From 9a6f8e599204aa37481f5f37a1b20938aee98d5c Mon Sep 17 00:00:00 2001
From: "Wang, FangYuan" <39615225+At1a8@users.noreply.github.com>
Date: Tue, 7 Jul 2026 14:57:57 +0800
Subject: [PATCH] [AMD] Fix DeepSeek V4 MTP accuracy issue (#30333)

---
 .../srt/mem_cache/deepseek_v4_compress_state.py       | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/python/sglang/srt/mem_cache/deepseek_v4_compress_state.py b/python/sglang/srt/mem_cache/deepseek_v4_compress_state.py
index 735eb0430..dff122578 100644
--- a/python/sglang/srt/mem_cache/deepseek_v4_compress_state.py
+++ b/python/sglang/srt/mem_cache/deepseek_v4_compress_state.py
@@ -129,7 +129,16 @@ class CompressStatePool:
             dtype=dtype, device=device, enable_memory_saver=enable_memory_saver
         )
         if not online:
-            self.kv_score_buffer[-1].clear()
+            if _is_hip and ratio == 128:
+                # Request-scoped C128 state is addressed by req_pool_idx (or a
+                # per-request ring).  The pool is allocated with torch.empty(),
+                # so a cold server can otherwise read uninitialized partial
+                # states before a request slot has been written for the first
+                # time.  Initialize all C128 rows to the empty-state sentinel;
+                # C4 keeps the historical last-row sentinel behavior.
+                self.kv_score_buffer.clear()
+            else:
+                self.kv_score_buffer[-1].clear()
 
     def _alloc_kv_score_buffer(
         self, *, dtype: torch.dtype, device: str, enable_memory_saver: bool