[https://nvbugs/6248769][fix] In KVCacheManager.__init__, after resolving max_attention_window_vec and…

tensorrt_llm/_torch/pyexecutor/resource_manager.py

@@ -7,7 +7,7 @@
 import os
 from abc import ABC, abstractmethod
 from collections import OrderedDict, defaultdict, deque
-from dataclasses import dataclass
+from dataclasses import dataclass, replace
 from typing import (TYPE_CHECKING, Dict, Iterable, List, NamedTuple, Optional,
                     Sequence, Set, Tuple, Union)
 
@@ -680,6 +680,18 @@ def append_to_kv_heads_per_layer(num_kv_heads_per_layer: List[int],
             layer_mask=layer_mask,
         )
 
+        # _resolve_max_attention_window_vec clamps each window to max_seq_len (a
+        # sliding window wider than the sequence never slides, so it degenerates
+        # to full attention).  Mirror that clamp onto the per-pool window_size so
+        # pool keys stay consistent with max_attention_window_vec; otherwise
+        # _build_layer_to_pool_idx can't map a layer whose window was clamped
+        # down (e.g. Ministral sliding_window=32768 with max_seq_len=512).
+        if self.pool_configurations:
+            self.pool_configurations = [
+                replace(pc, window_size=min(pc.window_size, self.max_seq_len))
+                for pc in self.pool_configurations
+            ]
+
         # Now that max_attention_window_vec is known, build layer -> pool_idx
         # from the (pre-clamp) pool_configurations.  Stays valid through the
         # window clamping below because that only rewrites per-pool

tests/integration/test_lists/waives.txt

@@ -14,7 +14,6 @@ accuracy/test_llm_api_autodeploy.py::TestGemma4MoE::test_bf16 SKIP (https://nvbu
 accuracy/test_llm_api_autodeploy.py::TestGemmaE2B::test_gemma4_e2b_it SKIP (https://nvbugs/6194934)
 accuracy/test_llm_api_autodeploy.py::TestMiniMaxM2::test_finegrained_fp8 SKIP (https://nvbugs/6158397)
 accuracy/test_llm_api_autodeploy.py::TestModelRegistryAccuracy::test_autodeploy_from_registry[google_gemma-3-1b-it-False] SKIP (https://nvbugs/6248764)
-accuracy/test_llm_api_autodeploy.py::TestModelRegistryAccuracy::test_autodeploy_from_registry[mistralai_Ministral-8B-Instruct-2410-False] SKIP (https://nvbugs/6248769)
 accuracy/test_llm_api_autodeploy.py::TestModelRegistryAccuracy::test_autodeploy_from_registry[nvidia_Llama-3.1-8B-Instruct-NVFP4-True] SKIP (https://nvbugs/6245279)
 accuracy/test_llm_api_autodeploy.py::TestNemotronNanoV3::test_accuracy[nvfp4-1-trtllm] SKIP (https://nvbugs/6200112)
 accuracy/test_llm_api_autodeploy.py::TestNemotronUltraV3::test_accuracy[nvfp4-8] SKIP (https://nvbugs/6248757)