From 047b7eba2130fe396efbdd0d274d22bd710db709 Mon Sep 17 00:00:00 2001
From: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
Date: Thu, 10 Sep 2026 11:46:52 -0700
Subject: [PATCH 1/9] [https://nvbugs/6713231][fix] Guard cross-KV prefix reuse
 for feature-driven encoder requests

Whisper requests carry encoder features, not encoder token ids, so the
request has no encoder unique tokens; with block reuse on (the default)
the C++ capacity scheduler dereferenced that empty optional and every
generation failed with "bad optional access". Skip the cross-reuse lookup
without encoder tokens, disable cross-pool reuse for such models, and
cover the default-reuse path with a Whisper integration test.

Signed-off-by: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
---
 .../batch_manager/capacityScheduler.cpp       | 24 ++++++++++++---
 tensorrt_llm/_torch/pyexecutor/_util.py       | 16 ++++++++++
 tensorrt_llm/inputs/registry.py               | 12 ++++++++
 .../llmapi/test_llm_api_pytorch_whisper.py    | 30 ++++++++++++++++++-
 .../test_lists/test-db/l0_h100.yml            |  1 +
 .../test_lists/test-db/l0_l40s.yml            |  1 +
 6 files changed, 79 insertions(+), 5 deletions(-)

diff --git a/cpp/tensorrt_llm/batch_manager/capacityScheduler.cpp b/cpp/tensorrt_llm/batch_manager/capacityScheduler.cpp
index 54f45259d3c8..518ba9e80334 100644
--- a/cpp/tensorrt_llm/batch_manager/capacityScheduler.cpp
+++ b/cpp/tensorrt_llm/batch_manager/capacityScheduler.cpp
@@ -37,6 +37,14 @@ using kv_cache_manager::BlockKeyHasher;
 namespace
 {
 
+/// Feature-driven encoders (e.g. Whisper) carry no encoder token ids, so the request has no
+/// encoder unique tokens to key cross-KV blocks on and cross prefix reuse cannot apply to it.
+bool hasEncoderUniqueTokens(LlmRequest const& req)
+{
+    auto const& encoderUniqueTokens = req.getEncoderUniqueTokens();
+    return encoderUniqueTokens.has_value() && encoderUniqueTokens.value() != nullptr;
+}
+
 std::tuple<std::unordered_set<BlockKey, BlockKeyHasher>, std::unordered_set<BlockKey, BlockKeyHasher>>
 prefillWithChunkedContextsAlreadyExecuting(RequestList const& activeRequests,
     kv_cache_manager::BaseKVCacheManager const& kvCacheManager,
@@ -59,7 +67,7 @@ prefillWithChunkedContextsAlreadyExecuting(RequestList const& activeRequests,
                     newlyContributedContextBlocks.insert(summary.firstNewBlock.value());
                 }
             }
-            if (crossKvCacheManager && crossKvCacheManager->isEnableBlockReuse())
+            if (crossKvCacheManager && crossKvCacheManager->isEnableBlockReuse() && hasEncoderUniqueTokens(*req))
             {
                 auto uniqueTokens = *(req->getEncoderUniqueTokens().value());
                 auto summary = crossKvCacheManager->analyzePrefixReuse(uniqueTokens, *req);
@@ -344,12 +352,20 @@ std::tuple<RequestVector, RequestVector> GuaranteedNoEvictScheduler::impl(
                         if (crossKvCacheManager && crossKvCacheManager->isEnableBlockReuse()
                             && !crossKvCacheManager->getBlockManager().isVariableWindow())
                         {
-                            auto uniqueTokens = *(req->getEncoderUniqueTokens().value());
-                            crossSummary = crossKvCacheManager->analyzePrefixReuse(uniqueTokens, *req);
+                            if (hasEncoderUniqueTokens(*req))
+                            {
+                                auto uniqueTokens = *(req->getEncoderUniqueTokens().value());
+                                crossSummary = crossKvCacheManager->analyzePrefixReuse(uniqueTokens, *req);
+                            }
+                            else
+                            {
+                                // Nothing to look up: an empty summary means "no reusable cross blocks".
+                                crossSummary = kv_cache_manager::PrefixReuseSummary{};
+                            }
                         }
                     }
                     else if (isEncoderInit && crossKvCacheManager && crossKvCacheManager->isEnableBlockReuse()
-                        && !crossKvCacheManager->getBlockManager().isVariableWindow())
+                        && !crossKvCacheManager->getBlockManager().isVariableWindow() && hasEncoderUniqueTokens(*req))
                     {
                         // Encoder admission only needs the cross summary for reuse ordering.
                         auto uniqueTokens = *(req->getEncoderUniqueTokens().value());
diff --git a/tensorrt_llm/_torch/pyexecutor/_util.py b/tensorrt_llm/_torch/pyexecutor/_util.py
index 50576b75e62b..7e417a292a09 100644
--- a/tensorrt_llm/_torch/pyexecutor/_util.py
+++ b/tensorrt_llm/_torch/pyexecutor/_util.py
@@ -25,6 +25,8 @@
                                  is_sm_100f, prefer_pinned,
                                  str_dtype_to_binding, torch_dtype_to_str)
 from tensorrt_llm.inputs.multimodal import MultimodalParams
+from tensorrt_llm.inputs.registry import \
+    input_processor_requires_encoder_features
 
 # isort: off
 from tensorrt_llm.llmapi.llm_args import (
@@ -1962,6 +1964,10 @@ def _split_kv_cache_budget_for_draft(
     def _is_encoder_decoder(self) -> bool:
         return self._model_engine.model.model_config.is_encoder_decoder
 
+    def _encoder_input_is_features(self) -> bool:
+        return input_processor_requires_encoder_features(
+            type(self._model_engine.model))
+
     @staticmethod
     def _get_config_int_attr(config, names: tuple[str, ...]) -> Optional[int]:
         for name in names:
@@ -2089,6 +2095,16 @@ def _split_kv_cache_budget_for_cross(
 
         self_kv_cache_config = base_kv_cache_config.model_copy()
         cross_kv_cache_config = base_kv_cache_config.model_copy()
+        if (cross_kv_cache_config.enable_block_reuse
+                and self._encoder_input_is_features()):
+            # Cross blocks are keyed on encoder token ids; a feature-driven
+            # encoder has none, so reuse can never hit and the C++ scheduler
+            # must not be asked for a cross prefix-reuse summary.
+            logger.info(
+                "Disabling block reuse for the cross-KV cache: the encoder "
+                "takes feature tensors, so requests carry no encoder tokens "
+                "to key cross blocks on.")
+            cross_kv_cache_config.enable_block_reuse = False
         split_any_budget = False
 
         free_fraction = base_kv_cache_config.free_gpu_memory_fraction
diff --git a/tensorrt_llm/inputs/registry.py b/tensorrt_llm/inputs/registry.py
index 0963bd0367df..ae5d2c8bdad3 100644
--- a/tensorrt_llm/inputs/registry.py
+++ b/tensorrt_llm/inputs/registry.py
@@ -1061,6 +1061,18 @@ def support_multimodal_disaggregated(model_cls: Type[nn.Module]):
     return model_cls
 
 
+def input_processor_requires_encoder_features(
+        model_cls: Type[nn.Module]) -> bool:
+    """Whether ``model_cls``'s input processor feeds the encoder a feature tensor.
+
+    Feature-driven encoders (e.g. Whisper's audio encoder) take no encoder
+    token ids, so such requests have nothing to key cross-KV blocks on.
+    """
+    processor_cls = INPUT_PROCESSOR_REGISTRY._input_processors_cls_by_model_type.get(
+        model_cls)
+    return bool(getattr(processor_cls, "requires_encoder_features", False))
+
+
 def register_input_processor(
         processor_cls: Type[InputProcessor],
         model_type: str,
diff --git a/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py b/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py
index 8fd9e5f5855b..2dc76d548648 100644
--- a/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py
+++ b/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py
@@ -116,6 +116,7 @@ def _make_llm(
     cuda_graph_batch_sizes: list[int] | None = None,
     tensor_parallel_size: int = 1,
     encoder_graphs: bool = False,
+    enable_block_reuse: bool = False,
 ) -> LLM:
     """Build a Whisper LLM for the test matrix, optionally with encoder CUDA graphs."""
     # CudaGraphConfig captures the decode step; the enc-dec encoder step opts in
@@ -154,7 +155,7 @@ def _make_llm(
         disable_overlap_scheduler=True,  # overlap scheduler unsupported
         enable_chunked_prefill=False,
         kv_cache_config=KvCacheConfig(
-            enable_block_reuse=False,
+            enable_block_reuse=enable_block_reuse,
             free_gpu_memory_fraction=_FREE_GPU_MEMORY_FRACTION,
             cross_kv_cache_fraction=_CROSS_KV_CACHE_FRACTION,
             use_kv_cache_manager_v2=use_kv_cache_manager_v2,
@@ -248,6 +249,33 @@ def test_whisper_pytorch_transcribe_end_to_end(monkeypatch):
 ]
 
 
+def test_whisper_pytorch_block_reuse_enabled(monkeypatch):
+    """Greedy transcription with KV block reuse left at its default (enabled).
+
+    Whisper requests carry encoder features, not encoder token ids, so the
+    cross-KV pool has nothing to key reuse on; the executor must still admit
+    and run them (https://nvbugs/6713231). Batch 2 co-schedules two
+    encoder-init requests, the shape that reached the unguarded cross-reuse
+    lookup in the C++ capacity scheduler.
+    """
+    monkeypatch.setenv("TLLM_WORKER_USE_SINGLE_PROCESS", "1")
+
+    model_path = _get_whisper_model_path()
+    wave, sample_rate = soundfile.read(_get_audio_path())
+    sampling_params = SamplingParams(temperature=0.0, max_tokens=_MAX_NEW_TOKENS)
+
+    with _make_llm(model_path, enable_block_reuse=True) as llm:
+        for batch_size in (1, 2):
+            outputs = llm.generate(
+                [_audio_prompt(wave, sample_rate) for _ in range(batch_size)],
+                sampling_params,
+            )
+            for output in outputs:
+                completion = output.outputs[0]
+                assert list(completion.token_ids) == _EXPECTED_GREEDY_OUTPUT_TOKEN_IDS
+                assert _EXPECTED_TRANSCRIPT_FRAGMENT in completion.text.lower()
+
+
 @pytest.mark.parametrize("torch_dtype,cuda_graph_batch_sizes,graphs_captured", _BEAM_SEARCH_CASES)
 def test_whisper_pytorch_beam_search(
     monkeypatch, torch_dtype, cuda_graph_batch_sizes, graphs_captured
diff --git a/tests/integration/test_lists/test-db/l0_h100.yml b/tests/integration/test_lists/test-db/l0_h100.yml
index d0ac04317b09..1c5b99ff8072 100644
--- a/tests/integration/test_lists/test-db/l0_h100.yml
+++ b/tests/integration/test_lists/test-db/l0_h100.yml
@@ -193,6 +193,7 @@ l0_h100:
   - llmapi/test_llm_api_pytorch_t5.py::test_t5_pytorch_generate_encoder_decoder_end_to_end[bf16-kv-v1-cuda-graph-on-beam2-t5-small]
   - llmapi/test_llm_api_pytorch_t5.py::test_t5_pytorch_generate_encoder_decoder_end_to_end[bf16-kv-v1-cuda-graph-on-greedy-overlap-t5-small]
   - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_transcribe_end_to_end
+  - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_block_reuse_enabled
   - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_feature_combinations[bf16-kv-v2-decoder-graphs-on-greedy]
   - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_beam_search[bf16-kv-v1-decoder-graphs-on-beam2]
   - test_e2e.py::test_trtllm_bench_iteration_log[PyTorch-streaming-meta-llama/Llama-3.1-8B-llama-3.1-model/Meta-Llama-3.1-8B]
diff --git a/tests/integration/test_lists/test-db/l0_l40s.yml b/tests/integration/test_lists/test-db/l0_l40s.yml
index 05361d0813f7..aa70fa119e8a 100644
--- a/tests/integration/test_lists/test-db/l0_l40s.yml
+++ b/tests/integration/test_lists/test-db/l0_l40s.yml
@@ -51,6 +51,7 @@ l0_l40s:
   # The encoder-graphs case also exercises decoder graphs, so it stands in for
   # a decoder-only case rather than adding to it; KV-v2 stays covered on H100.
   - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_transcribe_end_to_end
+  - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_block_reuse_enabled
   - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_feature_combinations[bf16-kv-v1-encoder-graphs-on-greedy]
 - condition:
     ranges:

From 3e80cd1d357d4ba0a7c024f3e5f7a081b775783c Mon Sep 17 00:00:00 2001
From: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
Date: Wed, 16 Sep 2026 14:33:12 -0700
Subject: [PATCH 2/9] [https://nvbugs/6713231][test] Cover feature encoder
 cross-KV scheduling

Signed-off-by: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
---
 .../batch_manager/capacitySchedulerTest.cpp   | 57 +++++++++++++++++++
 1 file changed, 57 insertions(+)

diff --git a/cpp/tests/unit_tests/batch_manager/capacitySchedulerTest.cpp b/cpp/tests/unit_tests/batch_manager/capacitySchedulerTest.cpp
index 804a7f2106f1..9c4689da4943 100644
--- a/cpp/tests/unit_tests/batch_manager/capacitySchedulerTest.cpp
+++ b/cpp/tests/unit_tests/batch_manager/capacitySchedulerTest.cpp
@@ -2608,6 +2608,21 @@ std::shared_ptr<LlmRequest> createEncoderInitRequest(
     EXPECT_EQ(req->getState(), LlmRequestState::kENCODER_INIT);
     return req;
 }
+
+std::shared_ptr<LlmRequest> createFeatureEncoderInitRequest(
+    int32_t promptLen, int32_t maxNewTokens, int32_t encoderInputLen, int32_t encoderOutputLen, uint64_t reqId)
+{
+    auto inputTokens = VecTokens(promptLen, 1);
+    auto executorReq = tensorrt_llm::executor::Request(inputTokens, maxNewTokens);
+    auto encoderInputFeatures = tensorrt_llm::executor::Tensor::cpu(
+        tensorrt_llm::executor::DataType::kFP32, {encoderInputLen, /*featureSize=*/1});
+    executorReq.setEncoderInputFeatures(std::move(encoderInputFeatures));
+    executorReq.setEncoderOutputLength(encoderOutputLen);
+    auto req = std::make_shared<LlmRequest>(reqId, executorReq);
+    EXPECT_EQ(req->getState(), LlmRequestState::kENCODER_INIT);
+    EXPECT_FALSE(req->getEncoderUniqueTokens().has_value());
+    return req;
+}
 } // namespace
 
 // GuaranteedNoEvict: a single encoder-init request is admitted without
@@ -2688,3 +2703,45 @@ TEST_F(CapacitySchedulerTest, EncoderInitDoesNotConsumeCrossPool)
         EXPECT_EQ(crossKvCacheManager->getNumFreeBlocks(), crossFreeBefore) << "policy=" << static_cast<int>(policy);
     }
 }
+
+TEST_F(CapacitySchedulerTest, FeatureEncoderWithoutTokensSkipsCrossPoolReuseAnalysis)
+{
+    SizeType32 const maxNumRequests = 1;
+    SizeType32 const tokensPerBlock = 10;
+    SizeType32 const selfMaxTokens = 100;
+    SizeType32 const crossMaxTokens = 20;
+    int32_t const promptLen = 10;
+    int32_t const encoderInputLen = 20;
+    int32_t const encoderOutputLen = 20;
+
+    auto kvCacheManager = getKvCacheManager(maxNumRequests, tokensPerBlock, selfMaxTokens, selfMaxTokens,
+        /*sinkTokenLength=*/0, /*enableReuse=*/true);
+    auto crossKvCacheManager = getKvCacheManager(maxNumRequests, tokensPerBlock, crossMaxTokens, crossMaxTokens,
+        /*sinkTokenLength=*/0, /*enableReuse=*/true, kv_cache_manager::CacheType::kCROSS);
+    auto peftCacheManager = getPeftCacheManager();
+    auto capacityScheduler
+        = CapacityScheduler(maxNumRequests, CapacitySchedulerPolicy::kGUARANTEED_NO_EVICT, kvCacheManager != nullptr,
+            /*twoStepsLookAhead=*/false, LlmRequestState::kENCODER_INIT, LlmRequestState::kGENERATION_COMPLETE);
+    auto req = createFeatureEncoderInitRequest(
+        promptLen, /*maxNewTokens=*/40, encoderInputLen, encoderOutputLen, /*reqId=*/1);
+    RequestList activeRequests{req};
+
+    auto expectRequestScheduled = [&]()
+    {
+        auto [fittingRequests, fittingDisaggGenInitRequests, pausedRequests]
+            = capacityScheduler(activeRequests, kvCacheManager, peftCacheManager, crossKvCacheManager);
+        ASSERT_EQ(fittingRequests.size(), 1u);
+        EXPECT_EQ(fittingRequests.front()->mRequestId, req->mRequestId);
+        EXPECT_TRUE(fittingDisaggGenInitRequests.empty());
+        EXPECT_TRUE(pausedRequests.empty());
+    };
+
+    // Whisper-like feature encoders have no token IDs to key cross-KV reuse.
+    // Exercise encoder admission, first decoder context, and chunked decoder
+    // context, which use the scheduler's three cross-prefix analysis paths.
+    expectRequestScheduled();
+    req->setState(LlmRequestState::kCONTEXT_INIT);
+    expectRequestScheduled();
+    req->setContextCurrentPosition(1);
+    expectRequestScheduled();
+}

From 203b75fed5203864139271fadb5e15e7bf6d6898 Mon Sep 17 00:00:00 2001
From: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
Date: Mon, 21 Sep 2026 12:00:51 -0700
Subject: [PATCH 3/9] [https://nvbugs/6713231][test] Clarify Whisper reuse
 coverage

Signed-off-by: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
---
 tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py b/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py
index 2dc76d548648..02f68d4c84ac 100644
--- a/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py
+++ b/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py
@@ -250,7 +250,7 @@ def test_whisper_pytorch_transcribe_end_to_end(monkeypatch):
 
 
 def test_whisper_pytorch_block_reuse_enabled(monkeypatch):
-    """Greedy transcription with KV block reuse left at its default (enabled).
+    """Greedy transcription with KV block reuse explicitly enabled.
 
     Whisper requests carry encoder features, not encoder token ids, so the
     cross-KV pool has nothing to key reuse on; the executor must still admit

From 7cb1ad898d481e21121e1c845549b583a381f619 Mon Sep 17 00:00:00 2001
From: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
Date: Tue, 22 Sep 2026 10:27:30 -0700
Subject: [PATCH 4/9] [https://nvbugs/6713231][fix] Disable unsafe Whisper KV
 reuse

Signed-off-by: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
---
 tensorrt_llm/_torch/pyexecutor/_util.py          | 16 +++++++++-------
 .../defs/llmapi/test_llm_api_pytorch_whisper.py  |  8 ++++----
 tests/integration/test_lists/test-db/l0_h100.yml |  2 +-
 tests/integration/test_lists/test-db/l0_l40s.yml |  2 +-
 .../executor/kv_cache/test_dual_pool_kv_cache.py | 12 ++++++++++++
 5 files changed, 27 insertions(+), 13 deletions(-)

diff --git a/tensorrt_llm/_torch/pyexecutor/_util.py b/tensorrt_llm/_torch/pyexecutor/_util.py
index 7e417a292a09..aba1557aa9d6 100644
--- a/tensorrt_llm/_torch/pyexecutor/_util.py
+++ b/tensorrt_llm/_torch/pyexecutor/_util.py
@@ -2095,15 +2095,17 @@ def _split_kv_cache_budget_for_cross(
 
         self_kv_cache_config = base_kv_cache_config.model_copy()
         cross_kv_cache_config = base_kv_cache_config.model_copy()
-        if (cross_kv_cache_config.enable_block_reuse
+        if (base_kv_cache_config.enable_block_reuse
                 and self._encoder_input_is_features()):
-            # Cross blocks are keyed on encoder token ids; a feature-driven
-            # encoder has none, so reuse can never hit and the C++ scheduler
-            # must not be asked for a cross prefix-reuse summary.
+            # Decoder self-KV is conditioned on the encoder output, while
+            # cross-KV is keyed on encoder token ids. Feature-driven encoders
+            # provide neither reusable token ids nor an input discriminator,
+            # so neither pool can safely reuse blocks between requests.
             logger.info(
-                "Disabling block reuse for the cross-KV cache: the encoder "
-                "takes feature tensors, so requests carry no encoder tokens "
-                "to key cross blocks on.")
+                "Disabling block reuse for the self- and cross-KV caches: "
+                "the encoder takes feature tensors, so requests carry no "
+                "encoder tokens or input discriminator to key cache entries.")
+            self_kv_cache_config.enable_block_reuse = False
             cross_kv_cache_config.enable_block_reuse = False
         split_any_budget = False
 
diff --git a/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py b/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py
index 02f68d4c84ac..e244e2df9c99 100644
--- a/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py
+++ b/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py
@@ -249,12 +249,12 @@ def test_whisper_pytorch_transcribe_end_to_end(monkeypatch):
 ]
 
 
-def test_whisper_pytorch_block_reuse_enabled(monkeypatch):
-    """Greedy transcription with KV block reuse explicitly enabled.
+def test_whisper_pytorch_block_reuse_requested(monkeypatch):
+    """Greedy transcription when KV block reuse is requested.
 
     Whisper requests carry encoder features, not encoder token ids, so the
-    cross-KV pool has nothing to key reuse on; the executor must still admit
-    and run them (https://nvbugs/6713231). Batch 2 co-schedules two
+    executor disables reuse for both KV pools and must still admit and run
+    them (https://nvbugs/6713231). Batch 2 co-schedules two
     encoder-init requests, the shape that reached the unguarded cross-reuse
     lookup in the C++ capacity scheduler.
     """
diff --git a/tests/integration/test_lists/test-db/l0_h100.yml b/tests/integration/test_lists/test-db/l0_h100.yml
index 1c5b99ff8072..bf417807f2cf 100644
--- a/tests/integration/test_lists/test-db/l0_h100.yml
+++ b/tests/integration/test_lists/test-db/l0_h100.yml
@@ -193,7 +193,7 @@ l0_h100:
   - llmapi/test_llm_api_pytorch_t5.py::test_t5_pytorch_generate_encoder_decoder_end_to_end[bf16-kv-v1-cuda-graph-on-beam2-t5-small]
   - llmapi/test_llm_api_pytorch_t5.py::test_t5_pytorch_generate_encoder_decoder_end_to_end[bf16-kv-v1-cuda-graph-on-greedy-overlap-t5-small]
   - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_transcribe_end_to_end
-  - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_block_reuse_enabled
+  - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_block_reuse_requested
   - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_feature_combinations[bf16-kv-v2-decoder-graphs-on-greedy]
   - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_beam_search[bf16-kv-v1-decoder-graphs-on-beam2]
   - test_e2e.py::test_trtllm_bench_iteration_log[PyTorch-streaming-meta-llama/Llama-3.1-8B-llama-3.1-model/Meta-Llama-3.1-8B]
diff --git a/tests/integration/test_lists/test-db/l0_l40s.yml b/tests/integration/test_lists/test-db/l0_l40s.yml
index aa70fa119e8a..6bd8a8f588d2 100644
--- a/tests/integration/test_lists/test-db/l0_l40s.yml
+++ b/tests/integration/test_lists/test-db/l0_l40s.yml
@@ -51,7 +51,7 @@ l0_l40s:
   # The encoder-graphs case also exercises decoder graphs, so it stands in for
   # a decoder-only case rather than adding to it; KV-v2 stays covered on H100.
   - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_transcribe_end_to_end
-  - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_block_reuse_enabled
+  - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_block_reuse_requested
   - llmapi/test_llm_api_pytorch_whisper.py::test_whisper_pytorch_feature_combinations[bf16-kv-v1-encoder-graphs-on-greedy]
 - condition:
     ranges:
diff --git a/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py b/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
index a6efac87bba1..9a483a957f3a 100644
--- a/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
+++ b/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
@@ -289,6 +289,18 @@ def test_split_free_fraction_when_budget_is_zero(self):
         assert self_config.free_gpu_memory_fraction == pytest.approx(0.4)
         assert config.free_gpu_memory_fraction == pytest.approx(0.8)
 
+    def test_feature_encoder_disables_reuse_for_both_pools(self):
+        """Feature inputs cannot safely key self- or cross-KV reuse."""
+        config = _make_kv_cache_config(cross_kv_cache_fraction=0.5)
+        creator = _make_creator(config, is_enc_dec=True)
+        creator._encoder_input_is_features = Mock(return_value=True)
+
+        self_config, cross_config = creator._split_kv_cache_budget_for_cross()
+
+        assert config.enable_block_reuse
+        assert not self_config.enable_block_reuse
+        assert not cross_config.enable_block_reuse
+
     def test_is_encoder_decoder_helper(self):
         dec_config = _make_model_config(is_encoder_decoder=False)
         dec_creator = _make_creator(_make_kv_cache_config(), model_config=dec_config)

From 38e2461df0f7643626ea5bdb900bc8c74d47ac43 Mon Sep 17 00:00:00 2001
From: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
Date: Wed, 23 Sep 2026 11:13:17 -0700
Subject: [PATCH 5/9] [https://nvbugs/6713231][test] Cover feature encoder
 manager configs

Signed-off-by: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
---
 .../kv_cache/test_dual_pool_kv_cache.py       | 25 +++++++++++++++++++
 1 file changed, 25 insertions(+)

diff --git a/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py b/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
index 25591e54be4a..ad87588c895b 100644
--- a/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
+++ b/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
@@ -696,6 +696,31 @@ def create_cross_manager(cross_cfg, *_args, **_kwargs):
             (pytest.approx(0.45), expected_split),
         ]
 
+    @pytest.mark.parametrize("use_kv_cache_manager_v2", [False, True])
+    def test_build_managers_disables_feature_encoder_reuse(
+        self, use_kv_cache_manager_v2: bool
+    ) -> None:
+        """Pass reuse-disabled configs to both feature-encoder managers."""
+        config = _make_kv_cache_config(
+            cross_kv_cache_fraction=0.5,
+            max_gpu_total_bytes=8 * (1 << 30),
+            use_kv_cache_manager_v2=use_kv_cache_manager_v2,
+        )
+        creator = _make_creator(config, is_enc_dec=True)
+        creator.configure_kv_cache_capacity = Mock()
+        creator._encoder_input_is_features = Mock(return_value=True)
+        creator._should_create_separate_draft_kv_cache = Mock(return_value=False)
+        creator._create_kv_cache_manager = Mock(return_value=Mock())
+        creator._create_cross_kv_cache_manager = Mock(return_value=Mock())
+
+        creator.build_managers({}, estimating_kv_cache=False)
+
+        self_config = creator._create_kv_cache_manager.call_args.kwargs["kv_cache_config_override"]
+        cross_config = creator._create_cross_kv_cache_manager.call_args.args[0]
+        assert config.enable_block_reuse
+        assert not self_config.enable_block_reuse
+        assert not cross_config.enable_block_reuse
+
     def test_build_managers_skips_cross_pool_for_decoder_only(self):
         creator = _make_creator(
             _make_kv_cache_config(

From 1ede81044c31883512b7c5673161879f0cb2896d Mon Sep 17 00:00:00 2001
From: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
Date: Wed, 23 Sep 2026 11:25:17 -0700
Subject: [PATCH 6/9] [https://nvbugs/6713231][test] Preserve token encoder
 reuse coverage

Signed-off-by: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
---
 .../executor/kv_cache/test_dual_pool_kv_cache.py      | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py b/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
index ad87588c895b..570b8ed847f5 100644
--- a/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
+++ b/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
@@ -302,6 +302,17 @@ def test_feature_encoder_disables_reuse_for_both_pools(self):
         assert not self_config.enable_block_reuse
         assert not cross_config.enable_block_reuse
 
+    def test_token_encoder_preserves_reuse_for_both_pools(self) -> None:
+        """Token inputs retain reusable identities for both KV pools."""
+        config = _make_kv_cache_config(cross_kv_cache_fraction=0.5)
+        creator = _make_creator(config, is_enc_dec=True)
+        creator._encoder_input_is_features = Mock(return_value=False)
+
+        self_config, cross_config = creator._split_kv_cache_budget_for_cross()
+
+        assert self_config.enable_block_reuse
+        assert cross_config.enable_block_reuse
+
     def test_is_encoder_decoder_helper(self):
         dec_config = _make_model_config(is_encoder_decoder=False)
         dec_creator = _make_creator(_make_kv_cache_config(), model_config=dec_config)

From ae3ae95622f2b6a7e58e0687ce39da3102b08313 Mon Sep 17 00:00:00 2001
From: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
Date: Wed, 23 Sep 2026 13:47:06 -0700
Subject: [PATCH 7/9] [https://nvbugs/6713231][fix] Disable paged context for
 feature encoders

Signed-off-by: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
---
 tensorrt_llm/_torch/pyexecutor/_util.py                     | 3 +++
 .../_torch/executor/kv_cache/test_dual_pool_kv_cache.py     | 6 ++++++
 2 files changed, 9 insertions(+)

diff --git a/tensorrt_llm/_torch/pyexecutor/_util.py b/tensorrt_llm/_torch/pyexecutor/_util.py
index 2fb318fd349a..7570b8bf840e 100644
--- a/tensorrt_llm/_torch/pyexecutor/_util.py
+++ b/tensorrt_llm/_torch/pyexecutor/_util.py
@@ -2372,6 +2372,9 @@ def _split_kv_cache_budget_for_cross(
                 "encoder tokens or input discriminator to key cache entries.")
             self_kv_cache_config.enable_block_reuse = False
             cross_kv_cache_config.enable_block_reuse = False
+            # The attention backend reads this shared runtime flag rather than
+            # either derived manager config when selecting paged-context FMHA.
+            self._model_engine.attn_runtime_features.cache_reuse = False
         split_any_budget = False
 
         free_fraction = base_kv_cache_config.free_gpu_memory_fraction
diff --git a/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py b/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
index 570b8ed847f5..0903070dca2e 100644
--- a/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
+++ b/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
@@ -161,6 +161,9 @@ def _make_creator(
     if model_config is None:
         model_config = _make_model_config(is_encoder_decoder=is_enc_dec)
     model_engine = _make_mock_model_engine(model_config)
+    model_engine.attn_runtime_features = SimpleNamespace(
+        cache_reuse=kv_cache_config.enable_block_reuse
+    )
 
     if manager_cls is None:
         manager_cls = (
@@ -301,6 +304,7 @@ def test_feature_encoder_disables_reuse_for_both_pools(self):
         assert config.enable_block_reuse
         assert not self_config.enable_block_reuse
         assert not cross_config.enable_block_reuse
+        assert not creator._model_engine.attn_runtime_features.cache_reuse
 
     def test_token_encoder_preserves_reuse_for_both_pools(self) -> None:
         """Token inputs retain reusable identities for both KV pools."""
@@ -312,6 +316,7 @@ def test_token_encoder_preserves_reuse_for_both_pools(self) -> None:
 
         assert self_config.enable_block_reuse
         assert cross_config.enable_block_reuse
+        assert creator._model_engine.attn_runtime_features.cache_reuse
 
     def test_is_encoder_decoder_helper(self):
         dec_config = _make_model_config(is_encoder_decoder=False)
@@ -731,6 +736,7 @@ def test_build_managers_disables_feature_encoder_reuse(
         assert config.enable_block_reuse
         assert not self_config.enable_block_reuse
         assert not cross_config.enable_block_reuse
+        assert not creator._model_engine.attn_runtime_features.cache_reuse
 
     def test_build_managers_skips_cross_pool_for_decoder_only(self):
         creator = _make_creator(

From e6b325788f2e2bdb38839363909ec990b594fabd Mon Sep 17 00:00:00 2001
From: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
Date: Wed, 23 Sep 2026 14:04:22 -0700
Subject: [PATCH 8/9] [https://nvbugs/6713231][test] Cover Whisper registry and
 C++ scheduler

Signed-off-by: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
---
 .../defs/llmapi/test_llm_api_pytorch_whisper.py        |  9 +++++++--
 .../unittest/_torch/modeling/test_modeling_whisper.py  | 10 +++++++++-
 2 files changed, 16 insertions(+), 3 deletions(-)

diff --git a/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py b/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py
index e244e2df9c99..f3860c8461e0 100644
--- a/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py
+++ b/tests/integration/defs/llmapi/test_llm_api_pytorch_whisper.py
@@ -117,6 +117,7 @@ def _make_llm(
     tensor_parallel_size: int = 1,
     encoder_graphs: bool = False,
     enable_block_reuse: bool = False,
+    use_python_scheduler: bool = True,
 ) -> LLM:
     """Build a Whisper LLM for the test matrix, optionally with encoder CUDA graphs."""
     # CudaGraphConfig captures the decode step; the enc-dec encoder step opts in
@@ -166,7 +167,7 @@ def _make_llm(
         # 1500 encoder positions every Whisper request produces.
         max_input_len=_ENCODER_OUTPUT_LEN,
         max_num_tokens=2 * _ENCODER_OUTPUT_LEN,
-        scheduler_config=SchedulerConfig(use_python_scheduler=True),
+        scheduler_config=SchedulerConfig(use_python_scheduler=use_python_scheduler),
         tensor_parallel_size=tensor_parallel_size,
         **encoder_kwargs,
         **dtype_kwargs,
@@ -264,7 +265,11 @@ def test_whisper_pytorch_block_reuse_requested(monkeypatch):
     wave, sample_rate = soundfile.read(_get_audio_path())
     sampling_params = SamplingParams(temperature=0.0, max_tokens=_MAX_NEW_TOKENS)
 
-    with _make_llm(model_path, enable_block_reuse=True) as llm:
+    with _make_llm(
+        model_path,
+        enable_block_reuse=True,
+        use_python_scheduler=False,
+    ) as llm:
         for batch_size in (1, 2):
             outputs = llm.generate(
                 [_audio_prompt(wave, sample_rate) for _ in range(batch_size)],
diff --git a/tests/unittest/_torch/modeling/test_modeling_whisper.py b/tests/unittest/_torch/modeling/test_modeling_whisper.py
index feb993d258a8..96abd9baf336 100644
--- a/tests/unittest/_torch/modeling/test_modeling_whisper.py
+++ b/tests/unittest/_torch/modeling/test_modeling_whisper.py
@@ -25,7 +25,15 @@
 import torch
 from transformers import WhisperConfig, WhisperFeatureExtractor
 
-from tensorrt_llm._torch.models.modeling_whisper import WhisperLogMelFrontend
+from tensorrt_llm._torch.models.modeling_whisper import (
+    WhisperForConditionalGeneration,
+    WhisperLogMelFrontend,
+)
+from tensorrt_llm.inputs.registry import input_processor_requires_encoder_features
+
+
+def test_whisper_input_processor_requires_encoder_features():
+    assert input_processor_requires_encoder_features(WhisperForConditionalGeneration)
 
 
 def _synthetic_waveform_batch(n_samples: int, seed: int = 1234) -> np.ndarray:

From 4d337807276ee49a139361a9ba85adc48b58f8b4 Mon Sep 17 00:00:00 2001
From: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
Date: Wed, 23 Sep 2026 14:16:50 -0700
Subject: [PATCH 9/9] [https://nvbugs/6713231][fix] Detect feature encoders
 after compilation

Signed-off-by: Pranav Shrestha <254760092+pranav-nvidia@users.noreply.github.com>
---
 tensorrt_llm/_torch/pyexecutor/_util.py       |  7 +++----
 tensorrt_llm/inputs/registry.py               | 12 ------------
 .../kv_cache/test_dual_pool_kv_cache.py       | 19 ++++++++++++++-----
 .../_torch/modeling/test_modeling_whisper.py  |  8 ++------
 4 files changed, 19 insertions(+), 27 deletions(-)

diff --git a/tensorrt_llm/_torch/pyexecutor/_util.py b/tensorrt_llm/_torch/pyexecutor/_util.py
index 7570b8bf840e..c48adab79919 100644
--- a/tensorrt_llm/_torch/pyexecutor/_util.py
+++ b/tensorrt_llm/_torch/pyexecutor/_util.py
@@ -25,8 +25,6 @@
                                  is_sm_100f, prefer_pinned,
                                  str_dtype_to_binding, torch_dtype_to_str)
 from tensorrt_llm.inputs.multimodal import MultimodalParams
-from tensorrt_llm.inputs.registry import \
-    input_processor_requires_encoder_features
 
 # isort: off
 from tensorrt_llm.llmapi.llm_args import (
@@ -2230,8 +2228,9 @@ def _is_encoder_decoder(self) -> bool:
         return self._model_engine.model.model_config.is_encoder_decoder
 
     def _encoder_input_is_features(self) -> bool:
-        return input_processor_requires_encoder_features(
-            type(self._model_engine.model))
+        return bool(
+            getattr(self._model_engine.input_processor,
+                    "requires_encoder_features", False))
 
     @staticmethod
     def _get_config_int_attr(config, names: tuple[str, ...]) -> Optional[int]:
diff --git a/tensorrt_llm/inputs/registry.py b/tensorrt_llm/inputs/registry.py
index f63a0f62a640..985075db25ff 100644
--- a/tensorrt_llm/inputs/registry.py
+++ b/tensorrt_llm/inputs/registry.py
@@ -1074,18 +1074,6 @@ def support_multimodal_disaggregated(model_cls: Type[nn.Module]):
     return model_cls
 
 
-def input_processor_requires_encoder_features(
-        model_cls: Type[nn.Module]) -> bool:
-    """Whether ``model_cls``'s input processor feeds the encoder a feature tensor.
-
-    Feature-driven encoders (e.g. Whisper's audio encoder) take no encoder
-    token ids, so such requests have nothing to key cross-KV blocks on.
-    """
-    processor_cls = INPUT_PROCESSOR_REGISTRY._input_processors_cls_by_model_type.get(
-        model_cls)
-    return bool(getattr(processor_cls, "requires_encoder_features", False))
-
-
 def register_input_processor(
         processor_cls: Type[InputProcessor],
         model_type: str,
diff --git a/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py b/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
index 0903070dca2e..9e1fe4b878e4 100644
--- a/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
+++ b/tests/unittest/_torch/executor/kv_cache/test_dual_pool_kv_cache.py
@@ -50,6 +50,14 @@ class _FakeCudaStream:
     cuda_stream = 0
 
 
+class _FakeCompiledModel:
+    """Minimal torch.compile-style wrapper that changes the model's type."""
+
+    def __init__(self, original_model):
+        self._orig_mod = original_model
+        self.model_config = original_model.model_config
+
+
 class _FakeKVCacheManagerCpp:
     def __init__(self, **kwargs):
         self.kwargs = kwargs
@@ -141,6 +149,7 @@ def _make_mock_model_engine(model_config):
     engine.dtype = torch.bfloat16
     engine.is_draft_model = False
     engine.kv_cache_manager_key = ResourceManagerType.KV_CACHE_MANAGER
+    engine.input_processor = SimpleNamespace(requires_encoder_features=False)
     return engine
 
 
@@ -293,11 +302,12 @@ def test_split_free_fraction_when_budget_is_zero(self):
         assert self_config.free_gpu_memory_fraction == pytest.approx(0.4)
         assert config.free_gpu_memory_fraction == pytest.approx(0.8)
 
-    def test_feature_encoder_disables_reuse_for_both_pools(self):
-        """Feature inputs cannot safely key self- or cross-KV reuse."""
+    def test_wrapped_feature_encoder_disables_reuse_for_both_pools(self):
+        """Feature detection survives a torch.compile-style model wrapper."""
         config = _make_kv_cache_config(cross_kv_cache_fraction=0.5)
         creator = _make_creator(config, is_enc_dec=True)
-        creator._encoder_input_is_features = Mock(return_value=True)
+        creator._model_engine.model = _FakeCompiledModel(creator._model_engine.model)
+        creator._model_engine.input_processor.requires_encoder_features = True
 
         self_config, cross_config = creator._split_kv_cache_budget_for_cross()
 
@@ -310,7 +320,6 @@ def test_token_encoder_preserves_reuse_for_both_pools(self) -> None:
         """Token inputs retain reusable identities for both KV pools."""
         config = _make_kv_cache_config(cross_kv_cache_fraction=0.5)
         creator = _make_creator(config, is_enc_dec=True)
-        creator._encoder_input_is_features = Mock(return_value=False)
 
         self_config, cross_config = creator._split_kv_cache_budget_for_cross()
 
@@ -724,7 +733,7 @@ def test_build_managers_disables_feature_encoder_reuse(
         )
         creator = _make_creator(config, is_enc_dec=True)
         creator.configure_kv_cache_capacity = Mock()
-        creator._encoder_input_is_features = Mock(return_value=True)
+        creator._model_engine.input_processor.requires_encoder_features = True
         creator._should_create_separate_draft_kv_cache = Mock(return_value=False)
         creator._create_kv_cache_manager = Mock(return_value=Mock())
         creator._create_cross_kv_cache_manager = Mock(return_value=Mock())
diff --git a/tests/unittest/_torch/modeling/test_modeling_whisper.py b/tests/unittest/_torch/modeling/test_modeling_whisper.py
index 96abd9baf336..90b13a7daf7c 100644
--- a/tests/unittest/_torch/modeling/test_modeling_whisper.py
+++ b/tests/unittest/_torch/modeling/test_modeling_whisper.py
@@ -25,15 +25,11 @@
 import torch
 from transformers import WhisperConfig, WhisperFeatureExtractor
 
-from tensorrt_llm._torch.models.modeling_whisper import (
-    WhisperForConditionalGeneration,
-    WhisperLogMelFrontend,
-)
-from tensorrt_llm.inputs.registry import input_processor_requires_encoder_features
+from tensorrt_llm._torch.models.modeling_whisper import WhisperInputProcessor, WhisperLogMelFrontend
 
 
 def test_whisper_input_processor_requires_encoder_features():
-    assert input_processor_requires_encoder_features(WhisperForConditionalGeneration)
+    assert WhisperInputProcessor.requires_encoder_features
 
 
 def _synthetic_waveform_batch(n_samples: int, seed: int = 1234) -> np.ndarray: