[None][fix] Allow MiniMax-M3's to use separate draft KV cache in disagg

tensorrt_llm/_torch/pyexecutor/py_executor_creator.py

@@ -514,8 +514,23 @@ def create_py_executor(
         # WAR for https://nvbugs/5807902
         # Disable separate draft KV cache in disaggregated mode
         # Enable separate pool for None DI + Non-KVBM and Aggregated + KVBM
+        #
+        # MiniMax-M3 is exempt: its cache manager forbids sharing draft
+        # layers (supports_shared_draft_layers=False), so this WAR's
+        # shared-manager fallback does not exist for it. Both worker roles
+        # must keep the separate manager, or their target pool layouts
+        # diverge and disaggregated KV transfer breaks. Checked via the
+        # sparse-attention algorithm because the manager class is not yet
+        # resolved here ("minimax_m3" maps 1:1 to MiniMaxM3KVCacheManagerV2).
         if cache_transceiver_config is not None:
-            spec_config._allow_separate_draft_kv_cache = False
+            if is_minimax_m3(m3_sparse_config):
+                logger.warning(
+                    "Disaggregated MiniMax-M3 keeps the separate draft KV "
+                    "cache manager; draft-layer KV is not transferred, so "
+                    "generation-side acceptance is reduced until the drafter "
+                    "rebuilds its context window.")
+            else:
+                spec_config._allow_separate_draft_kv_cache = False
 
     # chunk_unit_size may be changed to 64 when using flash mla
     attn_runtime_features = AttentionRuntimeFeatures(