shared/tree_placeholder_guard_byteshark/tree_placeholder_guard.patch

diff --git a/tree_v2_ext.py b/tree_v2_ext.py
index 0000000..0000000 100644
--- a/tree_v2_ext.py
+++ b/tree_v2_ext.py
@@
 TREE_STATS_EVERY = int(os.environ.get("TREE_STATS_EVERY", "2048"))
+TREE_FAIL_ON_PLACEHOLDER_SPEC = (
+    os.environ.get("TREE_FAIL_ON_PLACEHOLDER_SPEC", "1") == "1"
+)
@@
         spec = scheduler_output.scheduled_spec_decode_tokens
         tree = False
         if SPEC_TREE_SPEC and len(spec) == 1:
             (toks,) = spec.values()
             ksch = len(toks)
+            if (
+                TREE_FAIL_ON_PLACEHOLDER_SPEC
+                and ksch
+                and any(int(tok) < 0 for tok in toks)
+            ):
+                _fail(
+                    "scheduler handed negative tree draft tokens before "
+                    f"GPU verify: len={ksch} head={list(toks)[:8]}"
+                )
             if ksch and ksch % TREE_W == 0:
                 K = ksch // TREE_W
                 R = 1 + ksch
diff --git a/sitecustomize.py b/sitecustomize.py
index 0000000..0000000 100644
--- a/sitecustomize.py
+++ b/sitecustomize.py
@@
 def _apply_samplerprep_proposer_patch(module: Any) -> None:
     proposer_cls = module.SpecDecodeBaseProposer
     original_prepare_next_token_ids_padded = proposer_cls.prepare_next_token_ids_padded
+    cache_stats = {"hit": 0, "miss_placeholder": 0}
@@
         if (
             FUSED_GREEDY_REJECTION_PREP
             and cached is not None
             and sampled_token_ids.shape[0] == gpu_input_batch.num_reqs
             and gpu_input_batch.num_reqs == 1
         ):
+            cache_stats["hit"] += 1
             return cached
+        if SPEC_TREE_SPEC and FUSED_GREEDY_REJECTION_PREP:
+            try:
+                has_placeholder = bool((sampled_token_ids < 0).any().item())
+            except Exception:
+                has_placeholder = False
+            if has_placeholder:
+                cache_stats["miss_placeholder"] += 1
+                raise RuntimeError(
+                    "tree-v2 placeholder row reached upstream "
+                    "prepare_next_token_ids_padded without a samplerprep cache "
+                    f"hit; shape={tuple(sampled_token_ids.shape)} "
+                    f"cache_stats={cache_stats}"
+                )
         return original_prepare_next_token_ids_padded(
             self,
             sampled_token_ids,