server: learn a past example: /learn's messages form serves its prompt through the ordinary completion path first, then steps

tools/server/server-context.cpp

@@ -2429,6 +2429,18 @@ struct server_context_impl {
     }
 
     void send_final_response(server_slot & slot) {
+        if (slot.task->params.learn_only) {
+            // a /learn that served its prompt first: the learn outcome is its whole reply
+            auto res = std::make_unique<server_task_result_learn>();
+            res->id = slot.task->id;
+            if (slot.learned) {
+                res->outcome = *slot.learned;
+            } else {
+                res->outcome.failure = "the served prompt finished without its learn step";
+            }
+            queue_results.send(std::move(res));
+            return;
+        }
         auto res = std::make_unique<server_task_result_cmpl_final>();
 
         res->id      = slot.task->id;
@@ -6450,6 +6462,28 @@ void server_routes::init_routes() {
         task.learn.engine    = json_value(body, "engine", std::string());
         task.learn.input     = json_value(body, "input", std::string());
         if (body.contains("messages")) {
+            // A past example is served on HER slot under HER stable set (ONE-RESIDENT §61, §71, §227):
+            // the slot picker would otherwise take another mind's idle slot and evict her cached prefix,
+            // and the server-wide default set would learn a correction against the wrong model (Fable
+            // on fork #89). Her slot is required; her set is required whenever the server holds any
+            // non-zero adapter besides the shadow being trained.
+            if (task.learn.id_slot < 0) {
+                res->error(format_error_response("a messages-form learn names her slot (\"id_slot\"): the slot whose cached prompt it serves",
+                                                 ERROR_TYPE_INVALID_REQUEST));
+                return res;
+            }
+            if (!body.contains("lora")) {
+                // the set the main loop published (this is an HTTP thread: never params_base itself)
+                for (const auto & adapter : ctx_server.lora_report_snapshot()) {
+                    const int32_t id = adapter.value("id", -1);
+                    if (id != task.learn.lora_id && adapter.value("scale", 0.0f) != 0.0f) {
+                        res->error(format_error_response("a messages-form learn names her stable adapter set (\"lora\", as a completion "
+                                                         "sends it): adapter " + std::to_string(id) + " serves at a non-zero scale here",
+                                                         ERROR_TYPE_INVALID_REQUEST));
+                        return res;
+                    }
+                }
+            }
             // the conversation she was served, its last message her reply (or the correction). The
             // served prompt is every earlier message through the SAME parser and template arguments
             // /v1/chat/completions used (chat_template_kwargs included: a thinking template renders a
@@ -6514,6 +6548,36 @@ void server_routes::init_routes() {
                 return res;
             }
             task.learn.reply.assign(tokens.begin() + prefix, tokens.end());
+            // A past example (S4 (f)): the slot may no longer hold this prompt. Serve it first
+            // through the ordinary completion path (prompt cache, checkpoints, the slot's own
+            // admission), one token, and learn as it finishes: when the slot still holds it, the
+            // cache makes that free; when it does not, it is one prefill. One prefill path, one
+            // learn path. The slot is the caller's choice: whose cached prompt it replaces.
+            server_task serve(SERVER_TASK_TYPE_COMPLETION);
+            serve.tokens = server_tokens(task.learn.prompt, false);
+            json serve_body = { { "n_predict", 1 }, { "temperature", 0.0 }, { "cache_prompt", true } };
+            if (body.contains("lora")) {
+                serve_body["lora"] = body.at("lora"); // her stable set: the prompt is served, and the shadow learns, beside it
+            }
+            serve.params = server_schema::eval_llama_cmpl_schema(ctx_server.vocab, params, meta->logit_bias_eog, serve_body);
+            serve.params.learn      = task.learn;
+            serve.params.learn_only = true;
+            serve.params.res_type   = TASK_RESPONSE_TYPE_NONE;
+            serve.id_slot           = task.learn.id_slot;
+            auto & rd = res->rd;
+            serve.id = rd.get_new_id();
+            rd.post_task(std::move(serve));
+            auto result = rd.next(req.should_stop);
+            if (!result) {
+                GGML_ASSERT(req.should_stop());
+                return res;
+            }
+            if (result->is_error()) {
+                res->error(result->to_json());
+                return res;
+            }
+            res->ok(result->to_json());
+            return res;
         } else if (body.contains("reply_tokens")) {
             task.learn.reply = body.at("reply_tokens").get<std::vector<llama_token>>();
         } else {

tools/server/server-task.h

@@ -183,6 +183,8 @@ struct task_params {
     // learn a reply on this request's slot when it finishes, before the slot is released (no other
     // request can take the slot in between): POST /learn's work, atomic with the turn
     std::optional<learn_request> learn;
+    // this completion exists only to serve its prompt for `learn`: its reply is the learn outcome
+    bool learn_only = false;
 };
 
 // struct for tracking the state of a task (e.g., for streaming)

tools/server/tests/learn-every-format.py

@@ -116,6 +116,66 @@ def landed(lora, outcome):
                       "next": nxt["step"], "stale_generation": stale_gen["failure"], "resumed": resumed["step"]}))
     sys.exit(0)
 
+# A past example (S4 (f); needs --jinja): learn a conversation the slot no longer holds. The engine
+# serves its prompt first through the ordinary completion path, then steps; a second learn of a
+# conversation the slot now holds reuses the cached prompt.
+if "--past-example" in sys.argv:
+    raw("/completion", {"prompt": "Something unrelated entirely: the weather in Lisbon", "id_slot": 0, "n_predict": 1, "temperature": 0})
+    def lesson(lora, messages):
+        step = fenced(lora)
+        return {"messages": messages, "id_slot": 0, "lora_id": lora, "lr": step["lr"], "execution": step["execution"],
+                "base_step": step["base_step"], "engine": ENGINE, "input": step["input"],
+                "chat_template_kwargs": {"enable_thinking": False}}
+    past = [{"role": "user", "content": "Which city is the capital of Kansas?"},
+            {"role": "assistant", "content": "Topeka is the capital of Kansas."}]
+    first = landed(sid, json.loads(raw("/learn", lesson(sid, past))))
+    assert first["success"] and first["n_labels"] > 1 and first["step"] == 1, first
+    again = landed(sid, json.loads(raw("/learn", lesson(sid, past))))
+    assert again["success"] and again["step"] == 2 and again["loss"] < first["loss"], (first, again)
+    print(json.dumps({"past_example": first, "again_cached": again}))
+    sys.exit(0)
+
+# A past example is learned under HER stable set, on HER slot (needs shadow-05.gguf and fresh copies
+# shadow-06.gguf, shadow-07.gguf in --train-dir). What this catches (Fable on fork #89): the serve
+# task carried no adapter set, so a mind with a stable gene learned her past example against the
+# server-wide default (the wrong model) and on whatever idle slot the picker chose.
+if "--past-example-stable-set" in sys.argv:
+    import urllib.error
+    # her gene: a shadow trained two steps and snapshotted (a fresh shadow is an identity adapter)
+    turn = {"prompt": "The capital of Kansas is", "id_slot": 0, "n_predict": 1, "temperature": 0}
+    for _ in range(2):
+        assert landed(sid, json.loads(raw("/completion", dict(turn, learn=fenced(sid, lr=0.05))))["learn"])["success"]
+    assert json.loads(raw("/lora-adapters/snapshot", {"id": sid, "name": "gene-test.gguf"}))["success"]
+    gene = json.loads(raw("/lora-adapters/load", {"name": "gene-test.gguf"}))["id"]
+    with_gene, without_gene = (json.loads(raw("/lora-adapters/load", {"name": n}))["id"] for n in ("shadow-06.gguf", "shadow-07.gguf"))
+    past = [{"role": "user", "content": "Which city is the capital of Kansas?"},
+            {"role": "assistant", "content": "Topeka is the capital of Kansas."}]
+    def lesson(shadow, **extra):
+        step = fenced(shadow)
+        body = {"messages": past, "id_slot": 0, "lora_id": shadow, "lr": step["lr"], "execution": step["execution"],
+                "base_step": step["base_step"], "engine": ENGINE, "input": step["input"],
+                "chat_template_kwargs": {"enable_thinking": False}}
+        body.update(extra)
+        return body
+    def refused(body, needle):
+        try:
+            raw("/learn", body)
+        except urllib.error.HTTPError as error:
+            detail = error.read().decode()
+            assert error.code == 400 and needle in detail, (error.code, detail)
+            return detail
+        raise AssertionError("not refused: " + needle)
+    refused(lesson(with_gene, id_slot=-1), "id_slot")
+    under = landed(with_gene, json.loads(raw("/learn", lesson(with_gene, lora=[{"id": gene, "scale": 1.0}]))))
+    plain = landed(without_gene, json.loads(raw("/learn", lesson(without_gene))))  # the gene serves at scale 0 here
+    assert under["success"] and plain["success"], (under, plain)
+    assert abs(under["loss"] - plain["loss"]) > 1e-4, ("her gene never reached the serve", under, plain)
+    # once the gene serves server-wide, a learn that does not name her set is refused, never guessed
+    raw("/lora-adapters", [{"id": gene, "scale": 1.0}])
+    why = refused(lesson(without_gene), "names her stable adapter set")
+    print(json.dumps({"under_her_gene": under["loss"], "without": plain["loss"], "refused": why[:120]}))
+    sys.exit(0)
+
 # Two minds on one serving context: each shadow keeps its own optimizer while the other learns in
 # between (needs shadow-05.gguf and a copy of it, shadow-06.gguf, in --train-dir).
 if "--two-minds" in sys.argv: