+        with pytest.raises(ValueError, match="Bad LP/v1 magic"):
+            parse_logprobs(bad)
