MUSA-0203: in-place query_start_loc update for captured loop replay

Add PR vllm-project#34880's matching in-place hunk for the captured replay path.
Without it the propose() loop reassigns:
  common_attn_metadata.query_start_loc = self.arange[: batch_size + 1]
which creates a fresh tensor each call. The captured FA kernel baked
in the pre-capture pointer at boot time and reads stale data at replay
-> draft tokens past position 0 collapse (Pos 1-4 fell to 7/3/1/0.6%
on yeahdongcn70 cookbook).

PR vllm-project#34880 changes to:
  common_attn_metadata.query_start_loc[: batch_size + 1] = self.arange[...]
which is an in-place write into the persistent buffer the captured
graph already references. Same fix for query_start_loc_cpu.

Test: reboot with VLLM_MUSA_DRAFT_FULL_WRAP=1 and rerun cookbook bench.
Expected: accept rate recovers, Pos 1-4 distribution matches eager.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: Xiaodong Ye

vllm_musa/patches/vllm__v1__spec_decode__llm_base_proposer.patch.py

@@ -192,6 +192,27 @@
                 batch_descriptor=batch_desc,
                 slot_mapping=self._get_slot_mapping(input_batch_size),"""
 
+# ---- Hunk 9b: propose() loop — in-place query_start_loc/query_start_loc_cpu ----
+# PR #34880 changes reassignment to in-place write so the captured loop graph's
+# data_ptr stays valid at replay. Reassignment created a fresh tensor each call
+# and the captured FA kernel read from the stale pre-capture pointer, which is
+# the root cause of Position 1-4 accept-rate collapse on Step 4.
+_OLD_QSL_REASSIGN = """        common_attn_metadata.query_start_loc = self.arange[: batch_size + 1]
+        common_attn_metadata.query_start_loc_cpu = torch.from_numpy(
+            self.token_arange_np[: batch_size + 1]
+        ).clone()"""
+
+_NEW_QSL_REASSIGN = """        # MUSA-0203 / PR #34880: in-place write keeps the captured graph's
+        # data_ptr valid across replays (the previous reassignment created a
+        # fresh tensor each call, stranding the captured FA kernel on a stale
+        # pointer and collapsing accept rate at positions > 0).
+        common_attn_metadata.query_start_loc[: batch_size + 1] = self.arange[
+            : batch_size + 1
+        ]
+        common_attn_metadata.query_start_loc_cpu[: batch_size + 1] = torch.from_numpy(
+            self.token_arange_np[: batch_size + 1]
+        ).clone()"""
+
 # ---- Hunk 11: full dummy_run rewrite — accept common_attn_metadata and
 # capture FULL-mode entries during boot. Without this, CUDAGraphWrapper
 # tries to capture at request time and trips validate_cudagraph_capturing_enabled.
@@ -360,6 +381,7 @@
     (_OLD_FFC_FIRST, _NEW_FFC_FIRST),
     (_OLD_PROPOSE_LOOP_DET, _NEW_PROPOSE_LOOP_DET),
     (_OLD_FFC_LOOP, _NEW_FFC_LOOP),
+    (_OLD_QSL_REASSIGN, _NEW_QSL_REASSIGN),
     (_OLD_DUMMY_RUN, _NEW_DUMMY_RUN),
     (_OLD_ISINSTANCE, _NEW_ISINSTANCE),
 ]