renderers 0.1.12.dev1__tar.gz → 0.1.12.dev3__tar.gz

{renderers-0.1.12.dev1 → renderers-0.1.12.dev3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: renderers
-Version: 0.1.12.dev1
+Version: 0.1.12.dev3
 Summary: Chat template renderers — deterministic message-to-token conversion for LLM training
 License-Expression: Apache-2.0
 License-File: LICENSE

{renderers-0.1.12.dev1 → renderers-0.1.12.dev3}/renderers/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.1.12.dev1'
-__version_tuple__ = version_tuple = (0, 1, 12, 'dev1')
+__version__ = version = '0.1.12.dev3'
+__version_tuple__ = version_tuple = (0, 1, 12, 'dev3')
 
 __commit_id__ = commit_id = None

{renderers-0.1.12.dev1 → renderers-0.1.12.dev3}/renderers/base.py

@@ -2095,7 +2095,7 @@ def build_trajectory_step(
         "completion_mask": [True] * len(completion_ids),
         "completion_logprobs": [0.0] * len(completion_ids),
         "routed_experts": None,
-        "kept_tokens": None,
+        "sampling_mask": None,
     }
     if (
         full_rendered.multi_modal_data is not None

{renderers-0.1.12.dev1 → renderers-0.1.12.dev3}/renderers/client.py

@@ -28,7 +28,6 @@ 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":"'
 # vLLM uses this value both when sampled-token evidence is missing and as a
 # lower-bound clamp, so receiving it cannot prove the real logprob was returned.
 VLLM_LOGPROB_SENTINEL = -9999.0
@@ -129,12 +128,9 @@ def _strip_base64_field(raw: bytes, prefix: bytes) -> tuple[bytes, memoryview |
 
 def parse_generate_response(raw: bytes) -> dict[str, Any]:
     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
 
 
@@ -374,7 +370,9 @@ async def generate(
     parsed = renderer.parse_response(completion_ids, tools=tools)
 
     routed_experts = choice.get("routed_experts")
-    kept_tokens = choice.get("kept_tokens")
+    # vLLM's native kept-set sampling masks (``--return-sampling-mask``):
+    # one list of surviving vocab ids per completion token.
+    sampling_mask = choice.get("sampling_mask")
 
     # /inference/v1/generate returns finish_reason in {"stop","length",...} —
     # never "tool_calls" (a chat-completions concept). Promote stop→tool_calls
@@ -393,6 +391,7 @@ async def generate(
 
     return {
         "request_id": data.get("request_id") or "",
+        "usage": data.get("usage"),
         "prompt_ids": list(effective_prompt_ids or prompt_ids),
         "renderer_prompt_ids": list(prompt_ids) if content_parts else None,
         "mm_placeholders": mm_placeholders,
@@ -403,7 +402,7 @@ async def generate(
         "tool_calls": parsed.tool_calls,
         "finish_reason": finish_reason,
         "routed_experts": routed_experts,
-        "kept_tokens": kept_tokens,
+        "sampling_mask": sampling_mask,
         # 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.

{renderers-0.1.12.dev1 → renderers-0.1.12.dev3}/tests/test_client.py

@@ -68,6 +68,7 @@ class _FakeClient:
     def __init__(self):
         self.calls = []
         self.base_url = "http://fake-host:8000/v1"
+        self.usage = None
         routed_experts = np.array([[[1]], [[2]]], dtype=np.uint8)
         self.choice = {
             "index": 0,
@@ -93,6 +94,8 @@ class _FakeClient:
             "request_id": "gen-test",
             "choices": [self.choice],
         }
+        if self.usage is not None:
+            payload["usage"] = self.usage
         if body and body.get("content_parts"):
             payload["prompt_token_ids"] = [1, 2, 2, 3]
             payload["mm_placeholders"] = {"image": [{"offset": 1, "length": 2}]}
@@ -114,8 +117,22 @@ def _run_generate(client, renderer=None):
     )
 
 
-def test_generate_builds_request_body_and_parses_response():
+@pytest.mark.parametrize(
+    "usage",
+    [
+        None,
+        {
+            "prompt_tokens": 3,
+            "completion_tokens": 2,
+            "total_tokens": 5,
+            "prompt_tokens_details": {"cached_tokens": 2},
+            "completion_tokens_details": {"reasoning_tokens": 1},
+        },
+    ],
+)
+def test_generate_builds_request_body_and_parses_response(usage):
     client = _FakeClient()
+    client.usage = usage
     renderer = _FakeRenderer()
 
     result = asyncio.run(
@@ -167,6 +184,7 @@ def test_generate_builds_request_body_and_parses_response():
     assert result["routed_experts"]["data"].tobytes() == base64.b64encode(b"\x01\x02")
     assert result["multi_modal_data"] is None
     assert result["request_id"] == "gen-test"
+    assert result["usage"] == usage
     # Per-token attribution from the renderer surfaces on the result so
     # downstream consumers (verifiers RendererClient → prime-rl) can
     # build selective loss masks without a second render pass.