[None][fix] Do not leak KV cache quantization into vision encoder

tensorrt_llm/_torch/models/modeling_radio.py

@@ -1024,10 +1024,11 @@ def __init__(self,
         self.model_config = copy.deepcopy(model_config)
         if self.model_config.quant_config is not None:
             if disable_quantization:
-                # The basic method `apply_quant_config_exclude_modules` in DecoderModelForCausalLM keeps the kv_cache_quant_algo so we also keep it here.
-                self.model_config.quant_config = QuantConfig(
-                    kv_cache_quant_algo=self.model_config.quant_config.
-                    kv_cache_quant_algo)
+                # Vision encoder runs with kv_cache_manager=None, so there is no KV cache to
+                # quantize. Keeping kv_cache_quant_algo would make FlashInfer raise:
+                # "FP8 KV cache is not supported without a KV cache manager" for FP8 LLM checkpoints
+                # that specify FP8 KV Cache.
+                self.model_config.quant_config = QuantConfig()
 
         self.model_config = dataclasses.replace(
             self.model_config, attn_backend=vision_attn_backend)

tests/integration/test_lists/test-db/l0_a10.yml

@@ -23,6 +23,7 @@ l0_a10:
   - unittest/_torch/modeling/test_modeling_cohere2.py
   - unittest/_torch/modeling/test_nemotron_nano_preprocessing.py
   - unittest/_torch/modeling/test_modeling_parakeet.py
+  - unittest/_torch/modeling/test_modeling_radio.py
   - unittest/_torch/sampler/test_trtllm_sampler.py
   - unittest/_torch/executor/test_async_transfer_manager.py
   - unittest/_torch/executor/test_scheduler_serializable_output.py

tests/unittest/_torch/modeling/test_modeling_radio.py

@@ -0,0 +1,95 @@
+# Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved.
+from unittest import mock
+
+import pytest
+import torch
+from transformers import PretrainedConfig
+
+from tensorrt_llm._torch import model_config as model_config_lib
+from tensorrt_llm._torch.models import modeling_radio
+from tensorrt_llm._torch.models.modeling_radio import RADIOVisionModel
+from tensorrt_llm.models.modeling_utils import QuantConfig
+from tensorrt_llm.quantization.mode import QuantAlgo
+
+_TINY_VIT = modeling_radio.VITTIMMConfig(
+    embed_dim=64,
+    depth=2,
+    num_attention_heads=2,
+    intermediate_size=128,
+    img_size=32,
+)
+
+
+def _make_vision_config():
+    """Minimal PretrainedConfig mimicking Nemotron-Nano-V3's `vision_config`.
+
+    Patterned on the `vision_config` block of config.json shipped with newer nemotron nano
+    multimodal models; fields pared down to what RADIOVisionModel reads.
+    """
+    config = PretrainedConfig()
+    config.patch_size = 16
+    config.adaptor_names = None
+    config.feature_normalizer_config = None
+    config.inter_feature_normalizer_config = None
+    config.max_resolution = 64
+    config.vitdet_window_size = None
+    config.preferred_resolution = (32, 32)
+    config.video_temporal_patch_size = 1
+    config.separate_video_embedder = True
+    config.torch_dtype = torch.bfloat16
+    config.args = {
+        "model": "vit_tiny_test",
+        "in_chans": None,
+        "input_size": None,
+        "drop": 0.0,
+        "cpe_max_size": 64,
+        "cls_token_per_teacher": False,
+        "teachers": [{"name": "dummy"}],
+        "register_multiple": None,
+        "cpe_num_registers": None,
+    }
+    return config
+
+
+@pytest.fixture
+def tiny_vit_config():
+    with mock.patch.dict(
+        modeling_radio.VIT_TIMM_CONFIG_BY_NAME,
+        {"vit_tiny_test": _TINY_VIT},
+    ):
+        yield
+
+
+def _make_fp8_model_config():
+    vision_config = _make_vision_config()
+    quant_config = QuantConfig(
+        quant_algo=QuantAlgo.FP8,
+        kv_cache_quant_algo=QuantAlgo.FP8,
+    )
+    return model_config_lib.ModelConfig(
+        pretrained_config=vision_config,
+        quant_config=quant_config,
+    )
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")
+def test_radio_fp8_parent_kv_cache_does_not_leak_into_vit(tiny_vit_config):
+    """When the parent LLM uses FP8 KV cache, the RADIO vision encoder must not inherit it.
+
+    A ViT has no KV cache (kv_cache_manager=None). If `kv_cache_quant_algo=FP8` leaks into the
+    vision tower, FlashInfer raises at forward time about it not being supported.
+    """
+    vision_model = RADIOVisionModel(_make_fp8_model_config(), disable_quantization=True)
+
+    device = torch.device("cuda")
+    dtype = torch.bfloat16
+    vision_model = vision_model.to(device).to(dtype)
+
+    # 32x32 image: multiple of `patch_size=16`, so `min_resolution_step` is satisfied.
+    pixel_values = torch.randn(1, 3, 32, 32, device=device, dtype=dtype)
+
+    with torch.inference_mode():
+        features = vision_model.forward(pixel_values)
+
+    assert features.shape[0] == 1
+    assert features.shape[-1] == _TINY_VIT.embed_dim