[None][fix] fix eagle3 accuracy issue on sm120 (#8944)

Signed-off-by: bhsueh <11360707+byshiue@users.noreply.github.com>

Author: byshiue

cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpp

@@ -375,7 +375,7 @@ void DecoderXQAImplJIT::runImpl(XQAParams const& xqaParams, KVCacheBuffer const&
             .mask = reinterpret_cast<SpecDecParams::MaskType const*>(xqaParams.spec_decoding_packed_mask)};
     };
 
-    constexpr uint32_t kMAX_NB_KERNEL_PARAMS = 15;
+    constexpr uint32_t kMAX_NB_KERNEL_PARAMS = 16;
     uint32_t idxNextParam = 0;
     void* kernelParams[kMAX_NB_KERNEL_PARAMS];
     auto appendParam = [&](auto* p) mutable

tensorrt_llm/_torch/attention_backend/trtllm.py

@@ -1076,9 +1076,9 @@ def update_spec_dec_param(
             spec_decoding_position_offsets = None
             spec_decoding_packed_mask = None
             spec_decoding_generation_lengths = None
-        # spec_dec mode should only be enabled for pre-Blackwell machines and when there's a spec-dec tree.
-        self.is_spec_decoding_enabled = is_spec_decoding_enabled and get_sm_version(
-        ) < 100
+        # spec_dec mode should only be enabled for non-sm100 machines and when there's a spec-dec tree.
+        self.is_spec_decoding_enabled = is_spec_decoding_enabled and (
+            get_sm_version() < 100 or get_sm_version() == 120)
 
         if get_sm_version() >= 100:
             if is_spec_dec_tree or is_spec_dec_dynamic_tree: