renderers 0.1.8.dev58__tar.gz → 0.1.8.dev61__tar.gz

{renderers-0.1.8.dev58 → renderers-0.1.8.dev61}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: renderers
-Version: 0.1.8.dev58
+Version: 0.1.8.dev61
 Summary: Chat template renderers — deterministic message-to-token conversion for LLM training
 License-Expression: Apache-2.0
 License-File: LICENSE

{renderers-0.1.8.dev58 → renderers-0.1.8.dev61}/renderers/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.1.8.dev58'
-__version_tuple__ = version_tuple = (0, 1, 8, 'dev58')
+__version__ = version = '0.1.8.dev61'
+__version_tuple__ = version_tuple = (0, 1, 8, 'dev61')
 
 __commit_id__ = commit_id = None

{renderers-0.1.8.dev58 → renderers-0.1.8.dev61}/renderers/base.py

@@ -2006,6 +2006,7 @@ def build_trajectory_step(
         "completion_mask": [True] * len(completion_ids),
         "completion_logprobs": [0.0] * len(completion_ids),
         "routed_experts": None,
+        "kept_tokens": None,
     }
     if (
         full_rendered.multi_modal_data is not None

{renderers-0.1.8.dev58 → renderers-0.1.8.dev61}/renderers/client.py

@@ -32,6 +32,7 @@ from renderers.base import (
 
 _request_logger = logging.getLogger("renderers.client")
 ROUTED_EXPERTS_DATA_PREFIX = b'"routed_experts":{"data":"'
+KEPT_TOKENS_IDS_PREFIX = b'"kept_tokens":{"ids":"'
 
 
 class OverlongPromptError(Exception):
@@ -121,23 +122,31 @@ async def _maybe_offload(renderer: Renderer | RendererPool, fn):
     return fn()
 
 
-def strip_routed_experts_data(raw: bytes) -> tuple[bytes, memoryview | None]:
-    data_start = raw.find(ROUTED_EXPERTS_DATA_PREFIX)
+def _strip_base64_field(raw: bytes, prefix: bytes) -> tuple[bytes, memoryview | None]:
+    """Splice a large base64 string field out of raw JSON bytes.
+
+    Avoids json-decoding megabytes of base64; the returned memoryview
+    references ``raw`` and is re-inserted into the parsed payload.
+    """
+    data_start = raw.find(prefix)
     if data_start < 0:
         return raw, None
 
-    data_start += len(ROUTED_EXPERTS_DATA_PREFIX)
+    data_start += len(prefix)
     data_end = raw.index(b'"', data_start)
-    routed_data = memoryview(raw)[data_start:data_end]
+    data = memoryview(raw)[data_start:data_end]
     stripped = raw[:data_start] + raw[data_end:]
-    return stripped, routed_data
+    return stripped, data
 
 
 def parse_generate_response(raw: bytes) -> dict[str, Any]:
-    stripped, routed_data = strip_routed_experts_data(raw)
+    stripped, routed_data = _strip_base64_field(raw, ROUTED_EXPERTS_DATA_PREFIX)
+    stripped, kept_ids_data = _strip_base64_field(stripped, KEPT_TOKENS_IDS_PREFIX)
     payload: dict[str, Any] = json.loads(stripped)
     if routed_data is not None:
         payload["choices"][0]["routed_experts"]["data"] = routed_data
+    if kept_ids_data is not None:
+        payload["choices"][0]["kept_tokens"]["ids"] = kept_ids_data
     return payload
 
 
@@ -293,6 +302,7 @@ async def generate(
     completion_logprobs = [float(c.get("logprob") or 0.0) for c in content_lp or []]
 
     routed_experts = choice.get("routed_experts")
+    kept_tokens = choice.get("kept_tokens")
 
     # /inference/v1/generate returns finish_reason in {"stop","length",...} —
     # never "tool_calls" (a chat-completions concept). Promote stop→tool_calls
@@ -319,6 +329,7 @@ async def generate(
         "tool_calls": parsed.tool_calls,
         "finish_reason": finish_reason,
         "routed_experts": routed_experts,
+        "kept_tokens": kept_tokens,
         # The mm sidecar consumed on the request side, surfaced back so
         # callers can persist it on the trajectory step for downstream
         # multi-turn bridging and training-sample construction.