ROCm · zaneni6 · Aug 17, 2026 · Aug 17, 2026 · Aug 17, 2026
diff --git a/src/CMakePresets.json b/src/CMakePresets.json
@@ -5,7 +5,7 @@
       "name": "common-default",
       "hidden": true,
       "cacheVariables": {
-        "FLM_VERSION": "1.0.1",
+        "FLM_VERSION": "1.0.2",
         "NPU_VERSION": "32.0.203.304"
       }
     },

diff --git a/src/inno/flm.iss b/src/inno/flm.iss
@@ -4,7 +4,7 @@
 
 AppName=flm
 
-AppVersion=1.0.1
+AppVersion=1.0.2
 
 AppPublisher=FastFlowLM
 

diff --git a/src/lib/xrt/dequant.dll b/src/lib/xrt/dequant.dll
diff --git a/src/lib/xrt/gemm.dll b/src/lib/xrt/gemm.dll
diff --git a/src/lib/xrt/gemma4e_npu.dll b/src/lib/xrt/gemma4e_npu.dll
diff --git a/src/lib/xrt/gemma_embedding.dll b/src/lib/xrt/gemma_embedding.dll
diff --git a/src/lib/xrt/gemma_npu.dll b/src/lib/xrt/gemma_npu.dll
diff --git a/src/lib/xrt/gemma_text_npu.dll b/src/lib/xrt/gemma_text_npu.dll
diff --git a/src/lib/xrt/gpt_oss_npu.dll b/src/lib/xrt/gpt_oss_npu.dll
diff --git a/src/lib/xrt/lfm2_npu.dll b/src/lib/xrt/lfm2_npu.dll
diff --git a/src/lib/xrt/libgemma4e_npu.so b/src/lib/xrt/libgemma4e_npu.so
diff --git a/src/lib/xrt/libqwen3_5_omni_npu.so b/src/lib/xrt/libqwen3_5_omni_npu.so
diff --git a/src/lib/xrt/libqwen3_5vl_npu.so b/src/lib/xrt/libqwen3_5vl_npu.so
diff --git a/src/lib/xrt/libqwen3_6_moe_npu.so b/src/lib/xrt/libqwen3_6_moe_npu.so
diff --git a/src/lib/xrt/llama_npu.dll b/src/lib/xrt/llama_npu.dll
diff --git a/src/lib/xrt/lm_head.dll b/src/lib/xrt/lm_head.dll
diff --git a/src/lib/xrt/mha.dll b/src/lib/xrt/mha.dll
diff --git a/src/lib/xrt/nanbeige_npu.dll b/src/lib/xrt/nanbeige_npu.dll
diff --git a/src/lib/xrt/phi4_npu.dll b/src/lib/xrt/phi4_npu.dll
diff --git a/src/lib/xrt/q4_npu_eXpress.dll b/src/lib/xrt/q4_npu_eXpress.dll
diff --git a/src/lib/xrt/qwen2_npu.dll b/src/lib/xrt/qwen2_npu.dll
diff --git a/src/lib/xrt/qwen2vl_npu.dll b/src/lib/xrt/qwen2vl_npu.dll
diff --git a/src/lib/xrt/qwen3_5_omni_npu.dll b/src/lib/xrt/qwen3_5_omni_npu.dll
diff --git a/src/lib/xrt/qwen3_5vl_npu.dll b/src/lib/xrt/qwen3_5vl_npu.dll
diff --git a/src/lib/xrt/qwen3_6_moe_npu.dll b/src/lib/xrt/qwen3_6_moe_npu.dll
diff --git a/src/lib/xrt/qwen3_npu.dll b/src/lib/xrt/qwen3_npu.dll
diff --git a/src/lib/xrt/qwen3vl_npu.dll b/src/lib/xrt/qwen3vl_npu.dll
diff --git a/src/lib/xrt/whisper_npu.dll b/src/lib/xrt/whisper_npu.dll
diff --git a/src/wix/flm.wxs b/src/wix/flm.wxs
@@ -23,7 +23,7 @@
 
   <Package Name="flm"
            Manufacturer="FastFlowLM"
-           Version="1.0.1"
+           Version="1.0.2"
            UpgradeCode="5A5B2872-1E63-4934-9F57-088554CF4959"
            Language="1033"
            Scope="perMachine"

diff --git a/src/xclbins/DeepSeek-R1-0528-Qwen3-8B-NPU2/layer.xclbin b/src/xclbins/DeepSeek-R1-0528-Qwen3-8B-NPU2/layer.xclbin
diff --git a/src/xclbins/Deepseek-R1-Distill-Llama-8B-NPU2/layer.xclbin b/src/xclbins/Deepseek-R1-Distill-Llama-8B-NPU2/layer.xclbin
diff --git a/src/xclbins/Llama-3.1-8B-NPU2/layer.xclbin b/src/xclbins/Llama-3.1-8B-NPU2/layer.xclbin
diff --git a/src/xclbins/Llama-3.2-1B-NPU2/layer.xclbin b/src/xclbins/Llama-3.2-1B-NPU2/layer.xclbin
diff --git a/src/xclbins/Llama-3.2-3B-NPU2/layer.xclbin b/src/xclbins/Llama-3.2-3B-NPU2/layer.xclbin
diff --git a/src/xclbins/Nanbeige4.1-3B-NPU2/layer.xclbin b/src/xclbins/Nanbeige4.1-3B-NPU2/layer.xclbin
diff --git a/src/xclbins/Phi4-mini-Instruct-NPU2/layer.xclbin b/src/xclbins/Phi4-mini-Instruct-NPU2/layer.xclbin
diff --git a/src/xclbins/Qwen3-0.6B-NPU2/layer.xclbin b/src/xclbins/Qwen3-0.6B-NPU2/layer.xclbin
diff --git a/src/xclbins/Qwen3-1.7B-NPU2/layer.xclbin b/src/xclbins/Qwen3-1.7B-NPU2/layer.xclbin
diff --git a/src/xclbins/Qwen3-4B-Instruct-2507-NPU2/layer.xclbin b/src/xclbins/Qwen3-4B-Instruct-2507-NPU2/layer.xclbin
diff --git a/src/xclbins/Qwen3-4B-NPU2/layer.xclbin b/src/xclbins/Qwen3-4B-NPU2/layer.xclbin
diff --git a/src/xclbins/Qwen3-4B-Thinking-2507-NPU2/layer.xclbin b/src/xclbins/Qwen3-4B-Thinking-2507-NPU2/layer.xclbin
diff --git a/src/xclbins/Qwen3-8B-NPU2/layer.xclbin b/src/xclbins/Qwen3-8B-NPU2/layer.xclbin
diff --git a/src/xclbins/Qwen3-VL-4B-Instruct-NPU2/layer.xclbin b/src/xclbins/Qwen3-VL-4B-Instruct-NPU2/layer.xclbin
diff --git a/src/xclbins/Qwen3.5-0.8B-NPU2/GateDeltaNet_prefill.xclbin b/src/xclbins/Qwen3.5-0.8B-NPU2/GateDeltaNet_prefill.xclbin
diff --git a/src/xclbins/Qwen3.5-0.8B-NPU2/layer.xclbin b/src/xclbins/Qwen3.5-0.8B-NPU2/layer.xclbin
diff --git a/src/xclbins/Qwen3.5-0.8B-NPU2/lm_head.xclbin b/src/xclbins/Qwen3.5-0.8B-NPU2/lm_head.xclbin
diff --git a/src/xclbins/Qwen3.5-0.8B-NPU2/mm.xclbin b/src/xclbins/Qwen3.5-0.8B-NPU2/mm.xclbin
diff --git a/src/xclbins/Qwen3.5-2B-NPU2/GateDeltaNet_prefill.xclbin b/src/xclbins/Qwen3.5-2B-NPU2/GateDeltaNet_prefill.xclbin
diff --git a/src/xclbins/Qwen3.5-2B-NPU2/layer.xclbin b/src/xclbins/Qwen3.5-2B-NPU2/layer.xclbin
diff --git a/src/xclbins/Qwen3.5-2B-NPU2/lm_head.xclbin b/src/xclbins/Qwen3.5-2B-NPU2/lm_head.xclbin
diff --git a/src/xclbins/Qwen3.5-2B-NPU2/mm.xclbin b/src/xclbins/Qwen3.5-2B-NPU2/mm.xclbin
diff --git a/src/xclbins/Qwen3.5-4B-NPU2/GateDeltaNet_prefill.xclbin b/src/xclbins/Qwen3.5-4B-NPU2/GateDeltaNet_prefill.xclbin
diff --git a/src/xclbins/Qwen3.5-4B-NPU2/layer.xclbin b/src/xclbins/Qwen3.5-4B-NPU2/layer.xclbin
diff --git a/src/xclbins/Qwen3.5-4B-NPU2/lm_head.xclbin b/src/xclbins/Qwen3.5-4B-NPU2/lm_head.xclbin
diff --git a/src/xclbins/Qwen3.5-4B-NPU2/mm.xclbin b/src/xclbins/Qwen3.5-4B-NPU2/mm.xclbin
diff --git a/src/xclbins/Qwen3.5-9B-NPU2/GateDeltaNet_prefill.xclbin b/src/xclbins/Qwen3.5-9B-NPU2/GateDeltaNet_prefill.xclbin
diff --git a/src/xclbins/Qwen3.5-9B-NPU2/layer.xclbin b/src/xclbins/Qwen3.5-9B-NPU2/layer.xclbin
diff --git a/src/xclbins/Qwen3.5-9B-NPU2/lm_head.xclbin b/src/xclbins/Qwen3.5-9B-NPU2/lm_head.xclbin
diff --git a/src/xclbins/Qwen3.5-9B-NPU2/mm.xclbin b/src/xclbins/Qwen3.5-9B-NPU2/mm.xclbin
diff --git a/src/xclbins/Qwen3.6-35B-A3B-NPU2/GateDeltaNet_prefill.xclbin b/src/xclbins/Qwen3.6-35B-A3B-NPU2/GateDeltaNet_prefill.xclbin
diff --git a/src/xclbins/Qwen3.6-35B-A3B-NPU2/dequant_mm.xclbin b/src/xclbins/Qwen3.6-35B-A3B-NPU2/dequant_mm.xclbin
diff --git a/src/xclbins/Qwen3.6-35B-A3B-NPU2/layer.xclbin b/src/xclbins/Qwen3.6-35B-A3B-NPU2/layer.xclbin
diff --git a/src/xclbins/Qwen3.6-35B-A3B-NPU2/lm_head.xclbin b/src/xclbins/Qwen3.6-35B-A3B-NPU2/lm_head.xclbin
diff --git a/src/xclbins/Qwen3.6-35B-A3B-NPU2/mm.xclbin b/src/xclbins/Qwen3.6-35B-A3B-NPU2/mm.xclbin