omnicache 0.1.0__tar.gz

omnicache-0.1.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

omnicache-0.1.0/MANIFEST.in

@@ -0,0 +1,6 @@
+include README.md LICENSE pyproject.toml
+recursive-include omnicache/configs *.json
+prune omnicache/server
+prune omnicache/tests
+prune omnicache/server_tests
+global-exclude __pycache__ *.pyc *.sh *.proto.* .venv

omnicache-0.1.0/PKG-INFO

@@ -0,0 +1,102 @@
+Metadata-Version: 2.4
+Name: omnicache
+Version: 0.1.0
+Summary: A unified-representation KV cache for vLLM: one model's KV cache becomes reusable by another model
+Author: omnicache contributors
+License: Apache-2.0
+Project-URL: Codec weights, https://www.modelscope.ai/models/Zachary777/omnicache-qwen3-codec
+Keywords: vllm,kv-cache,kv-connector,llm,inference
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Operating System :: POSIX :: Linux
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: vllm==0.15.1
+Requires-Dist: torch>=2.4
+Requires-Dist: numpy
+Requires-Dist: safetensors>=0.4
+Requires-Dist: transformers>=4.51
+Requires-Dist: huggingface_hub
+Requires-Dist: modelscope
+Provides-Extra: redis
+Requires-Dist: redis>=5; extra == "redis"
+Provides-Extra: test
+Requires-Dist: pytest>=8; extra == "test"
+Dynamic: license-file
+
+# omnicache
+
+A unified-representation KV cache for vLLM. One model reads a passage; another model answers questions about it
+without reading it again.
+
+The sender's vLLM instance encodes the KV cache of the passage into a shared representation `U` (1,024 dims per
+token, 2 KB/token in bf16) and writes it to a store. The receiver's vLLM instance, running a different model, decodes
+`U` into its own KV cache, writes it into its paged KV memory and prefills only the question.
+
+## Install
+
+```bash
+pip install omnicache            # pulls vllm==0.15.1
+```
+
+The base models are the official Qwen3 checkpoints. The codec weights (one encoder/decoder pair per model) are
+downloaded on first start from ModelScope
+([Zachary777/omnicache-qwen3-codec](https://www.modelscope.ai/models/Zachary777/omnicache-qwen3-codec)); each server
+fetches only its own model's part.
+
+## Start two models
+
+```bash
+KTC='{"kv_connector":"OmniCacheConnector","kv_role":"kv_both",
+      "kv_connector_extra_config":{"store":"/shared/omnicache_store"}}'
+ARGS="--attention-backend FLASH_ATTN --max-model-len 4096 --enforce-eager --no-async-scheduling"
+
+CUDA_VISIBLE_DEVICES=0 vllm serve Qwen/Qwen3-8B  --port 8001 $ARGS --kv-transfer-config "$KTC"
+CUDA_VISIBLE_DEVICES=1 vllm serve Qwen/Qwen3-14B --port 8002 $ARGS --kv-transfer-config "$KTC"
+```
+
+Both servers must point at the same `store`, a directory both can reach. `kv_both` lets each server send and
+receive. Optional keys: `"codec"` (a local bundle directory, a Hugging Face repo `owner/name`, or
+`modelscope://owner/name` / `modelscope.ai://owner/name`).
+
+## Talk
+
+```python
+from omnicache import Agent
+
+a = Agent("http://127.0.0.1:8001")   # Qwen3-8B
+b = Agent("http://127.0.0.1:8002")   # Qwen3-14B
+
+msg = a.send(open("contract.txt").read())                  # A prefills the passage and publishes U
+reply = b.ask(msg, "Who is liable if delivery is late?")   # B decodes U, prefills only the question
+print(reply.text)
+
+back = b.send(long_report)                                 # the other direction works the same way
+print(a.ask(back, "Summarise the report.").text)
+```
+
+`msg` holds only a key and the passage token ids (JSON-serialisable with `msg.to_json()`); `U` goes from server to
+server through the store and never passes through the client. `ask()` raises `OmniCacheMiss` when the receiver did
+not load `U`, so an answer from a native re-read is never reported as a hand-off.
+
+## How a message is matched
+
+The store key is `sha256` of the passage token ids, their count and the codec version. The sender computes it when it
+encodes; the receiver recomputes it from the first `T` tokens of its prompt and loads `U` only if the two agree. The
+same passage published twice is encoded once. Each sender's `U` is kept separately, and a receiver reuses the KV it
+decoded from a given sender's `U` for repeated questions on the same passage (vLLM prefix cache, isolated per
+passage and sender with `cache_salt`).
+
+## Scope of this release
+
+- Models: Qwen3-4B-Instruct-2507, Qwen3-8B, Qwen3-14B, any direction (they share one tokenizer).
+- Passage at the very start of the receiver prompt, bare text, at most 2,048 tokens together with the question.
+- Question format `"\n\nQuestion: ...\nAnswer:"`, completions API, non-streaming.
+- Single GPU per model (no tensor parallelism), FlashAttention backend, bf16 KV cache.
+- Store: a shared directory (a Redis backend is experimental).
+
+## License
+
+Apache-2.0

omnicache-0.1.0/README.md

@@ -0,0 +1,74 @@
+# omnicache
+
+A unified-representation KV cache for vLLM. One model reads a passage; another model answers questions about it
+without reading it again.
+
+The sender's vLLM instance encodes the KV cache of the passage into a shared representation `U` (1,024 dims per
+token, 2 KB/token in bf16) and writes it to a store. The receiver's vLLM instance, running a different model, decodes
+`U` into its own KV cache, writes it into its paged KV memory and prefills only the question.
+
+## Install
+
+```bash
+pip install omnicache            # pulls vllm==0.15.1
+```
+
+The base models are the official Qwen3 checkpoints. The codec weights (one encoder/decoder pair per model) are
+downloaded on first start from ModelScope
+([Zachary777/omnicache-qwen3-codec](https://www.modelscope.ai/models/Zachary777/omnicache-qwen3-codec)); each server
+fetches only its own model's part.
+
+## Start two models
+
+```bash
+KTC='{"kv_connector":"OmniCacheConnector","kv_role":"kv_both",
+      "kv_connector_extra_config":{"store":"/shared/omnicache_store"}}'
+ARGS="--attention-backend FLASH_ATTN --max-model-len 4096 --enforce-eager --no-async-scheduling"
+
+CUDA_VISIBLE_DEVICES=0 vllm serve Qwen/Qwen3-8B  --port 8001 $ARGS --kv-transfer-config "$KTC"
+CUDA_VISIBLE_DEVICES=1 vllm serve Qwen/Qwen3-14B --port 8002 $ARGS --kv-transfer-config "$KTC"
+```
+
+Both servers must point at the same `store`, a directory both can reach. `kv_both` lets each server send and
+receive. Optional keys: `"codec"` (a local bundle directory, a Hugging Face repo `owner/name`, or
+`modelscope://owner/name` / `modelscope.ai://owner/name`).
+
+## Talk
+
+```python
+from omnicache import Agent
+
+a = Agent("http://127.0.0.1:8001")   # Qwen3-8B
+b = Agent("http://127.0.0.1:8002")   # Qwen3-14B
+
+msg = a.send(open("contract.txt").read())                  # A prefills the passage and publishes U
+reply = b.ask(msg, "Who is liable if delivery is late?")   # B decodes U, prefills only the question
+print(reply.text)
+
+back = b.send(long_report)                                 # the other direction works the same way
+print(a.ask(back, "Summarise the report.").text)
+```
+
+`msg` holds only a key and the passage token ids (JSON-serialisable with `msg.to_json()`); `U` goes from server to
+server through the store and never passes through the client. `ask()` raises `OmniCacheMiss` when the receiver did
+not load `U`, so an answer from a native re-read is never reported as a hand-off.
+
+## How a message is matched
+
+The store key is `sha256` of the passage token ids, their count and the codec version. The sender computes it when it
+encodes; the receiver recomputes it from the first `T` tokens of its prompt and loads `U` only if the two agree. The
+same passage published twice is encoded once. Each sender's `U` is kept separately, and a receiver reuses the KV it
+decoded from a given sender's `U` for repeated questions on the same passage (vLLM prefix cache, isolated per
+passage and sender with `cache_salt`).
+
+## Scope of this release
+
+- Models: Qwen3-4B-Instruct-2507, Qwen3-8B, Qwen3-14B, any direction (they share one tokenizer).
+- Passage at the very start of the receiver prompt, bare text, at most 2,048 tokens together with the question.
+- Question format `"\n\nQuestion: ...\nAnswer:"`, completions API, non-streaming.
+- Single GPU per model (no tensor parallelism), FlashAttention backend, bf16 KV cache.
+- Store: a shared directory (a Redis backend is experimental).
+
+## License
+
+Apache-2.0

omnicache-0.1.0/omnicache/__init__.py

@@ -0,0 +1,27 @@
+"""OmniCache: cross-model latent hand-off for vLLM.
+
+A sender vLLM instance encodes the KV of a passage into the shared latent U (1,024 dims per token); a receiver
+instance of a different model decodes U into its own KV and skips the passage prefill.
+
+    from omnicache import Agent
+    a, b = Agent("http://127.0.0.1:8001"), Agent("http://127.0.0.1:8002")
+    msg = a.send(doc)
+    reply = b.ask(msg, "Who is the supplier?")
+
+The connector itself (omnicache.connector.OmniCacheConnector) is registered with vLLM by the
+`vllm.general_plugins` entry point (omnicache.plugin:register).
+"""
+__version__ = "0.1.0"
+
+_CLIENT = ("Agent", "Member", "Message", "Reply", "OmniCacheMiss")
+
+
+def __getattr__(name):  # lazy: vLLM imports this package in every process via the plugin entry point
+    if name in _CLIENT:
+        from .client import agent
+
+        return agent.Member if name == "Agent" else getattr(agent, name)
+    raise AttributeError(name)
+
+
+__all__ = list(_CLIENT)

omnicache-0.1.0/omnicache/client/agent.py

@@ -0,0 +1,179 @@
+"""Application-level API: one `Member` per vLLM endpoint, messages flow between any two members.
+
+    a = Member("http://host:8001", "qwen3_8b", ...)
+    b = Member("http://host:8002", "qwen3_14b", ...)
+    a.chat("hello")                         # ordinary generation, OmniCache not involved
+    msg = a.send(doc)                       # A prefills doc, encodes U, stores it; returns a small handle
+    reply = b.ask(msg, "Who is the supplier?")   # B decodes U into its own KV, reads only the question
+    back = b.send(long_report); a.ask(back, "...")   # reverse direction, same two calls
+
+Constraints (the codec's training distribution, enforced here):
+  * the passage is the very beginning of the receiver prompt, as bare text (no chat template before it);
+  * all members share the public token grid (Qwen3 tokenizer); cross-tokenizer members are out of MVP scope;
+  * ask() fails loudly (OmniCacheMiss) unless the receiver echoes load == "loaded", so a silent native read is never
+    reported as a latent hand-off.
+"""
+from __future__ import annotations
+
+import json
+import os
+import time
+from dataclasses import asdict, dataclass, field
+from typing import Any, Dict, List, Optional
+
+from .grid import Grid, cache_salt_for, key_for, question_suffix
+from .http import _completion, _post
+
+
+class OmniCacheMiss(RuntimeError):
+    """The receiver did not load the U (miss, bad salt, wrong T, store not visible, ...)."""
+
+
+@dataclass
+class Message:
+    """What travels between applications: a key into the U store plus the passage token ids (the receiver prompt
+    must start with exactly these ids). Small: ~4 bytes per token plus a header."""
+    id: str
+    sender: str
+    T: int
+    passage_ids: List[int]
+    codec_sha256: str
+    publish_s: float = 0.0
+    echo: Dict[str, Any] = field(default_factory=dict)
+
+    def to_json(self) -> str:
+        return json.dumps(asdict(self))
+
+    @classmethod
+    def from_json(cls, s: str) -> "Message":
+        return cls(**json.loads(s))
+
+
+@dataclass
+class Reply:
+    text: str
+    loaded: bool            # the passage KV came from the sender's U (fresh decode or reuse of an earlier decode)
+    latency_s: float
+    echo: Dict[str, Any]
+    raw: Dict[str, Any] = field(repr=False, default_factory=dict)
+
+
+DEFAULT_GRID_TOKENIZER = "Qwen/Qwen3-8B"     # public grid; all Qwen3 members tokenise identically
+
+
+def _served_model(url: str, api_key: Optional[str], timeout: float) -> Dict[str, Any]:
+    import urllib.request
+
+    req = urllib.request.Request(url + "/v1/models", headers={"Authorization": f"Bearer {api_key}"} if api_key else {})
+    with urllib.request.urlopen(req, timeout=timeout) as r:
+        data = json.loads(r.read().decode("utf-8"))["data"]
+    return data[0]
+
+
+class Member:
+    """One vLLM endpoint running the OmniCache connector.
+
+    Only `url` is required: the served model id (and its local path, used as tokenizer when it exists) is read from
+    /v1/models; the member name and codec version of a sender come back in its publish echo.
+    """
+
+    def __init__(self, url: str, name: Optional[str] = None, tokenizer_path: Optional[str] = None,
+                 codec_sha256: Optional[str] = None, store=None, api_key: Optional[str] = None,
+                 timeout: float = 600.0, model_id: Optional[str] = None):
+        self.url = url.rstrip("/")
+        self.api_key = api_key
+        self.timeout = timeout
+        info: Dict[str, Any] = {}
+        if model_id is None or tokenizer_path is None:
+            info = _served_model(self.url, api_key, timeout)
+        self.model = model_id or info.get("id")
+        if tokenizer_path is None:
+            root = info.get("root")
+            tokenizer_path = root if root and os.path.isdir(root) else DEFAULT_GRID_TOKENIZER
+        self.grid = Grid(tokenizer_path)
+        self.name = name                        # member name; learnt from the first publish echo when None
+        self.codec_sha256 = codec_sha256        # learnt from the first echo when None
+        self.store = store                      # optional FileUStore: enables sha-pinned salts + visibility wait
+
+    # ---- ① ordinary chat, untouched by the connector -------------------------------------------------------
+    def chat(self, text: str, max_tokens: int = 256, **kw) -> str:
+        payload = {"model": self.model, "messages": [{"role": "user", "content": text}], "max_tokens": max_tokens,
+                   "temperature": kw.pop("temperature", 0.0), **kw}
+        r = _post(self.url + "/v1/chat/completions", payload, self.timeout, self.api_key)
+        return r["choices"][0]["message"]["content"]
+
+    # ---- ② publish a passage as U ---------------------------------------------------------------------------
+    def send(self, doc: str, wait_s: float = 10.0) -> Message:
+        ids = self.grid.passage_ids(doc)
+        T = len(ids)
+        payload = {"model": self.model, "prompt": ids, "max_tokens": 1, "temperature": 0.0,
+                   "kv_transfer_params": {"omnicache": {"op": "send", "T": T}}}
+        t0 = time.perf_counter()
+        r = _post(self.url + "/v1/completions", payload, self.timeout, self.api_key)
+        echo = (r.get("kv_transfer_params") or {}).get("omnicache", {})
+        if echo.get("store") not in ("written", "queued", "exists"):
+            raise RuntimeError(f"{self.name or self.url} did not publish U: {echo}")
+        self.name = self.name or echo.get("member")
+        self.codec_sha256 = self.codec_sha256 or echo.get("codec_sha256")
+        key = key_for(ids, T, self.codec_sha256)
+        if echo.get("u_key") and echo["u_key"] != key:
+            raise RuntimeError(f"server key {echo['u_key']} != client key {key} (tokenizer grid or codec mismatch)")
+        if self.store is not None:                           # make sure the U is visible before handing out the id
+            deadline = time.monotonic() + wait_s
+            while self.store.stat(key, self.name) is None:
+                if time.monotonic() > deadline:
+                    raise RuntimeError(f"U {key} from {self.name} not visible in the store after {wait_s}s")
+                time.sleep(0.05)
+        return Message(id=key, sender=self.name, T=T, passage_ids=ids, codec_sha256=self.codec_sha256,
+                       publish_s=time.perf_counter() - t0, echo=echo)
+
+    # ---- ③ answer a question about a received message ---------------------------------------------------------
+    def _salt(self, msg: Message) -> str:
+        u_sha = None
+        if self.store is not None:
+            ent = self.store.stat(msg.id, msg.sender)
+            u_sha = ent.u_sha256 if ent is not None else None
+        return cache_salt_for(msg.id, msg.sender, u_sha)
+
+    def ask(self, msg: Message, question: str, max_tokens: int = 64, first_line: bool = True, strict: bool = True,
+            lang: str = "en") -> Reply:
+        if self.codec_sha256 and msg.codec_sha256 != self.codec_sha256:
+            raise ValueError(f"message encoded with codec {msg.codec_sha256[:12]}, this member uses {self.codec_sha256[:12]}")
+        q_ids = self.grid.question_suffix_ids(question, lang)
+        payload: Dict[str, Any] = {"model": self.model, "prompt": list(msg.passage_ids) + q_ids, "max_tokens": max_tokens,
+                                   "temperature": 0.0,
+                                   "kv_transfer_params": {"omnicache": {"op": "recv", "T": msg.T, "sender": msg.sender, "u_key": msg.id}},
+                                   "cache_salt": self._salt(msg)}
+        if first_line:
+            payload["stop_token_ids"] = self.grid.newline_token_ids()
+        t0 = time.perf_counter()
+        r = _completion(self.url + "/v1/completions", payload, self.timeout, self.api_key, False)
+        dt = time.perf_counter() - t0
+        echo = (r.get("kv_transfer_params") or {}).get("omnicache", {})
+        # "loaded": U decoded into fresh blocks this request. "local_hit": the blocks were decoded from the same U by an
+        # earlier ask and reused through the prefix cache; they sit under the sender-qualified cache_salt, which only
+        # connector loads use, so they are U-derived, never a native read of the passage.
+        loaded = echo.get("load") in ("loaded", "local_hit") and int(echo.get("T") or 0) == msg.T
+        if strict and not loaded:
+            raise OmniCacheMiss(f"{self.name} did not load U {msg.id[:16]} from {msg.sender}: {echo}")
+        text = r["text"].split("\n", 1)[0].strip() if first_line else r["text"]
+        return Reply(text=text, loaded=loaded, latency_s=dt, echo=echo, raw=r)
+
+    # ---- references for evaluation ---------------------------------------------------------------------------
+    def read_and_ask(self, doc: str, question: str, max_tokens: int = 64, lang: str = "en") -> Reply:
+        """Receiver reads the passage itself (native upper bound)."""
+        p, q = self.grid.build(doc, question, lang)
+        payload = {"model": self.model, "prompt": p + q, "max_tokens": max_tokens, "temperature": 0.0,
+                   "stop_token_ids": self.grid.newline_token_ids()}
+        t0 = time.perf_counter()
+        r = _completion(self.url + "/v1/completions", payload, self.timeout, self.api_key, False)
+        return Reply(text=r["text"].split("\n", 1)[0].strip(), loaded=False, latency_s=time.perf_counter() - t0, echo={}, raw=r)
+
+    def ask_blind(self, question: str, max_tokens: int = 64, lang: str = "en") -> Reply:
+        """No context at all (floor)."""
+        q = self.grid.question_suffix_ids(question, lang)
+        payload = {"model": self.model, "prompt": q, "max_tokens": max_tokens, "temperature": 0.0,
+                   "stop_token_ids": self.grid.newline_token_ids()}
+        t0 = time.perf_counter()
+        r = _completion(self.url + "/v1/completions", payload, self.timeout, self.api_key, False)
+        return Reply(text=r["text"].split("\n", 1)[0].strip(), loaded=False, latency_s=time.perf_counter() - t0, echo={}, raw=r)