verifiers 0.3.2.dev102__py3-none-any.whl → 0.3.2.dev105__py3-none-any.whl

verifiers/v1/cli/dashboard/eval.py

@@ -297,10 +297,10 @@ def _push_footer(push: "PushState | None") -> Group | None:
         line.append(f"  not closed out: {push.error}", style="red")
     elif push.error:
         line = Text(f"Trace push failed ({push.error})", style="red", overflow="fold")
-    elif push.incomplete and push.url:
-        # Closed out, but the uploader lost records: what landed is there, say what didn't.
-        line = Text(f"Traces pushed ({push.url})", style="white", overflow="fold")
-        line.append(f"  incomplete: {push.incomplete}", style="red")
+    elif (incomplete := push.incomplete) and push.url:
+        label = "Traces pushed" if push.finished else "Pushing traces"
+        line = Text(f"{label} ({push.url})", style="white", overflow="fold")
+        line.append(f"  incomplete: {incomplete}", style="red")
     elif not push.finished:
         line = Text(f"Pushing traces ({push.url})", style="dim", overflow="fold")
     else:

verifiers/v1/serve/delta.py

@@ -4,12 +4,12 @@ The worker streams a served episode as it grows. Each trace announces its own ch
 (`Trace.notify`, fired by the rollout at every phase change and by the interception proxy
 after every recorded turn); the `DeltaStreamer` then diffs the run's live traces against
 what it has already sent and ships only the new part — the trace header once, then
-appended nodes / calls / errors, semantic links landing on earlier nodes, the scalar
+appended nodes / calls / errors, semantic links and routing-row repairs on earlier nodes, the scalar
 fields whose value changed (timing spans, stop condition, rewards, ...), and the
 `pending` preview — the messages of the request in flight that no node holds yet, so a
 watcher sees a tool result before the model has answered it. The `Trace` is
-append-only at turn granularity (a turn's nodes are committed complete, with their
-tokens), so apart from the preview, which a committed turn repeats, every byte of the
+append-only at turn granularity except for links and the last routing row of a node,
+which the next prefill can repair. Apart from these and the preview, every byte of the
 episode crosses the wire once and the stream costs about what a single reply would; the
 reply that ends the run carries only the episode head and per-trace counts the client
 checks its assembly against. A cursor advances only once its delta is on the wire, so a
@@ -29,8 +29,10 @@ from collections.abc import Awaitable, Callable
 from typing import TYPE_CHECKING, Any, Self
 
 import msgpack
+import numpy as np
 from pydantic import BaseModel
 
+from verifiers.v1.graph import _decode_ndarray, _encode_ndarray
 from verifiers.v1.serve.encoding import msgpack_encoder
 
 if TYPE_CHECKING:
@@ -73,7 +75,7 @@ def pack(payload: Any) -> bytes:
 
 
 def unpack(data: bytes) -> Any:
-    # `links` is keyed by node index (int).
+    # Node updates are keyed by node index (int).
     return msgpack.unpackb(data, raw=False, strict_map_key=False)
 
 
@@ -90,6 +92,8 @@ class TraceCursor:
         self.sent = dict.fromkeys(LIST_FIELDS, 0)
         self.links: list[int] = []
         """Per sent node, how many of its semantic links went out with or after it."""
+        self.final_rows: dict[int, bytes] = {}
+        """Per node that carries routing, its final row as packed when last sent."""
         self.scalars: dict[str, bytes] = {}
         self.pending: bytes = pack([])
 
@@ -157,6 +161,31 @@ class DeltaStreamer:
                 else:
                     self.cursors[trace_id] = cursor
 
+    def _maybe_add_routing_repairs(
+        self,
+        delta: dict[str, Any],
+        trace: Trace,
+        cursor: TraceCursor,
+        sent_nodes: int,
+    ) -> None:
+        """Add final rows repaired since they were sent, keyed by node index.
+
+        Record every node's current row on the cursor. `sent_nodes` is the pre-flush
+        count, so a node first sent in this delta is never reported as a repair.
+        """
+        repairs: dict[int, dict] = {}
+        for index, node in enumerate(trace.nodes):
+            if node.routed_experts is None:
+                continue
+            row = _encode_ndarray(node.routed_experts[-1:])
+            packed = pack(row)
+            if cursor.final_rows.get(index) != packed:
+                cursor.final_rows[index] = packed
+                if index < sent_nodes:
+                    repairs[index] = row
+        if repairs:
+            delta["routing_repairs"] = repairs
+
     def diff(self) -> list[tuple[str, dict, TraceCursor | None]]:
         """Each trace's delta against its sent cursor, with the cursor as it stands once
         that delta is sent (None for a discard). Nothing here is committed: `flush`
@@ -183,6 +212,7 @@ class DeltaStreamer:
                     cursor.links[index] = len(node_links)
             if links:
                 delta["links"] = links
+            self._maybe_add_routing_repairs(delta, trace, cursor, cursor.sent["nodes"])
             for field in LIST_FIELDS:
                 items = getattr(trace, field)
                 sent = cursor.sent[field]
@@ -220,6 +250,20 @@ class EpisodeAssembly:
     def __init__(self) -> None:
         self.traces: dict[str, dict] = {}
 
+    def _maybe_apply_routing_repairs(self, delta: dict[str, Any], trace: dict) -> None:
+        """Replace repaired final rows in nodes the client already holds.
+
+        Rebuild each array: a repair can widen its dtype, and decoded rows alias
+        the original delta's bytes, which must remain unchanged for consumers.
+        """
+        for index, row in (delta.get("routing_repairs") or {}).items():
+            node = trace["nodes"][int(index)]
+            node["routed_experts"] = _encode_ndarray(
+                np.concatenate(
+                    [_decode_ndarray(node["routed_experts"])[:-1], _decode_ndarray(row)]
+                )
+            )
+
     def apply(self, delta: dict) -> None:
         trace_id = delta["trace"]
         if delta.get("discard"):
@@ -234,6 +278,7 @@ class EpisodeAssembly:
             }
         for index, links in (delta.get("links") or {}).items():
             trace["nodes"][int(index)]["semantic_parents"].extend(links)
+        self._maybe_apply_routing_repairs(delta, trace)
         # a later `links` delta grows a node's semantic_parents in place, so the node
         # is copied: the delta stays as it was when the caller received it
         if "nodes" in delta:

verifiers/v1/utils/platform.py

@@ -24,9 +24,11 @@ class PushState:
 
     run: pr.Run | None = None
     error: str | None = None
-    incomplete: str | None = None
-    """Set when the run closed out but the uploader lost records on the way: the run
-    exists and holds what did land, so the footer says so rather than "failed"."""
+
+    @property
+    def incomplete(self) -> str | None:
+        """Current upload errors and record losses, including while the run is active."""
+        return _losses(self.run) if self.run is not None else None
 
     @property
     def url(self) -> str | None:
@@ -157,9 +159,8 @@ def _close(
         if state.error is None:
             state.error = f"{type(e).__name__}: {e}"
     else:
-        state.incomplete = _losses(run)
-        if state.incomplete:
-            logger.warning("--push: %s, but %s", status.value, state.incomplete)
+        if incomplete := state.incomplete:
+            logger.warning("--push: %s, but %s", status.value, incomplete)
         if run.url:
             # The run's own status: `finish()` is a no-op once another caller closed it.
             logger.info("--push: %s -> %s", run.status.value, run.url)

{verifiers-0.3.2.dev102.dist-info → verifiers-0.3.2.dev105.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev102
+Version: 0.3.2.dev105
 Summary: Verifiers: Environments for LLM Reinforcement Learning
 Project-URL: Homepage, https://github.com/primeintellect-ai/verifiers
 Project-URL: Documentation, https://github.com/primeintellect-ai/verifiers
@@ -35,7 +35,7 @@ Requires-Dist: msgpack>=1.1.2
 Requires-Dist: numpy>=2.1.0
 Requires-Dist: openai<3.0.0,>=2.54.0
 Requires-Dist: prime-pydantic-config[toml]>=0.4.3
-Requires-Dist: prime-runs>=0.1.2
+Requires-Dist: prime-runs>=0.1.3
 Requires-Dist: prime-sandboxes>=0.2.39
 Requires-Dist: prime-tunnel>=0.1.8
 Requires-Dist: pydantic>=2.12.3

{verifiers-0.3.2.dev102.dist-info → verifiers-0.3.2.dev105.dist-info}/RECORD

@@ -29,7 +29,7 @@ verifiers/v1/cli/resume.py,sha256=hqV5AxxqeIInJeEZQerQznuijhMNijFEFcHZIltfLdE,47
 verifiers/v1/cli/validate.py,sha256=K7mCz-ZI7sq69MQPg6QvreHOuEj-bgMln69a28kDTvA,17179
 verifiers/v1/cli/dashboard/__init__.py,sha256=v-baMxQuWxOCsbU7-p_jj2Q9BUnTN-TWi1q2hK6rU2s,198
 verifiers/v1/cli/dashboard/base.py,sha256=kUP93zJSIVLptSbnWX6MOy8kGg63fpeAzPqakCpPyDc,3547
-verifiers/v1/cli/dashboard/eval.py,sha256=D4iX69Ou8hHcv9_Bn2sPKF-9Zg89q00XLvNaLCajlNg,38369
+verifiers/v1/cli/dashboard/eval.py,sha256=HEXPEegNXdgU9zAIuHzsWsoSIHyiOcDZh4tCGFNxKAk,38353
 verifiers/v1/cli/dashboard/replay.py,sha256=_X5up9MJsRbd_sWe4LgNTPNT8k1ip4xNpo50hxI3Ljk,2697
 verifiers/v1/cli/dashboard/validate.py,sha256=xrpK3Y90JsoDCQYLLvvsnJ4CPRMvSSJIghfHK62E6FQ,3687
 verifiers/v1/cli/eval/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -149,7 +149,7 @@ verifiers/v1/runtimes/docker/__init__.py,sha256=2YJtjOVSA4OgxXgeIlFW3Ps01KEfLNT_
 verifiers/v1/runtimes/docker/egress.py,sha256=iFH_1XotlEM-OCaW7GNu3TWVEq1sqesKYweyVBX8g90,14360
 verifiers/v1/serve/__init__.py,sha256=wOKuwfugzmDT4yIniwzCXlf4G6tv6Cg8Kc--WMa38Bs,553
 verifiers/v1/serve/client.py,sha256=E2y2x4qqYveWRWC9VZYQEu19aMEryPWTdXnYzbq5uTA,9829
-verifiers/v1/serve/delta.py,sha256=UnDhGqXdJPeRgRJ26j8xkHqU_HtBfk8TdzAvNezaWKA,11525
+verifiers/v1/serve/delta.py,sha256=qhOUOYEt1wVzeaJ9IhjgU6ks5BkYOEtcFfGnuYxFEEs,13501
 verifiers/v1/serve/encoding.py,sha256=hBZFucAZK9riXOV3DaHcskq9zHTT8gVGkoFVOPfnr30,2499
 verifiers/v1/serve/pool.py,sha256=bT-FOlIuOiBtOasuPB-p9887tkpEqvQUJBgn1Q45ph0,15912
 verifiers/v1/serve/server.py,sha256=i6XrY_PMAapkrP54tYX8-K6iGmOmrR7BbQ-3Qhn3w4s,9885
@@ -185,13 +185,13 @@ verifiers/v1/utils/loaders.py,sha256=2TrgZK3MzKub1GnfaTVIQhlwF6q4qATXZCjicH6nT9c
 verifiers/v1/utils/logging.py,sha256=-zgaM-HZsgxumZ4wYNlRMZepp6JanCSjeL2wSVulwJM,2296
 verifiers/v1/utils/memory.py,sha256=ZkIvGk6uITAH5sKon65LifKPbvZr8mJ__PVc23FZpOQ,1835
 verifiers/v1/utils/paths.py,sha256=it_4JWf_8bPZe4TRcyvL7_KJ4fsbEsxhhIY_xA8Ikus,308
-verifiers/v1/utils/platform.py,sha256=F4qMxWJY5MwtJ7ozASwBrpcU8u3lXvHIc5rg-rviDOQ,7181
+verifiers/v1/utils/platform.py,sha256=YwTHVA1Onn74lG6S3sdHus9OlP-rAbwkk9Po7h-BjpE,7159
 verifiers/v1/utils/prime.py,sha256=UTYRjp9cbjNb6CVmBHda-1wWZAIIfxNOmTyNuT7_wL4,970
 verifiers/v1/utils/retries.py,sha256=Y2ZgrAjn-qNkRKZP_RVNL_7EK0iaRcZeCa404lpGYi4,5417
 verifiers/v1/utils/score.py,sha256=493yJVMw8teCu9JxapxMFPFzI0hNUqdo0Y2nGW4kckk,6200
 verifiers/v1/utils/version.py,sha256=-obEo_-l9-D8FLef4hYxncOe-uJpxrM1g2Hig_37Sgs,1607
-verifiers-0.3.2.dev102.dist-info/METADATA,sha256=bgoqfsVCE7mh66RzwnfvlX0Jcxbi05QJO_JMjrSPUpI,4238
-verifiers-0.3.2.dev102.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
-verifiers-0.3.2.dev102.dist-info/entry_points.txt,sha256=iugElcdWPKbQM7uFF0lZ8iUpHsNr17-BwEAAjJWxV3U,259
-verifiers-0.3.2.dev102.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
-verifiers-0.3.2.dev102.dist-info/RECORD,,
+verifiers-0.3.2.dev105.dist-info/METADATA,sha256=VP57HoS62DnWGpDZUu3u0quWYfjoEmFTCFrwrGgPjfE,4238
+verifiers-0.3.2.dev105.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
+verifiers-0.3.2.dev105.dist-info/entry_points.txt,sha256=iugElcdWPKbQM7uFF0lZ8iUpHsNr17-BwEAAjJWxV3U,259
+verifiers-0.3.2.dev105.dist-info/licenses/LICENSE,sha256=v0RrUsdV3IDoZhrRce297IXS3xMHNJ-_LdLpFAUWb9k,1072
+verifiers-0.3.2.dev105.dist-info/RECORD,,