test(mlx-fused): refresh stale fused-loop test fakes and expectations

tests/backends/mlx/test_fused_specdecode.py

@@ -51,6 +51,12 @@ def append_token(self, token_id):
         self.appends.append(token_id)
         return self.next_token_logits
 
+    def last_aux_torch_slice(self, start=0, end=None):
+        # Mirror MLXRestoredIncrementalVerifier.last_aux_torch_slice: per-aux-layer
+        # torch rows of the most recent forward_block, sliced [start:end].
+        aux = self._last_aux or [torch.zeros(1, self.hidden)]
+        return [a[start:end] for a in aux]
+
 
 class _FakeDrafter:
     def __init__(self, drafts):
@@ -89,13 +95,16 @@ def test_fused_loop_full_acceptance():
     res = fsd.fused_specdecode_generate(
         adapter, drafter, gen_tokens=5, block_size=4, eos_ids=(),
         **_loop_kwargs(drafter))
-    # Block1: candidate=[100,101,102] all accepted (3) + correction 103.
-    # Block2: candidate=[104] accepted (1) + correction 105 -> truncated to 5.
+    # Block1: candidate=[100,101,102] fully accepted (3). On FULL acceptance the
+    # loop reuses block_logits[-1] (=103) as the next distribution and does NOT
+    # append a correction token. next=103.
+    # Block2: L=2 -> candidate=[103,200]; accept 103 (1), reject 200, correction
+    #   =104 appended -> commit [103,104]; total 5 tokens.
     assert res["tokens"] == [100, 101, 102, 103, 104]
     assert res["blocks"] == 2
     assert res["mean_accept_len"] == 2.0          # (3 + 1) / 2
     assert adapter.commits[0] == (3, 3)           # block1 verify-commit
-    assert adapter.appends == [103, 105]          # one correction per block
+    assert adapter.appends == [104]               # only block2's correction
     # capture flag toggled on during loop, off after.
     assert adapter._capture_aux is False
     # context K/V extended once per block.
@@ -122,9 +131,27 @@ def test_fused_loop_stops_on_eos():
     res = fsd.fused_specdecode_generate(
         adapter, drafter, gen_tokens=50, block_size=4, eos_ids=(103,),
         **_loop_kwargs(drafter))
-    # correction 103 is EOS -> stop after first block.
+    # Block1 fully accepts [100,101,102] (no correction appended on full accept),
+    # leaving next=103. Block2's bonus is then 103 (EOS), committed and stopped.
     assert res["tokens"] == [100, 101, 102, 103]
-    assert res["blocks"] == 1
+    assert res["blocks"] == 2
+
+
+def test_fused_loop_greedy_fallback_on_low_acceptance():
+    adapter = _FakeAdapter(prompt_len=5, first_token=100)
+    # Each block accepts only the bonus (drafts mismatch the verifier), so after
+    # 2 blocks mean acceptance = 1.0 < 1.5 and the loop switches to plain greedy
+    # to finish the budget (no aux capture, no drafter extension past the blocks).
+    drafter = _FakeDrafter(drafts=[[999, 999, 999], [999, 999, 999]])
+    res = fsd.fused_specdecode_generate(
+        adapter, drafter, gen_tokens=10, block_size=4, eos_ids=(),
+        **_loop_kwargs(drafter))
+    # blocks 1-2 commit [100,101] then [102,103]; greedy fallback adds 104..109.
+    assert res["tokens"] == [100, 101, 102, 103, 104, 105, 106, 107, 108, 109]
+    assert res["blocks"] == 2              # only the speculative blocks are counted
+    assert res["mean_accept_len"] == 1.0   # (1 + 1) / 2
+    assert adapter._capture_aux is False   # turned off for the greedy tail
+    assert drafter.extend_calls == 2       # extended only during the spec blocks
 
 
 # =========================================================================== #
@@ -273,8 +300,10 @@ def as_linear(self, h): return "L"
     adapter._capture_aux = True
     logits = adapter.forward_block([7, 8])
     assert logits == "ROW"                       # _Model returns row at [0]
-    # aux = [hs[1]] = [layer-0 output]; bridged after stripping batch ([0]).
-    assert adapter._last_aux == [("torch", ("row", 0))]
+    # aux = [hs[1]] = [layer-0 output], captured LAZILY in MX (_last_aux_mx);
+    # _last_aux stays None and the torch bridge happens on demand.
+    assert adapter._last_aux is None
+    assert adapter.last_aux_torch_slice() == [("torch", ("row", 0))]
 
     # commit_or_truncate trims by (forwarded - accepted) and advances _past_len
     adapter.commit_or_truncate(forwarded=2, accepted=1)