support qwen3

csrc/ktransformers_ext/CMakeLists.txt

@@ -151,7 +151,6 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
         message(STATUS "Compiler and/or CPU do NOT support AVX512F")
         set(HAS_AVX512 False)
     endif()
-    set(HAS_AVX512 False)
 
     # check AMX
     string(FIND "${LSCPU_OUTPUT}" "amx" COMPILER_SUPPORTS_AMX)

ktransformers/optimize/optimize_rules/Qwen2-serve-amx.yaml

@@ -0,0 +1,96 @@
+- match:
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeRotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.RotaryEmbedding
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    name: "^lm_head$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+
+# - match:
+#     name: "^model\\.layers\\..*$"  # regular expression 
+#     class: torch.nn.Linear  # only match modules matching name and class simultaneously
+#   replace:
+#     class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+#     kwargs:
+#       generate_device: "cuda"
+#       prefill_device: "cuda"
+#       generate_op: "VLinearMarlin"
+#       prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\.(?!.*mlp\\.shared_expert_gate).*$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      generate_op: "VLinearMarlin"
+      prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\..*\\.mlp$"
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeSparseMoeBlock
+  replace:
+    class: ktransformers.operators.experts.KQwen2MoeSparseMoeBlockV2     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    name: "^model\\.layers\\..*\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExpertsV2     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op: "KExpertsCPU"
+      out_device: "cuda"
+      backend: "AMXInt8" # or "AMXBF16" or "llamafile" (default)
+  recursive: False # don't recursively inject submodules of this module
+- match:
+    name: "^model\\.layers\\..*\\.self_attn$"
+  replace:
+    class: ktransformers.operators.balance_serve_attention.KQwen2MoeAttention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+- match:
+    name: "^model$"
+  replace:
+    class: "ktransformers.operators.models.KQwen2MoeModel"
+    kwargs:
+      per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
+- match:
+    name: "^model.embed_tokens"
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cpu"
+      prefill_device: "cpu"
+
+- match:
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeRMSNorm
+  replace:
+    class: ktransformers.operators.layernorm.KQwen2MoeRMSNorm
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeMLP
+  replace:
+    class:  ktransformers.operators.mlp.KQwen2MoeMLP
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"

ktransformers/optimize/optimize_rules/Qwen2-serve.yaml

@@ -56,7 +56,6 @@
       generate_device: "cpu"
       generate_op: "KExpertsCPU"
       out_device: "cuda"
-      backend: "AMXInt8" # or "AMXBF16" or "llamafile" (default)
   recursive: False # don't recursively inject submodules of this module
 - match:
     name: "^model\\.layers\\..*\\.self_attn$"

ktransformers/optimize/optimize_rules/Qwen3Moe-serve-amx.yaml

@@ -0,0 +1,96 @@
+- match:
+    class: ktransformers.models.modeling_qwen2_moe.Qwen2MoeRotaryEmbedding
+  replace:
+    class: ktransformers.operators.RoPE.RotaryEmbedding
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    name: "^lm_head$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      generate_op: "VLinearMarlin"
+      prefill_op: "KLinearTorch"
+
+# - match:
+#     name: "^model\\.layers\\..*$"  # regular expression 
+#     class: torch.nn.Linear  # only match modules matching name and class simultaneously
+#   replace:
+#     class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+#     kwargs:
+#       generate_device: "cuda"
+#       prefill_device: "cuda"
+#       generate_op: "VLinearMarlin"
+#       prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\.(?!.*mlp\\.shared_expert_gate).*$"  # regular expression 
+    class: torch.nn.Linear  # only match modules matching name and class simultaneously
+  replace:
+    class: ktransformers.operators.linear.KTransformersLinear  # optimized Kernel on quantized data types
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+      generate_op: "KLinearMarlin"
+      prefill_op: "KLinearTorch"
+- match:
+    name: "^model\\.layers\\..*\\.mlp$"
+    class: ktransformers.models.modeling_qwen3_moe.Qwen3MoeSparseMoeBlock
+  replace:
+    class: ktransformers.operators.experts.KQwen3MoeSparseMoeBlockV2     # mlp module with custom forward function
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    name: "^model\\.layers\\..*\\.mlp\\.experts$"
+  replace:
+    class: ktransformers.operators.experts.KTransformersExpertsV2     # custom MoE Kernel with expert paralleism
+    kwargs:
+      prefill_device: "cuda"
+      prefill_op: "KExpertsTorch"
+      generate_device: "cpu"
+      generate_op: "KExpertsCPU"
+      out_device: "cuda"
+      backend: "AMXBF16" # or "AMXBF16" or "llamafile" (default)
+  recursive: False # don't recursively inject submodules of this module
+- match:
+    name: "^model\\.layers\\..*\\.self_attn$"
+  replace:
+    class: ktransformers.operators.balance_serve_attention.KQwen3MoeAttention # optimized MLA implementation
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+- match:
+    name: "^model$"
+  replace:
+    class: "ktransformers.operators.models.KQwen2MoeModel"
+    kwargs:
+      per_layer_prefill_intput_threshold: 0 # 0 is close layer wise prefill
+- match:
+    name: "^model.embed_tokens"
+  replace:
+    class: "default"
+    kwargs:
+      generate_device: "cpu"
+      prefill_device: "cpu"
+
+- match:
+    class: ktransformers.models.modeling_qwen3_moe.Qwen3MoeRMSNorm
+  replace:
+    class: ktransformers.operators.layernorm.KQwen3MoeRMSNorm
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"
+
+- match:
+    class: ktransformers.models.modeling_qwen3_moe.Qwen3MoeMLP
+  replace:
+    class:  ktransformers.operators.mlp.KQwen2MoeMLP
+    kwargs:
+      generate_device: "cuda"
+      prefill_device: "cuda"

ktransformers/optimize/optimize_rules/Qwen3Moe-serve.yaml

@@ -56,7 +56,6 @@
       generate_device: "cpu"
       generate_op: "KExpertsCPU"
       out_device: "cuda"
-      backend: "AMXBF16" # or "AMXBF16" or "llamafile" (default)
   recursive: False # don't recursively inject submodules of this module
 - match:
     name: "^model\\.layers\\..*\\.self_attn$"

ktransformers/server/args.py

@@ -94,7 +94,6 @@ class ArgumentParser:
         parser.add_argument("--user_algorithm", type=str, default=self.cfg.user_algorithm)
         parser.add_argument("--force_think", action=argparse.BooleanOptionalAction, type=bool, default=self.cfg.user_force_think)
         parser.add_argument("--use_cuda_graph", action=argparse.BooleanOptionalAction, type=bool, default=self.cfg.use_cuda_graph)
-        # parser.add_argument("--use_cuda_graph", action=argparse.BooleanOptionalAction, type=bool, default=False)
 
         # web config
         parser.add_argument("--web_cross_domain", type=bool, default=self.cfg.web_cross_domain)

ktransformers/server/backend/interfaces/balance_serve.py

@@ -56,8 +56,8 @@ ktransformer_rules_dir = (
     os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "..", "..", "./optimize/optimize_rules/") 
 )
 default_optimize_rules = {
-    "DeepseekV3ForCausalLM": ktransformer_rules_dir + "Moonlight-16B-A3B-serve.yaml",
-    # "DeepseekV3ForCausalLM": ktransformer_rules_dir + "DeepSeek-V3-Chat-serve.yaml",
+    # "DeepseekV3ForCausalLM": ktransformer_rules_dir + "Moonlight-16B-A3B-serve.yaml",
+    "DeepseekV3ForCausalLM": ktransformer_rules_dir + "DeepSeek-V3-Chat-serve.yaml",
     "Qwen2MoeForCausalLM": ktransformer_rules_dir + "Qwen2-serve.yaml",
     "Qwen3MoeForCausalLM": ktransformer_rules_dir + "Qwen3Moe-serve.yaml",
 }

ktransformers/server/balance_serve/inference/model_runner.py

@@ -85,7 +85,7 @@ class ModelRunner:
         elif isinstance(self.model, KQwen2MoeForCausalLM) or isinstance(self.model, KQwen3MoeForCausalLM):
             self.model.flash_infer_attn_plan(batch, self.bsz_tensor_buf, self.num_tokens_tensor_buf,
                                              num_q_heads=self.model.config.num_attention_heads, num_kv_heads=self.model.config.num_key_value_heads,
-                                             head_dim=128, 
+                                             head_dim=self.model.config.head_dim if hasattr(self.model.config, 'head_num') else self.model.config.hidden_size // self.model.config.num_attention_heads, 
                                              page_size=self.model.cache.page_size, causal=True,
                                              q_data_type=torch.bfloat16, kv_data_type=torch.bfloat16, cuda_graph_idx=cuda_graph_idx)
         else: