[https://nvbugs/6665906][fix] Restore Mistral Large 3 multimodal inference

cpp/kernels/fmha_v2/setup.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2020-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2020-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -2199,8 +2199,10 @@ def selected_mask_types(kspec):
             custom_mask = '0'
         # encoder models (head_size = 32 / 64 / 128) need packed_qkv input layout + padding mask.
         elif kspec.input_layout == InputLayout.PACKED_QKV:
-            # NOTE: 72/80 are added for vision transformer
-            if kspec.head_size not in [32, 64, 72, 80, 128]:
+            # NOTE: 72/80 are added for vision transformers. Pixtral's head size
+            # 104 is needed on SM100, where it falls back from TRTLLM-GEN.
+            if (kspec.head_size not in [32, 64, 72, 80, 128]
+                    and not (kspec.sm == 100 and kspec.head_size == 104)):
                 padding_mask = '0'
         # only cross attention (head_size = 32/64/128) needs contiguous_q_kv input layout + padding mask / custom_mask.
         elif kspec.input_layout == InputLayout.CONTIGUOUS_Q_KV:
@@ -6904,10 +6906,10 @@ def enumerate_kernels():
                   and kspec.cross_mha     == False
                   and kspec.flash_attention == True
                   and kspec.input_layout != InputLayout.SEPARATE_Q_K_V)
-                  # Gemma3 VL support.
+                  # Gemma3 VL and Pixtral VL support.
                   or  (kspec.sm           == 100
                   and kspec.dtype         in ['fp16', 'bf16', 'fp16_fp32', 'e4m3', 'e4m3_fp32']
-                  and kspec.head_size     == 72
+                  and kspec.head_size     in [72, 104]
                   and kspec.head_size_v   == 0
                   and kspec.sage_block_sizes is None
                   and kspec.version       == 2

cpp/tensorrt_llm/kernels/fmhaDispatcher.cpp

@@ -49,10 +49,12 @@ QkvLayout AttentionInputLayoutToQkvLayout(AttentionInputLayout layout)
 
 FmhaDispatcher::FmhaDispatcher(MHARunnerFixedParams fixedParams)
     : mFixedParams(fixedParams)
-    // TRTLLM-GEN only supports power of 2 head sizes (and 80 with padding).
-    // The exception will fall back to fmha v2.
+    // TRTLLM-GEN has no head-size 72 kernel or packed-QKV head-size 104 vision kernel.
+    // Keep mixed Q/KV precisions on TRTLLM-GEN: fmha v2 requires matching types.
     // Please update fmha_v2/setup.py if you want to add more supported head sizes.
-    , mUseTllmGen(tensorrt_llm::common::isSM100Family() && fixedParams.headSize != 72)
+    , mUseTllmGen(tensorrt_llm::common::isSM100Family() && fixedParams.headSize != 72
+          && !(fixedParams.headSize == 104 && fixedParams.attentionInputLayout == AttentionInputLayout::PACKED_QKV
+              && fixedParams.dataType == fixedParams.dataTypeKv))
     , mMultiProcessorCount(tensorrt_llm::common::getMultiProcessorCount())
 {
     if (mUseTllmGen)