[None][feat] Add isolated NVFP4 Kimi K3 targets with fused FP8 attention

tensorrt_llm/_torch/_experimental/modeling_v2/README.md

@@ -1,7 +1,6 @@
 # ModelingV2
 
-One self-contained modeling codebase per deployment target, beside the
-built-in model zoo rather than inside it.
+Explicit deployment targets beside the built-in model zoo.
 
 Where `_torch/models/modeling_deepseekv3.py` is one class serving V3, V3-Lite,
 R1 and V3.2 across every GPU generation and parallel topology,
@@ -11,6 +10,43 @@ entries, sharing nothing with its siblings, and trusted through accuracy gates
 instead of shared abstractions. The one-to-one correspondence between
 `models/<x>/` here and `modeling_<x>.py` there is the point of the exercise.
 
+Kimi K3 keeps its original MXFP4/SM100 deployments independent. The NVFP4
+implementation is a separate copy under `models/kimi_k3_vl/kimi_k3_nvfp4/`,
+shared by thin SM103 and SM107 target classes. NVFP4 retains checkpoint FP8
+projection weights and activations: decode fuses KDA's gate, convolution,
+recurrence and output normalization, and MLA's query/absorption/cache writes
+and attention/V-B/output gate. The ordinary FP8 GEMMs remain responsible for
+input and output projections. MLA's fused decode uses a BF16 KV cache with
+64-token pages, or composes the fused FP8 projections with the native FP8
+and NVFP4 cache/attention kernels. NVFP4 KV requires 128-token pages and the
+V2 cache manager; its high-precision tail and packed scales retain the native
+layout. The KDA fused decode and verify accept FP32 or BF16 recurrent-state
+storage, with FP32 recurrence arithmetic and FP32 per-token draft records.
+These KV and SSM storage choices are independent of the NVFP4 expert weights.
+The fused Q-B projection follows the selected GEMM backend's FP8 recipe,
+including weight resmoothing and UE8M0 activation scales where required.
+
+| MLA KV storage | KDA SSM storage | KDA recurrence / draft records |
+|---|---|---|
+| BF16 | FP32 or BF16 | FP32 |
+| FP8 E4M3 | FP32 or BF16 | FP32 |
+| NVFP4 | FP32 or BF16 | FP32 |
+
+For NVFP4 KV and BF16 recurrent-state storage:
+
+```yaml
+modeling_v2: require
+enable_chunked_prefill: false
+kv_cache_config:
+  dtype: nvfp4
+  tokens_per_block: 128
+  mamba_ssm_cache_dtype: bfloat16
+  use_kv_cache_manager_v2: true
+```
+
+The NVFP4 expert kernels remain separate from the MXFP4/MXFP8 kernels.
+Every hardware and quantization target requires its own accuracy gates.
+
 ## Using it
 
 ```python

tensorrt_llm/_torch/_experimental/modeling_v2/models/kimi_k3_vl/kimi_k3_mxfp4__sm_103__tp16_moetp16ep1/__init__.py

@@ -0,0 +1,3 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""Kimi K3 MXFP4 / sm_103 / tp16_moetp16ep1."""

tensorrt_llm/_torch/_experimental/modeling_v2/models/kimi_k3_vl/kimi_k3_mxfp4__sm_103__tp16_moetp16ep1/modeling.py

@@ -0,0 +1,21 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""Kimi K3 MXFP4, SM103, attention TP16 and MoE TP16 x EP1."""
+
+from tensorrt_llm._torch.model_config import ModelConfig
+from tensorrt_llm.quantization.mode import QuantAlgo
+
+from ..modeling import K3TargetSpec, ModelingV2KimiK3
+from ..modeling import _check_construction as _check_target_construction
+
+_TARGET = K3TargetSpec(sm=(10, 3), quant_algo=QuantAlgo.W4A8_MXFP4_MXFP8, moe_tp=16, moe_ep=1)
+
+
+def _check_construction(model_config: ModelConfig) -> None:
+    _check_target_construction(model_config, _TARGET)
+
+
+class ModelingV2KimiK3Mxfp4Sm103Tp16Moetp16ep1(ModelingV2KimiK3):
+    """Concrete deployment of the shared Kimi K3 text model."""
+
+    target_spec = _TARGET

tensorrt_llm/_torch/_experimental/modeling_v2/models/kimi_k3_vl/kimi_k3_mxfp4__sm_103__tp16_moetp16ep1/weights.py

@@ -0,0 +1,6 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""The shared Kimi K3 text checkpoint loader."""
+
+from ..weights import PREDICTED_NON_LOAD as PREDICTED_NON_LOAD
+from ..weights import load as load

tensorrt_llm/_torch/_experimental/modeling_v2/models/kimi_k3_vl/kimi_k3_mxfp4__sm_103__tp16_moetp4ep4/__init__.py

@@ -0,0 +1,3 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""Kimi K3 MXFP4 / sm_103 / tp16_moetp4ep4."""

tensorrt_llm/_torch/_experimental/modeling_v2/models/kimi_k3_vl/kimi_k3_mxfp4__sm_103__tp16_moetp4ep4/modeling.py

@@ -0,0 +1,21 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""Kimi K3 MXFP4, SM103, attention TP16 and MoE TP4 x EP4."""
+
+from tensorrt_llm._torch.model_config import ModelConfig
+from tensorrt_llm.quantization.mode import QuantAlgo
+
+from ..modeling import K3TargetSpec, ModelingV2KimiK3
+from ..modeling import _check_construction as _check_target_construction
+
+_TARGET = K3TargetSpec(sm=(10, 3), quant_algo=QuantAlgo.W4A8_MXFP4_MXFP8, moe_tp=4, moe_ep=4)
+
+
+def _check_construction(model_config: ModelConfig) -> None:
+    _check_target_construction(model_config, _TARGET)
+
+
+class ModelingV2KimiK3Mxfp4Sm103Tp16Moetp4ep4(ModelingV2KimiK3):
+    """Concrete deployment of the shared Kimi K3 text model."""
+
+    target_spec = _TARGET

tensorrt_llm/_torch/_experimental/modeling_v2/models/kimi_k3_vl/kimi_k3_mxfp4__sm_103__tp16_moetp4ep4/weights.py

@@ -0,0 +1,6 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""The shared Kimi K3 text checkpoint loader."""
+
+from ..weights import PREDICTED_NON_LOAD as PREDICTED_NON_LOAD
+from ..weights import load as load

tensorrt_llm/_torch/_experimental/modeling_v2/models/kimi_k3_vl/kimi_k3_nvfp4/__init__.py

@@ -0,0 +1,3 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+"""NVFP4 Kimi K3 targets with FP8 fused attention on SM103 and SM107."""