refactor: unify rollout onto inference engine KV-cache path

- RolloutGenerator now delegates prefill/decode to InferenceScheduler.run_batch (sync API, no background thread), sharing one KV-cache code path with the inference server and eliminating O(n^2) recompute in rollout
- Add sample(return_logprobs=) and Executor.execute_decode(return_logprobs=) to expose behaviour-policy log-probs through the engine; Task gains output_logprobs
- RolloutResult now subclasses RawRollout (adds rewards only), removing duplicated fields
- RolloutRunner.__call__ returns (result, is_fresh) instead of relying on object identity, removing the fragile refresh-detection contract
- Remove O(n^2) generate_responses helper and dead code (_tokenize_prompts, unused old_model arg)
- train_context.py wires InferenceScheduler directly instead of hand-rolling SamplingPipeline
- Tests: +11 covering return_logprobs, run_batch, and KV-cache-backed rollout semantics; 404 pass

tests/inference/test_sample.py

@@ -231,3 +231,54 @@ def test_sample_with_frequency_penalty():
     )
     assert tokens.shape == (1,)
     assert 0 <= tokens[0] < logits.size(-1)
+
+
+def test_sample_return_logprobs_shape():
+    """``return_logprobs=True`` returns ``[batch]`` logprobs aligned to tokens."""
+    logits = torch.tensor([[1.0, 2.0, 3.0], [3.0, 2.0, 1.0]])
+    out = sample(logits, temperature=1.0, return_logprobs=True)
+    tokens, logprobs = out
+    assert tokens.shape == (2,)
+    assert logprobs.shape == (2,)
+
+
+def test_sample_return_logprobs_nonpositive():
+    """Probabilities never exceed 1, so logprobs are always ≤ 0."""
+    torch.manual_seed(0)
+    logits = torch.randn(4, 50)
+    _, logprobs = sample(
+        logits, temperature=0.8, top_k=20, top_p=0.9, return_logprobs=True
+    )
+    assert torch.all(logprobs <= 1e-5)
+
+
+def test_sample_return_logprobs_greedy_path():
+    """Greedy decode (temperature 0) also returns logprobs."""
+    logits = torch.tensor([[1.0, 5.0, 2.0]])
+    tokens, logprobs = sample(logits, temperature=0.0, return_logprobs=True)
+    assert tokens[0].item() == 1
+    # log p(token=1) should equal log_softmax(logits)[1]
+    expected = torch.log_softmax(logits.float(), dim=-1)[0, 1]
+    assert torch.allclose(logprobs[0], expected, atol=1e-5)
+
+
+def test_sample_return_logprobs_matches_manual_computation():
+    """Returned logprob equals log_softmax(transformed_logits)[token]."""
+    torch.manual_seed(1)
+    logits = torch.randn(2, 30)
+    tokens, logprobs = sample(logits, temperature=0.7, top_p=0.95, return_logprobs=True)
+    # Recompute with the same pipeline
+    from astrai.inference.sample import (
+        SamplingPipeline,
+        TemperatureStrategy,
+        TopPStrategy,
+    )
+
+    pipeline = SamplingPipeline([TemperatureStrategy(0.7), TopPStrategy(0.95)])
+    transformed = pipeline.apply(logits.clone())
+    expected = torch.gather(
+        torch.log_softmax(transformed.float(), dim=-1),
+        -1,
+        tokens.unsqueeze(-1),
+    ).squeeze(-1)
+    assert torch.allclose(logprobs, expected, atol=1e-5)

tests/inference/test_scheduler.py

@@ -191,3 +191,124 @@ def test_prefill_skips_fully_cached_tasks(mock_model_and_tokenizer):
     task_id = scheduler.add_task("short prompt", stream_callback=lambda t: None)
     scheduler.stop()
     assert task_id.startswith("task_")
+
+
+def _make_real_scheduler(device):
+    """Build a scheduler backed by a tiny real model for run_batch tests."""
+    from astrai.config.model_config import AutoRegressiveLMConfig
+    from astrai.model.transformer import AutoRegressiveLM
+
+    class _Tok:
+        stop_ids = [2]
+
+        def encode(self, texts, **_):
+            if isinstance(texts, str):
+                texts = [texts]
+            return [[b for b in t.encode("utf-8")] for t in texts]
+
+        def decode(self, ids, skip_special_tokens=True):
+            return bytes(b for b in ids if b > 2 or not skip_special_tokens).decode(
+                "utf-8", errors="ignore"
+            )
+
+    cfg = AutoRegressiveLMConfig(
+        vocab_size=200,
+        dim=16,
+        n_heads=2,
+        n_kv_heads=1,
+        dim_ffn=32,
+        max_len=64,
+        n_layers=2,
+        norm_eps=1e-5,
+    )
+    model = AutoRegressiveLM(cfg).to(device=device).eval()
+    tokenizer = _Tok()
+    scheduler = InferenceScheduler(
+        model=model,
+        tokenizer=tokenizer,
+        max_batch_size=8,
+        max_seq_len=64,
+        max_prompt_len=64,
+    )
+    return scheduler, tokenizer, model
+
+
+def test_run_batch_returns_token_sequences():
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    scheduler, _tok, _model = _make_real_scheduler(device)
+    try:
+        prompts = [[10, 20, 30], [5, 6, 7, 8]]
+        results = scheduler.run_batch(prompts, max_tokens=4, temperature=1.0)
+        assert len(results) == 2
+        for ids in results:
+            assert isinstance(ids, list)
+            assert len(ids) <= 4
+            assert all(0 <= i < 200 for i in ids)
+    finally:
+        scheduler.stop()
+
+
+def test_run_batch_return_logprobs_aligned():
+    """return_logprobs=True gives (token_ids, logprobs) tuples with equal len."""
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    scheduler, _tok, _model = _make_real_scheduler(device)
+    try:
+        prompts = [[10, 20, 30, 40]]
+        results = scheduler.run_batch(
+            prompts, max_tokens=5, temperature=1.0, return_logprobs=True
+        )
+        assert len(results) == 1
+        token_ids, logprobs = results[0]
+        assert len(token_ids) == len(logprobs)
+        assert all(lp <= 1e-5 for lp in logprobs)  # logprobs ≤ 0
+    finally:
+        scheduler.stop()
+
+
+def test_run_batch_respects_max_tokens():
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    scheduler, _tok, _model = _make_real_scheduler(device)
+    try:
+        prompts = [[10, 20, 30]]
+        results = scheduler.run_batch(prompts, max_tokens=3, temperature=1.0)
+        assert len(results[0]) <= 3
+    finally:
+        scheduler.stop()
+
+
+def test_run_batch_stop_id_terminates():
+    """A token matching stop_ids terminates generation for that prompt."""
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    scheduler, _tok, _model = _make_real_scheduler(device)
+    try:
+        prompts = [[10, 20, 30]]
+        results = scheduler.run_batch(prompts, max_tokens=32, temperature=1.0)
+        # If stop token 2 was produced, it is the last token
+        if results[0] and results[0][-1] == 2:
+            # No tokens after stop should exist (since we terminate)
+            assert 2 not in results[0][:-1]
+    finally:
+        scheduler.stop()
+
+
+def test_run_batch_empty_prompts():
+    """Empty prompt list yields empty result list."""
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    scheduler, _tok, _model = _make_real_scheduler(device)
+    try:
+        assert scheduler.run_batch([], max_tokens=4) == []
+    finally:
+        scheduler.stop()
+
+
+def test_run_batch_too_long_prompt_skipped():
+    """A prompt longer than max_seq_len yields an empty result slot."""
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    scheduler, _tok, _model = _make_real_scheduler(device)
+    try:
+        long = list(range(100))  # > max_seq_len=64
+        results = scheduler.run_batch([long, [10, 20]], max_tokens=2)
+        assert results[0] == []
+        assert len(results[1]) <= 2
+    finally:
+        scheduler.stop()